epublications.uef.fi · DISSERTATIONS | TIMO PEKKALA | MULTIMODAL PREDICTION OF DEMENTIA AND BRAIN...

DIS

SE

RT

AT

ION

S | T

IMO

PE

KK

AL

A | M

UL

TIM

OD

AL

PR

ED

ICT

ION

OF

DE

ME

NT

IA A

ND

BR

AIN

PA

TH

OL

OG

Y | N

o 563

uef.fi

PUBLICATIONS OF THE UNIVERSITY OF EASTERN FINLAND

Dissertations in Health Sciences

ISBN 978-952-61-3379-9ISSN 1798-5706



TIMO PEKKALA

MULTIMODAL PREDICTION OF DEMENTIA AND BRAIN PATHOLOGY

Dementia causes a considerable burden on individuals and societies, and interventions

at earlier stages should be developed. In this thesis dementia and related neuropathology are predicted in elderly cognitively healthy individuals in order to identify high-risk

individuals for interventions and to enrich targetable pathologies in trial populations. Also, the study investigates the associations

of markers of early type 2 diabetes and brain amyloid deposition, a hallmark of

Alzheimer’s disease.

TIMO PEKKALA

MULTIMODAL PREDICTION OF DEMENTIAAND BRAIN PATHOLOGY

Timo Pekkala

MULTIMODAL PREDICTION OF DEMENTIAAND BRAIN PATHOLOGY

To be presented by permission of theFaculty of Health Sciences, University of Eastern Finland

for public examination in Kuopioon April 28th, 2020, at 12 o’clock noon

Publications of the University of Eastern FinlandDissertations in Health Sciences

No 563

Institute of Clinical Medicine, NeurologySchool of Medicine, Faculty of Health Sciences

University of Eastern FinlandKuopio

2020

Series EditorsProfessor Tomi Laitinen, M.D., Ph.D.

Institute of Clinical Medicine, Clinical Physiology and Nuclear MedicineFaculty of Health Sciences

Associate professor (Tenure Track) Tarja Kvist, Ph.D.Department of Nursing Science

Faculty of Health Sciences

Professor Kai Kaarniranta, M.D., Ph.D.Institute of Clinical Medicine, Ophthalmology


Associate Professor (Tenure Track) Tarja Malm, Ph.D.A.I. Virtanen Institute for Molecular Sciences


Lecturer Veli-Pekka Ranta, Ph.D.School of Pharmacy


Distributor:University of Eastern Finland

Kuopio Campus LibraryP.O. Box 1627

FI-70211 Kuopio, Finlandwww.uef.fi/kirjasto

Name of the printing office/kirjapainoGrano Oy, 2020

ISBN: 978-952-61-3379-9 (print/nid.)ISBN: 978-952-61-3380-5 (PDF)

ISSNL: 1798-5706ISSN: 1798-5706

ISSN: 1798-5714 (PDF)

Author’s address: Institute of Clinical Medicine, NeurologyUniversity of Eastern FinlandKUOPIOFINLAND

Doctoral programme: Doctoral Programme of Clinical Research

Supervisors: Associate Professor Alina Solomon, M.D., Ph.D.Institute of Clinical Medicine, NeurologyUniversity of Eastern FinlandKUOPIOFINLAND

Anette Hall, Ph.D.Institute of Clinical Medicine, NeurologyUniversity of Eastern FinlandKUOPIOFINLAND

Docent Tiia Ngandu, M.D., Ph.D.Public Health Promotion UnitFinnish Institute for Health and Welfare (THL)HELSINKIFINLAND

Professor Hilkka Soininen, M.D., Ph.D.Institute of Clinical Medicine, NeurologyUniversity of Eastern FinlandKUOPIOFINLAND

Reviewers: Chinedu Udeh-Momoh, Ph.D.Ageing Epidemiology Research Unit, School of Public HealthImperial College LondonLONDONUNITED KINGDOM

Tamlyn J. Watermeyer, Ph.D.Centre for Clinical Brain SciencesThe University of EdinburghEDINBURGHUNITED KINGDOM

Opponent: Professor Lefkos Middleton, M.D., Ph.D.Ageing Epidemiology Research Unit, School of Public HealthImperial College LondonLONDONUNITED KINGDOM

Pekkala, TimoMultimodal prediction of dementia and brain pathologyKuopio: University of Eastern FinlandPublications of the University of Eastern FinlandDissertations in Health Sciences 563. 2020, 120 p.ISBN: 978-952-61-3379-9 (print)ISSNL: 1798-5706ISSN: 1798-5706ISBN: 978-952-61-3380-5 (PDF)ISSN: 1798-5714 (PDF)

ABSTRACT

Dementia and associated brain pathology take years to develop. Effective interven-tions to prevent dementia have not been found, in part because interventions aretargeted at individuals in a relatively late stage of dementia progression. This thesisaims to develop prediction models for identifying persons at risk at an earlier stage.Prediction targets included incident dementia as well as common brain pathologiesunderlying progressive cognitive disorders in different elderly age cohorts. An ad-ditional aim was to investigate the association of blood markers of type two diabetes(DM2) and brain amyloid deposition, a hallmark of Alzheimer’s disease (AD).

Dementia was predicted in the Finnish population based Cardiovascular Risk Fac-tors, Aging and Dementia (N=709 and 1,009) and Vantaa 85+ (N=245) study popula-tions of cognitively healthy younger-old individuals (mean age 70 years) and oldest-old individuals (88 years), respectively. Multimodal predictors were used to predictincident dementia over a period of five to ten years using a Disease State Index (DSI)machine learning system. Incidences of common brain pathology were predicted ina Vantaa 85+ subpopulation (N=163, 89 years) over a four year follow up, and theprevalence of brain amyloid deposition on positron emission tomography (PET) waspredicted in a Finnish Geriatric Intervention Study to Prevent Cognitive Impairmentand Disability (FINGER) subpopulation (N=48) of cognitively healthy younger-oldindividuals (71 years) with elevated cardiovascular risks and cognition at or slightlybelow population norms. Both prediction models were built using the DSI. A furtherFINGER-PET subpopulation (N=41) was used for the analysis of blood DM2 markersusing a logistic regression.

Prediction of dementia in the younger-old population succeeded well (area underthe curve 0.75–0.79), and in the oldest-old population almost at the same level (0.73).Predictors of dementia for the younger old and the oldest old were different, withage and vascular health achieving less effective predictions for the older cohort. Forthe oldest old, dementia could be predicted more accurately than most types of brainpathology (0.61–0.72). Amyloid deposition was predicted well for the younger old(0.78) using among other modalities magnetic resonance imaging, but the predictionresults were better than for the oldest old even without imaging. Cognition was

7

a better predictor of dementia than pathology, and the apolipoprotein E genotypewas a better predictor of pathology than dementia. Out of the DM2 markers, lowlevels of insulin resistance markers and a low concentration of plasminogen activatorinhibitor-1 were associated with a positive brain amyloid deposition status.

These results indicate that at-risk persons could be identified years before a diag-nosis of dementia is given, and interventions could be targeted at those who benefitthe most. Different risk factors may have to be considered when targeting dementiaor specific pathologies. Prediction models for brain pathology—especially amyloid—could be used to enrich study populations with persons with a specific pathology tosave costs and invasive assessments in clinical trials.

Medical Subject Headings: Aged; Alzheimer Disease; Amyloid; Brain/pathology; Cognition;Cognitive Dysfunction; Decision Support Systems, Clinical; Dementia; Diabetes Mellitus,Type 2; Early Medical Intervention; Incidence; Longitudinal Studies; Neuropathology; RiskFactors

8

Pekkala, TimoDementian ja aivopatologian monityyppinen ennustaminenKuopio: Ita-Suomen yliopistoPublications of the University of Eastern FinlandDissertations in Health Sciences 563. 2020, 120 s.ISBN: 978-952-61-3379-9 (nid.)ISSNL: 1798-5706ISSN: 1798-5706ISBN: 978-952-61-3380-5 (PDF)ISSN: 1798-5714 (PDF)

TIIVISTELMA

Dementia ja sen taustalla vaikuttavat aivojen patologiset muutokset kehittyvat usei-den vuosien aikana. Tehokkaita keinoja dementian ehkaisemiseksi ei viela ole loy-detty. Osin tama saattaa johtua siita, etta ehkaisytoimia on tahan mennessa tutkit-tu dementian melko myohaisessa kehitysvaiheessa. Tama vaitostyo pyrki kehitta-maan ennustemalleja, joilla voitaisiin aiemmin tunnistaa henkilot, joilla on suuren-tunut dementiariski. Tyossa pyrittiin ennustamaan toisaalta dementian ja toisaaltayleisimpien aivojen dementiaan liitettyjen patologisten muutosten ilmaantumista.Ennustemalleja sovellettiin eri-ikaisiin vanhuusian kohortteihin. Lisaksi vaitostyos-sa selvitettiin tyypin kaksi diabeteksen verimerkkiaineiden pitoisuuksien yhteytta ai-vojen amyloidiproteiinikertymien esiintyvyyteen. Amyloidiproteiinin kertyminenaivokudokseen on yksi Alzheimerin taudin tyypillisista muutoksista.

Dementian ilmaantuvuutta ennustettiin kahden suomalaisen vaestopohjaisen tut-kimuksen aineistolla. Cardiovascular Risk Factors, Aging and Dementia -tutkimuk-sen (N=709 ja 1 009) koehenkilot olivat kognitiivisesti terveita keskimaarin 70-vuo-tiaita nuoria ikaantyneita ja Vantaa 85+ -tutkimuksen (N=245) henkilot taas keski-maarin 88-vuotiaita vanhoja ikaantyneita. Malleilla ennustettiin ilmaantuvuutta vi-idesta kymmeneen vuoden ajanjaksolla ja ennustetekijoina kaytettiin eri terveydenosa-alueilta mitattuja monityyppisia tekijoita. Mallit toteutettiin Disease State Index(DSI) koneoppimisjarjestelmalla. Aivojen patologisten muutosten ilmaantumista en-nustettiin Vantaa 85+ -tutkimuksen ruumiinavausosapopulaatiossa (N=163, ika kes-kimaarin 89 vuotta) keskimaarin neljan vuoden seurantajaksolla. Amyloidiprotei-inin esiintymista positroniemissiotomografiassa (PET) ennustettiin Finnish GeriatricIntervention Study to Prevent Cognitive Impairment and Disability -tutkimuksen(FINGER) pienessa osapopulaatiossa (N=48). FINGER-tutkimuksen koehenkilot olivalittu siten, etta he olivat kognitiivisesti terveita, mutta heidan kognition tasonsaoli mittauksissa vaestokeskiarvon mukainen tai hieman heikompi. Lisaksi heilla olisuurentunut sydan- ja verisuonisairauksien riski. Myos patologian ennustemallit pe-rustuivat DSI-jarjestelmaan. Tyypin kaksi diabeteksen merkkiaineiden ja amyloidinsuhdetta tutkittiin logistisella regressiolla hieman pienemmassa FINGER-tutkimuk-sen osajoukossa (N=41).

9

Dementian ennustaminen nuorten ikaantyneiden ryhmassa onnistui hyvin (AUC0.75–0.79) ja vanhojen ikaantyneiden ryhmassa lahes yhta hyvin (0.73). Tarkeim-mat ennustetekijat poikkesivat toisistaan eri ikaryhmissa: ika ja verisuonielimistonterveydentila olivat huonompia ennustetekijoita vanhojen ikaantyneiden ryhmassa.Vanhojen ikaantyneiden ryhmassa dementian ilmaantumista pystyttiin ennustamaantarkemmin kuin useimpien patologisten muutosten ilmaantumista (0.61–0.72). Amy-loidiproteiinin esiintymista aivokuvantamisessa pystyttiin ennustamaan hyvin nuor-ten ikaantyneiden ryhmassa (0.78), kun ennustetekijana kaytettiin muun muassa ai-vojen magneettikuvaustuloksia. Tulokset olivat tosin parempia nuorten ikaantynei-den ryhmassa verrattuna vanhojen ikaantyneiden ryhmaan, vaikka magneettikuvaus-tuloksia ei olisi ollut kaytettavissa. Kognition taso ennusti paremmin dementiankuin aivopatologian ilmaantuvuutta. Apolipoproteiini E:n genotyyppi taas ennustiparemmin patologian kuin dementian ilmaantuvuutta. Tyypin kaksi sokeritaudinmerkkiaineista matalaan insuliiniresistenssiin viittaavat merkkiainepitoisuudet ja ma-tala PAI-1-pitoisuus (plasminogeeni aktivaattori-1:n inhibiittori) olivat yhteydessa po-sitiiviseen amyloidiloydokseen.

Nama tulokset osoittivat, etta suuremman dementiariskin henkilot voidaan tun-nistaa vuosia ennen sairastumista. Tana diagnoosia edeltavana ajanjaksona voitaisiintoteuttaa interventioita niille, jotka niista eniten hyotyisivat. Kohdennettavat riskite-kijat tulisi valita sen mukaan, pyritaanko ehkaisemaan dementian tai tiettyjen patolo-gisten muutosten ilmaantumista. Patologisten muutosten ennustemalleilla voitaisiinrikastaa tutkimuspotilaita tietyn patologian—etenkin amyloidiproteiinin—suhteenkustannusten ja kajoavien toimenpiteiden vahentamiseksi kliinisissa tutkimuksissa.

Yleinen suomalainen asiasanasto: aikuistyypin diabetes; aivot; Alzheimerin tauti; dementia;ennaltaehkaisy; ennusteet; ikaantyneet; ilmaantuvuus; kognitio; pitkittaistutkimus; paatok-sentukijarjestelmat; riskitekijat

10

ACKNOWLEDGEMENTS

This doctoral thesis was carried out in the Institute of Clinical Medicine, Departmentof Neurology at the University of Eastern Finland, Kuopio, during the years 2014–2020. This project has been done in collaboration with the Public Health PromotionUnit at the National Institute for Health and Welfare in Finland and Division of Clin-ical Geriatrics, NVS, Karolinska Institute in Sweden. Many individuals have sup-ported this work and have made invaluable contributions. I would especially wantto thank the following people.

My main supervisor, Associate Professor Alina Solomon, for introducing me toclinical research and guiding my work by engaging with me in intellectually stimu-lating conversations, and also for the remarkable flexibility that combining medicalstudies and research work has required.

My co-supervisor, Doctor Anette Hall, who I shared a workroom with for theentire duration of the project. Your uncompromising attention to the project, thedetailed and knowledgeable feedback you have given, and the constant general en-couragement over these years have truly been priceless. Not to mention your insightsinto local life in Kuopio. The restaurant tips were especially appreciated.

My co-supervisor, Docent Tiia Ngandu, for the thorough and intelligent contri-bution bringing the project together especially in the later stages. And also for theinspiring discussions of career paths and of the future.

My co-supervisor, Professor Hilkka Soininen, who first introduced me to the ideaof research in the field of neurology. You bringing me together with Alina and Anettewas insightful, and your speedy and accurate commentary on the work in progresshas been essential.

Doctors Chinedu Udesh-Momoh and Tamlyn Watermeyer, for the thorough re-view of the thesis and for your valuable comments and criticism.

All the co-authors, for your inputs on the analysis and efforts with the manuscriptpreparations. Especially Miia Kivipelto for enabling collaboration between Helsinkiand Kuopio, Jyrki Lotjonen for all the help with the DSI, and Francesca Mangialaschefor help with the biomarker assay data.

My wonderful colleagues at the university for their encouragement, frequent livelylunches, and new ideas. Members of the Nordic Brain Network for the opportunityto show my work at different stages, and for the honest criticism and feedback. Andmany friends at medical school in Kuopio and back home in Helsinki for all theirfantastic support.

The Department of Neurology staff and especially Esa Koivisto and Mari Tikka-nen for all the technical help and practical advice.

My parents, siblings, their spouses, and their offspring. Thank you Anne and Karifor nudging me in the right direction in the early years. Jussi and Anna, Anni andKalle, thank you for the cottage life and for taking my mind off school and academiaevery now and then. My grandmother Paula, you always trusted in me, and your

11

positive outlook really helped getting it all together. Hemmo and Vaara, seeing youbabies grow up brought tremendous joy to your uncle. And also my in-law familyAuvinen, thank you for making me feel at home in Kuopio.

My fiance Timo A., for introducing me to the Savonian people and agreeing tomove to Kuopio for the years we spent there. And also for understanding the careerchange from engineering and banking to medicine. Having you by my side made itall worthwhile.

This work was funded and supported by Doctoral Program of Clinical Research atthe University of Eastern Finland, European Research Council grant 804371, Academyof Finland, Finnish Social Insurance Institution, Juho Vainio Foundation, Yrjo Jahns-son Foundation, Finnish Cultural Foundation; Alzheimer’s Research and Preven-tion Foundation; Swedish Research Council, Alzheimerfonden, Region StockholmALF and NSV grants, Center for Innovative Medicine (CIMED) at Karolinska Insti-tutet, Knut and Alice Wallenberg Foundation, Stiftelsen Stockholms sjukhem, Ko-nung Gustaf V:s och Drottning Victorias Frimurarstiftelse (Sweden); Joint Program ofNeurodegenerative Disorders – prevention (MIND-AD); EVO/VTR grants for Kuo-pio, Oulu, and Turku University Hospitals, and Seinajoki Central Hospital and OuluCity Hospital. The funders had no role in the study design, data collection and anal-ysis, or preparation and publishing of the manuscripts of the original publications ofthis thesis. The author declares no conflicts of interest.

Helsinki, April 2020

Timo Pekkala

12

LIST OF ORIGINAL PUBLICATIONS

This dissertation is based on the following original publications:

I Pekkala T, Hall A, Lotjonen J, Mattila J, Soininen H, Ngandu T, Laatikainen T,Kivipelto M and Solomon A. Development of a late-life dementia predictionindex with supervised machine learning in the population-based CAIDE study.Journal of Alzheimer’s Disease 55: 1055–1067, 2017.

II Hall A, Pekkala T, Polvikoski T, van Gils M, Kivipelto M, Lotjonen J, Mattila J,Kero M, Myllykangas L, Makela M, Oinas M, Paetau A, Soininen H, Tanska-nen M and Solomon A. Prediction models for dementia and neuropathology inthe oldest old: the Vantaa 85+ cohort study. Alzheimer’s Research & Therapy11, 2019.

III Pekkala T, Hall A, Ngandu T, van Gils M, Helisalmi S, Hanninen T, Kemp-painen N, Liu Y, Lotjonen J, Paajanen T, Rinne J O, Soininen H, Kivipelto M,Solomon A. Detecting amyloid positivity in elderly with increased risk of cog-nitive decline. Submitted to journal for publication.

IV Pekkala T, Hall A, Mangialasche F, Kemppainen N, Mecocci P, Ngandu T, RinneJ O, Soininen H, Tuomilehto J, Kivipelto M and Solomon A. Association of pe-ripheral insulin resistance and other markers of type 2 diabetes mellitus withbrain amyloid deposition in healthy individuals at risk of dementia. Submittedto journal for publication.

The publications were adapted with the permission of the copyright owners.

13

CONTENTS

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

TIIVISTELMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 REVIEW OF THE LITERATURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.1 Cognitive decline and dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.2 Alzheimer’s disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.2.1 Clinical presentation and criteria for clinical diagnosis . . . . . . . . . . . . . 242.2.2 Pathophysiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.2.3 AD biomarkers and biomarker-based diagnosis . . . . . . . . . . . . . . . . . . . . 27

2.3 Vascular cognitive disorders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.3.1 Clinical presentation and diagnostic criteria . . . . . . . . . . . . . . . . . . . . . . . . . 292.3.2 Pathophysiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.4 Lewy body dementias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.5 Other dementias. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332.6 Risk factors for dementia and brain pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.6.1 Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342.6.2 Risk genes and causal mutations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.6.3 Cardiovascular risk factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.6.4 Insulin resistance and diabetes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.6.5 Lifestyle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412.6.6 Psychosocial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.7 A multimodal approach to dementia risk management . . . . . . . . . . . . . . . . . . . . . 432.8 Dementia risk models and scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.8.1 Definition of the prediction problem in a medical context . . . . . . . . . . 442.8.2 Diagnostics of a prediction model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462.8.3 Statistical methods underlying prediction models . . . . . . . . . . . . . . . . . . . 472.8.4 Prognostic prediction of dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482.8.5 Prediction of brain amyloid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522.8.6 Prediction models in prevention trials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3 AIMS OF THE STUDY. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

4 SUBJECTS AND METHODS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594.1 The CAIDE study of younger old individuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594.2 The Vantaa 85+ study of oldest old individuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

15

4.3 The FINGER trial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644.4 Disease State Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664.5 Data analysis and prediction models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.5.1 Prediction and validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684.5.2 Association analysis in FINGER IR/DM cohort . . . . . . . . . . . . . . . . . . . . . . 69

5 RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715.1 Predicting incident dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

5.1.1 Population characteristics, and predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715.1.2 Dementia prediction in the younger old (CAIDE) . . . . . . . . . . . . . . . . . . . . 735.1.3 Dementia prediction in the older old (Vantaa 85+) . . . . . . . . . . . . . . . . . . 755.1.4 Dementia and neuropathology at death . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.2 Predicting brain pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 765.2.1 Longitudinal prediction of pathology (Vantaa 85+) . . . . . . . . . . . . . . . . . . 765.2.2 Diagnostic prediction of brain amyloid (FINGER) . . . . . . . . . . . . . . . . . . . 785.2.3 Biomarkers of DM, and brain amyloid (FINGER) . . . . . . . . . . . . . . . . . . . . 80

6 DISCUSSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 836.1 Prediction of incident dementia in the younger old . . . . . . . . . . . . . . . . . . . . . . . . . . 836.2 Prediction of incident dementia in the oldest old . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 856.3 Prediction of brain amyloid and AD-type pathology . . . . . . . . . . . . . . . . . . . . . . . . . 86

6.3.1 Frequency of amyloid beta, APOE ε4, and dementia . . . . . . . . . . . . . . 866.3.2 Amyloid beta prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 876.3.3 Associations of metabolic markers of diabetes and amyloid beta 896.3.4 Prediction of other AD and amyloid related pathology . . . . . . . . . . . . . 90

6.4 Prediction of other brain pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 916.4.1 Vascular pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 916.4.2 Hippocampal sclerosis, TDP-43 protein, and α-synuclein . . . . . . . . . 92

6.5 Prediction of dementia versus brain pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

7 CONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

8 FUTURE PERSPECTIVES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

16

ABBREVIATIONS

Aβ amyloid beta protein

AD Alzheimer’s disease

ADDTC Alzheimer’s Disease Diagnostic and Treatment Centers

ADL activities in daily living

ADNI Alzheimer’s Disease Neuroimaging Initiative

AF atrial fibrillation

ANU-ADRI Australian National University Alzheimer’s Disease Risk Index

APOE Apolipoprotein E gene

APP amyloid precursor protein

AT(N) amyloid, tau, and neurodegeneration

AUC area under the receiver operating characteristic curve

BACE beta-secretase

BBB blood–brain barrier

BDI Beck Depression Inventory

BMI body mass index

BP blood pressure

CAA cerebral amyloid angiopathy

CAD coronary artery disease

CAIDE Cardiovascular Risk Factors, Aging and Dementia

CARTS cerebral age-related TDP-43 with sclerosis

CDR clinical dementia rating

CERAD Consortium to Establish a Registry for Alzheimer’s Disease

CHD coronary heart disease

CI confidence interval

CN cognitively normal

17

CRF cardiorespiratory fitness

CSF cerebrospinal fluid

CVD cerebrovascular disease

DBP diastolic blood pressure

DLB dementia with Lewy bodies

DM diabetes mellitus

DSI Disease State Index

DSM-5 fifth revision of the DSM

DSM-IV-TR Diagnostic and Statistical Manual of Mental Disorders, 4th, re-vised, edition

DT decision tree

EPAD European Prevention of Alzheimer’s Dementia

FCSRT-FR free and cued selective reminding test, free recall

FINGER Finnish Geriatric Intervention Study to Prevent Cognitive Im-pairment and Disability

FINMONICA Finnish part of Monitoring Trends and Determinants in Cardio-vascular Disease

FTD frontotemporal dementia

GIP gastric inhibitory polypeptide

GLP-1 glucagon-like peptide-1

HbA1c glycated hemoglobin

HDL high-density lipoprotein

HIV human immunodeficiency virus

HOMA-IR homeostasis model assessment for insulin resistance

HS hippocampal sclerosis

IADL Instrumental Activities of Daily Living

ICD-10 International Classification of Diseases 10th revision

ICH intracerebral hemorrhage

IDE insulin-degrading enzyme

18

In-MINDD Innovative Midlife Intervention for Dementia Deterrence

IR insulin resistance

IWG International Working Group

LDL low-density lipoprotein

LIBRA Lifestyle for Brain health

LR logistic regression

MAPT Multidomain Alzheimer’s Prevention Trial

MCI mild cognitive impairment

MMSE Mini-Mental State examination

mNTB modified version of the Neuropsychological Test Battery

MRI magnetic resonance imaging

MTA medial temporal atrophy

NCD neurocognitive disorder

NIA-AA National Institute on Aging and the Alzheimer’s Association

NIA-RIA National Institute for Aging and Ronald and Nancy Reagan In-stitute of the Alzheimer’s Association

NINCDS-ADRDA National Institute of Neurological and Communicative Disordersand Stroke and the Alzheimer’s Disease and Related DisordersAssociation

NINDS-AIREN National Institute of Neurological Disorders and Stroke–Associa-tion Internationale pour la Recherche et l’Enseignement en Neu-rosciences

NSAID nonsteroidal anti-inflammatory drug

PAI-1 plasminogen activator inhibitor-1

PC Principal component

PCA principal components analysis

PD-D Parkinson’s disease dementia

PET positron emission tomography

PIB Pittsburgh compound B

PPA primary progressive aphasia

19

PPV positive predictive value

PreDIVA Prevention of Dementia by Intensive Vascular Care

RCT randomized controlled trial

RF random forest model

ROC Receiver operating characteristics

SBP systolic blood pressure

SMQ Subjective Memory Questionnaire

SPMSQ Short portable mental status questionnaire

SPRINT Systolic Blood Pressure Intervention Trial

SVM support vector machine

TDP-43 TAR DNA-binding protein with molecular weight 43 kDa

THIN Taiwanese Health Improvement Network

TIA transient ischemic attack

TRIPOD Transparent Reporting of a multivariable prediction model forIndividual Prognosis Or Diagnosis

VaD vascular dementia

VASCOG International Society of Vascular Behavioural and Cognitive Dis-orders

VCD vascular cognitive disorders

VCI vascular cognitive impairment

WM white matter

20

1 INTRODUCTION

Despite efforts to develop disease-modifying interventions to prevent Alzheimer’sdisease (AD), dementia and the underlying diseases are still prevalent. Addition-ally, with populations growing older in many parts of the world, total prevalence isexpected to grow further. There are, however, signs that the age-specific incidencemight be decreasing in some regions (Seblova et al., 2018). Why this is exactly is notclear, but improvements in living standards and reductions of certain risk factors ofdementia in the population may be partly responsible. Research into modifiable riskfactors of dementia has remained in focus as AD drug trials have so far failed, andtrials aiming to prevent cognitive decline and dementia have gained in importance.Interventions targeting single risk factors have often not proven successful, but newmultimodal interventions have shown promise (Kivipelto et al., 2018). Globally, de-mentia prevention has been set as a priority, and the World Health Organization hasjust recently published public health guidelines for prevention that are suitable for in-tegration into multifaceted health promotion initiatives (World Health Organization,2019).

A problem with both disease-modifying and preventive interventions is the longtime frame of dementia development. Intervention would probably have to be under-taken at an early stage to be effective. Recognizing at-risk individuals up to decadesearlier is challenging, although risk scores have been used to this end. One objectiveof this thesis is to build and validate such models for the purposes of future trials.An important feature of such models is also to communicate the determinants ofrisk, which may be beneficial at an individual level.

Progress is being made not only on the epidemiological level with risk factorsand their mitigation, but also with the pathophysiology of AD and other primarydementias. Measuring brain pathology via imaging and other markers offers earlyinformation on the disease process, and may also indicate disease severity more pre-cisely than the clinical state. Additionally, measurement of pathology can be a usefulindicator of intervention efficacy. One of the objectives of this thesis is to predict thepresence and incidence of brain pathology. Such a prediction tool could be usefulin guiding persons for further investigations, or for example to invite persons to anintervention trial targeting that specific pathology.

Type two diabetes, like dementia, is a growing problem in modern aging societies,and the two share risk factors. Diabetes is thought to be a risk factor for dementia,but causality and the possible mechanisms are not yet fully clear. For example, dia-betes increases vascular brain pathology, but findings have been conflicting regard-ing direct causal associations with AD pathology. Other shared, rather than direct,causal factors have also been suggested to underlie the diabetes–AD association. E.g.hypercortisolemia associated with early stages of AD may provoke disturbances inglucose metabolism (Notarianni, 2017). Pre-diabetes with elevated insulin resistancemay share pathological pathways with AD, and understanding these may aid in de-

21

signing interventions. Thus, this thesis also investigates the association of metabolicchanges preceding DM2 and brain amyloid accumulation.

22

2 REVIEW OF THE LITERATURE

2.1 COGNITIVE DECLINE AND DEMENTIA

The aging process involves changes in brain function and cognition, but these normalchanges allow an individual to age with autonomy and a well-functioning everydaylife. Cognitive impairment in this context is seen as a deviation from this path. Classi-cally, mild cognitive impairment (MCI) is a term used to describe early steps towardsthe pathological, where a subjective experience of cognitive decline can be backed upwith objectively measured impairment of cognition (Roberts and Knopman, 2013).In the case of progressive cognitive disorders, MCI usually progresses to dementia,which in turn is characterized by considerable functional disability due to increasingcognitive impairment.

The term dementia referring to a form of extreme mental incapacity goes back tothe 1520s, and accounts of dementia as a mental state go back to antiquity. The termdementia has historically been used for both senile dementia—dementia occurring inold age—and for dementia due to a somatic or psychiatric cause such as schizophre-nia or syphilis. Contemporary clinical practice tends towards retiring the term dueto the associated stigma in favour of neurocognitive disorder, or in the Finnish case theterm muistisairaus, memory disorder. Furthermore, as knowledge about the underlyingdiseases progresses, more disease-specific terminology is being increasingly used.

The diagnostic criteria for dementia—irrespective of etiology—in different diag-nostic systems have evolved in the past decades. The Diagnostic and Statistical Man-ual of Mental Disorders, 4th, revised, edition from 2000 (DSM-IV-TR; American Psy-chiatric Association, 2000), characterizes dementia as deterioration of cognition inmultiple domains. Memory deficit is a required criterion, in addition to impairmentof language skills, impairment of motor function, agnosia, or impaired executivefunctioning. The impairment should represent a decline from the previous level andbe so severe that occupational or social functioning is harmed, that is, activities indaily living (ADL) are impaired. DSM-IV-TR emphasizes cognitive testing for deter-mining the deficits. The International Classification of Diseases 10th revision (ICD-10;World Health Organization, 1993) defines dementia similarly primarily as a deficit ofmemory.

In the fifth revision of the DSM from 2013 (DSM-5; American Psychiatric Asso-ciation, 2013) six distinct domains of cognition are specified, and memory deficit isnot a requirement anymore. The term dementia has also been rephrased as a majorneurocognitive disorder (NCD). To represent less severe cognitive impairment that haspreviously been characterized as MCI and prodromal dementia, a new diagnosticcategory of mild NCD was introduced. The diagnosis of major NCD as opposed tomild NCD requires a lack of independent ADL.

The National Institute on Aging and the Alzheimer’s Association (NIA-AA) work-group criteria for all-cause dementia (McKhann et al., 2011) require neuropsychiatric

23

symptoms that interfere with at least two practical categories of daily living. Thecriteria allow deficits to be determined based on patient or informant history and asimple clinical assessment. Similarly to the DSM-5, a memory deficit is not an abso-lute requirement.

Neuropsychological testing is used to quantify deficits in cognition (Salmon andBondi, 2009), and validated methods exist to screen for and to assess the severity ofdementia. The Clinical Dementia Rating (CDR; Berg, 1988; Morris, 1993) is commonlyused to identify dementia stages from very mild to severe dementia on a four ladderscale. The assessment is based on an interview that focuses on memory and five othercognitive domains with the emphasis on memory deficits. The Mini-Mental State ex-amination (MMSE; Folstein et al., 1975) is used for screening for cognitive decline indifferent outpatient care settings and to gauge the development of diagnosed mem-ory disorders. It consists of a 19-item-long test battery that tests several cognitive do-mains and can be administered with little training. The quality of cognitive deficits iscommonly measured using The Consortium to Establish a Registry for Alzheimer’sDisease (CERAD; Morris et al., 1989) neuropsychological battery, which is well suitedas a first line of assessment for persons with suspected AD. More nuanced neuropsy-chological assessments are used to either characterize very mild symptoms or to per-form differential diagnostics (Salmon and Bondi, 2009).

Dementia, as defined above, is thought of as a syndrome that is distinctly removedfrom healthy aging. The syndrome is defined as a symptomatic entity, and it can haveany of the several specific underlying pathologies as a cause. The following chaptersintroduce the main causative pathologies, of which Alzheimer’s disease is the mostcommon and widely known.

2.2 ALZHEIMER’S DISEASE

Alzheimer’s disease was first defined as a clinical dementia entity. The first sectiongives an overview of the classical phenotype-oriented diagnostic frameworks. Then,a summary of AD pathology and associated biomarkers and examination possibilitiesis given. The last section presents newer diagnostic frameworks that use biomarkerdata at their core to define AD.

2.2.1 Clinical presentation and criteria for clinical diagnosis

The National Institute of Neurological and Communicative Disorders and Strokeand the Alzheimer’s Disease and Related Disorders Association (NINCDS-ARDRA)guidelines for clinical AD diagnosis (McKhann et al., 1984) are based on strictly clini-cal findings for probable AD diagnoses, and neuropathological evidence is needed forthe diagnosis of definite AD. Deficits in two or more domains of cognition are requiredand neurophysiological testing is emphasized in determining the deficits. The age ofdisease onset should be at 40–90 years. A category of possible AD was introduced todescribe atypical disease presentations with no other likely cause. DSM-IV-TR, simi-lar to NINCDS-ARDRA, recognizes dementia of the Alzheimer’s type with a disease-

24

specific requirement of advancing cognitive decline with a gradual onset. Early-onsetAD-dementia (<65 years) is recognized as an additional entity, as are four subtypequalifiers (delirium, delusions, depression, uncomplicated). Impairment of ADL isrequired, in contrast to the NINCDS-ADRDA guidelines where ADL impairment islisted as supportive for probable AD.

In 2011 the NIA-AA workgroup (McKhann et al., 2011) published updated diag-nostic criteria for dementia with the aim of incorporating state-of-the-art scientificknowledge on AD as causative for dementia. The core NIA-AA dementia criteria(section 2.1) are mandated for AD diagnosis. The suggested diagnostic procedure rec-ognizes several levels of diagnostic certainty and variability in phenotype specificallynot requiring a strict amnestic representation. As opposed to the NINCDS-ARDRAguidelines, the criteria are more specific to AD and information on biomarkers andgenetics can optionally be included. DSM-5 was updated with the NIA-AA criteriain mind, and with its new category of NCDs it introduced mild NCD as a parallelto the NIA-AA’s MCI due to AD. In DSM-5, attributing NCD to AD still requires amemory deficit, and biomarkers do not play a role.

More evidence on the pathophysiology of AD and a better understanding of thedecades-long disease development process have led to the need to set diagnostic cri-teria for AD that do not necessitate full-blown dementia. Such criteria could be usedto identify prodromal AD, a disease state preceding dementia. The intention would beto diagnose a specific disease, and not to start with incident dementia and retroac-tively phenotype the dementia. That is, the idea would be to move away from thetraditional two-stage diagnosis. New criteria incorporate biomarker information, asubject introduced in more detail in the following chapters. An International Work-ing Group (IWG1; Dubois et al., 2007) revised the NINCDS-ADRDA criteria by con-centrating on AD-specific cognitive deficits accompanied by supporting biomarkerfindings indicative of AD disease progress. The criteria are research-focused, requireequipment for biomarker analysis, and are tuned to be more specific than earlier cri-teria. The IWG1 criteria for probable AD are summarized as follows:

Core set of criteria A: Presence of an episodic memory impairment that 1) is re-ported to have progressed gradually over a period of at least six months, 2) can beverified objectively by testing, and 3) can be the solitary symptom or can be asso-ciated with other cognitive deficits. At least one supportive feature associated withknown AD pathology is additionally required: B) a specific form of brain atrophy,C) biomarker evidence in the cerebrospinal fluid, D) specific changes in amyloid pro-tein neuroimaging, or E) AD autosomal mutation in the family. Exclusion criteriainclude early onset or prominent non-AD symptoms, focal neurological symptomsor early extrapyramidal symptoms, and other sufficiently severe neurological con-ditions. Diagnosis of definite AD is warranted by the IWG1 criteria if the clinicalevidence is supported by either histopathological findings or the patient is shown tohave an AD autosomal mutation.

The older criteria have been used both in research and in clinical practice overthe last 30–40 years. The newer criteria are more aimed at research. At least in Fin-

25

land, making use of other biomarker information than structural brain imaging hasbeen constrained to more challenging cases requiring detailed differential diagnoses(Seppala et al., 2013).

Better understanding of the disease and improved technology in imaging, for ex-ample, have emphasized new challenges in defining AD. The current debate revolvesaround defining AD as a biological disease entity with a certain pathological cascadeversus defining the disease in terms of clinical symptoms that also covers early stagesof the disease. The latter approach coincides with the approach of the more recent di-agnostic frameworks presented in this section, and the former relies more stronglyon detailed biomarker profiles introduced in the following section.

2.2.2 Pathophysiology

Recognition of AD as a discrete disease identity was coupled with finding distinctpathological lesions in the brain of the first patient to be diagnosed with Alzheimer’sdisease, Auguste D. Under the microscope Dr. Alois Alzheimer observed senileplaques and neurofibrillary tangles that he recognized as a separate entity from vas-cular lesions. The senile—or neuritic—plaques consist of extracellular aggregatedamyloid beta (Aβ) peptides and are indeed typical to AD. The neurofibrillary tanglesare formed by the aggregation of phosphorylated tau protein in microtubules insideneurons, a process which is not entirely specific to AD. Vascular lesions have laterbeen presumed not to be linked to AD itself, but to cause cognitive decline and de-mentia independently. Other microscopic findings include general loss of neuronsand amyloid angiopathy. (Erkinjuntti et al., 2015; Engelhardt and Grinberg, 2015;Bondi et al., 2017)

Scientific research into AD has as of yet not produced a consensus on the exactpathway for the occurrence of pathologic changes, or even if the recognized patho-logical changes are causative of the disease or if they are themselves downstreameffects. Recent epidemiological studies suggest Alzheimer’s dementia pathology tobe heterogenous with a good share of cases being attributable to non-AD-type patho-logical profiles (Boyle et al., 2019). Amyloid pathology is hypothesized as the first-mover process, and neuritic plaques can be found in the brain decades before the ap-pearance of first symptoms (Jack et al., 2013). The amyloid precursor protein (APP)is a membrane-bound protein found in neurons as well as other tissues. The purposeof APP is not fully understood. As the name suggests, the protein is best known forthe end products of its cleavage, namely Aβ peptides. APP is cleaved by β secretases(BACE1 and -2) and γ secretase. Aβ40 and Aβ42 are the most common resultingoligomers, and Aβ42 is prone to misfolding and thus implicated in AD pathology.Aβ peptides are water soluble and can be found in cerebrospinal fluid (CSF), urine,and plasma. With increasing age, these Aβ peptides are known to aggregate as dif-fuse plaques in the neocortex. In AD, however, plaque formation is distinctly char-acterized by dense plaque cores surrounded by a detritus of dead neurons withina more diffuse plaque. These are neuritic plaques. Astrocytes and microglial cellsare often associated with neuritic plaques indicating an inflammatory response. Aβ

26

is also known to accumulate in the walls of small arteries to cause cerebral amyloidangiopathy (CAA), a condition known to predispose patients towards bleeds andischemia. (Erkinjuntti et al., 2015)

Aβ oligomers are thought to have neurotoxic and proinflammatory effects. Theseeffects may be partially responsible for the hyperphosphorylation of tau, a proteintypically stabilizing the structure of microtubules in cells. Hyperphosphorylated tauis dysfunctional and aggregates pathologically in helical filaments inside the cell.These neurofibrillary tangles are insoluble and disturb the functioning of the cell re-sulting ultimately in neuronal loss. Tau and hyperphosphorylated tau are howeversoluble, and can be measured in CSF, for instance. (Erkinjuntti et al., 2015)

Markers of both amyloid and tau pathology are used to reach a more accurateAD diagnosis, as the purely clinical diagnostic criteria of the previous section onlywarrant a likely diagnosis. The sensitivity of a clinical diagnosis is in the range 71–87% (Beach et al., 2012). Tau pathology is thought to progress in a manner bettermatching the clinical presentation of AD. The Braak staging (Braak and Braak, 1991)is a six step staging classification describing the spread of tangles from the entorhinalcortex through the limbic system—including the hippocampus—to the associativeand primary visual cortices. The patient is thought to be symptomatic first at thelimbic stages III–IV, and stages V–VI are usually associated with clinical AD. Thepattern of neuritic plaque formation is different, starting usually from smaller areasof the cortex and spreading throughout the cortex into subcortical and subtentorialstructures. This spreading pattern corresponds less reliably to the clinical stage ofthe disease than the spread of tangles. Therefore Aβ pathology is commonly quan-tified using the CERAD criteria for neuropathology (Mirra et al., 1991) that simplyreport the frequency of neuritic plaques in the neocortex (class 0 for none and A–Cfor sparse–frequent). The CERAD frequency is sometimes adapted by consideringclinical data such as age. The two types of pathology are combined under the Na-tional Institute for Aging and Ronald and Nancy Reagan Institute of the Alzheimer’sAssociation (NIA-RIA) guidelines (NIA-RIA, 1997), which determine three stages ofAD probability based purely on pathological findings.

2.2.3 AD biomarkers and biomarker-based diagnosis

Several biomarkers for characteristic AD pathologies are in use today. In-vivo mark-ers for AD pathology can provide support in diagnosis making in conjunction withthe clinical presentation. The development of biomarkers is keenly ongoing with aimto identify AD in its earliest stages when AD pathology is present but no symptomshave yet appeared. The established biomarkers of AD are proxies for Aβ pathology,tau pathology, and general neurodegeneration. Whereas earlier these pathologiescould be assessed by analyzing CSF, nowadays imaging modalities have made it pos-sible to gauge the brain less invasively and gain topographical information. Positronemission tomography (PET) using common amyloid-binding ligands (e.g. Pittsburghcompound B, PIB) correlate well with post-mortem Aβ findings (Dubois et al., 2014).Amyloid PET is somewhat lacking in specificity in regard to AD, as a number of

27

subjects show amyloid-PET positivity with no symptoms of AD (Dubois et al., 2014).The focus on the wider AD disease course has led to the development of crite-

ria for the disease stages preceding the distinct cognitive deficits of AD dementia.Historically, MCI has covered conditions where subjects suffer mild subjective andobjective cognitive symptoms that do not affect ADL. The term is agnostic to etiol-ogy and does not imply any kind of progression of the impairment. Preclinical ADand prodromal AD (Dubois, 2000) are terms that are used to specifically describe thestages of AD that precede dementia: in the preclinical phase AD-specific pathologyexists, but no symptoms are present. In prodromal AD symptoms appear, but not atan intensity warranting diagnosis of dementia. Patients would typically be classifiedas having MCI. In prodromal AD the emphasis is on the biomarker profile.

Further development of definitions and criteria for presymptomatic AD (Duboisand Albert, 2004; Dubois et al., 2010; Sperling et al., 2011) is ongoing. Recently, theInternational Working Group criteria from 2014 (IWG2) published criteria for twoentities of preclinical AD (Dubois et al., 2014). The first criterion focuses on thosewho are asymptomatic at risk of AD with either Aβ on PET or with both Aβ andtau abnormalities in CSF. Here the priority of Aβ is evident, and imaging is seenas more reliable than CSF analysis. Second, presymptomatic AD is defined in termsof genetic susceptibility in the form of one of the three autosomal dominant genesor other proven genes. Notably, in the IWG2 criteria the biomarkers for Aβ and taurepresent diagnostic markers—that is, upstream stages of the disease process—andare preferred over downstream progression markers such as cortical atrophy or brainglucose metabolism. Additionally, the criteria enable the diagnosis of AD withoutrestrictions to the phenotype. To further clarify the terminology, in 2016 the termpreclinical AD was suggested to be defined in terms of particularly high AD risk withboth Aβ and tau pathology present, as opposed to asymptomatic-at-risk representinga state of lower risk with only one type of pathology present (Dubois et al., 2016).

The NIA-AA workgroup updated the 2011 criteria in 2018 with the aim to defineAD as a biological entity purely in terms of the pathologic disease progress (Jacket al., 2018). The definition relies on the biomarker status as defined by the amyloid,tau, and neurodegeneration (AT(N)) status. Table 1 summarizes this grouping ofbiomarkers. The CSF total tau is seen here as a marker of neurodegeneration ratherthan a marker of tauopathy as opposed to the IWG2 criteria. The framework definesan AD continuum in terms of the AT(N) profile by requiring a positive Aβ finding—apriority as with the IWG2 criteria—and letting T and N vary.

Some biomarkers are better suited for monitoring disease progression than fordiagnostics, as they represent downstream changes and lack specificity. Measure-ment of medial temporal atrophy (MTA) using magnetic resonance imaging (MRI) isa good marker for the development of AD dementia in prodromal AD, and longitudi-nal MRI measurements are good predictors of disease progression. Hypometabolismon PET is a good tool for differential diagnostics and the determination of AD thesubtype as well as a good estimate of the remaining brain function in AD patientswith high cognitive reserve. (Dubois et al., 2014)

28

The recent advances in biomarker research and clinical studies in cohorts rep-resenting pre-AD or very mild AD individuals have led to a discussion on how toconceptualize AD. Dubois et al. (2018) frame the staging of AD around symptoms,defining the preclinical stage of sporadic AD as asymptomatic at risk. Here the phasebefore the AD-threshold—symptoms—is an at-risk state, not a part of the disease.Jack and Vemuri (2018) on the other hand, with the AT(N) classification scheme aimto frame AD as a biological entity that takes its pathological course and only in theend manifests itself as a clinical syndrome. In this framework, the at-risk stage ofDubois et al. corresponds with ongoing AD in the preclinical stage.

2.3 VASCULAR COGNITIVE DISORDERS

2.3.1 Clinical presentation and diagnostic criteria

Impaired blood supply to the brain may lead to cognitive deficits and dementia, aspectrum of disorders called vascular cognitive impairment (VCI; Erkinjuntti andGauthier, 2009). Current guidelines identify several subtypes of vascular disordersbased on arterial anatomy and disease etiology. Disease presentation varies accord-ing to etiology: disease in the large vessels typically causes severe symptoms abrupt-ly, and cognitive deficits may be more or less apparent after treatment and/or reha-bilitation. Additionally smaller, initially subclinical, events may cause lesions, suchas white matter lesions, that eventually lead to gradual cognitive deficits. Risk fac-

Table 1: AD biomarkers in the NIA-AA 2018 guidelines.

Type ofpathology

Biomarker Biomarker positivefinding

Topographyincluded

Other conditions linkedto biomarker positivity

Aβ CSF Aβ42 Low concentration No HIV encephalitis,multiple system atrophy

Amyloid PET High ligand uptake Yes Acute traumatic braininjury

Tau CSF p-tau High concentration No —

Tau PET High ligand uptake Yes Unknown

Neurode-generation

CSF total tau High concentration No Acute traumatic braininjury, stroke, CJD

Metabolic PET Low uptake inAD-typical pattern

Yes CVD, corticobasaldegeneration, PPA

Structural MRI Expert assessmentof atrophy

Yes CVD, epilepsy, anoxia,hippocampal sclerosis

Table adapted from Jack et al. (2016, 2018). Key: Aβ amyloid β protein, CJD Creutzfeldt–Jakobdisease, CSF cerebrospinal fluid, CVD cerebrovascular disease, HIV human immunodeficiencyvirus, MRI magnetic resonance imaging, PET positron emission tomography, p-tau phospho-rylated tau protein, PPA primary progressive aphasia.

29

tors for vascular health are also risk factors for VCI via their effect on arteries andpossibly also through other mechanisms (Erkinjuntti et al., 2015). These risk factorsare discussed in more detail in section 2.6.

Vascular dementia (VaD) has been recognized since the 1960s as a disease entityand nowadays it is seen as a part of the VCI spectrum. The International Society forVascular Behavioral and Cognitive Disorders (VASCOG; Sachdev et al., 2014) pro-posed improvements on the then current guidelines on VaD, resulting in expandednew guidelines for vascular cognitive disorders (VCD). The four sets of commonlyused criteria for VaD, The National Institute of Neurological Disorders and StrokeAssociation Internationale pour la Recherche et l’Enseignement en Neurosciences(NINDS-AIREN; Roman et al., 1993), the State of California Alzheimer’s Disease Di-agnostic and Treatment Centers (ADDTC; Chui et al., 1992), the DSM-IV (AmericanPsychiatric Association, 1994), and the ICD-10 (World Health Organization, 1993) cri-teria, start off with a classical notion of dementia with memory impairment, diagnosisbeing further specified in various ways by stroke history details, neuroimaging, andthe specific features of cognitive impairment. The ADDTC criteria differ from theothers somewhat. The 2014 VASCOG criteria improve on these criteria by modifyingthe cognitive domain criteria to better take into account the frontal-executive-typedeficits over memory deficits, recognize pre-dementia-level cognitive disability, de-fine impairment due to mixed etiology, and define the types of vascular pathologymore broadly.

According to the VASCOG criteria a diagnosis requires one or more of the fol-lowing cognitive domains to be affected: attention and processing speed, frontal-executive function, learning and memory, language, visuoconstructional-perceptu-al ability, body conception, and social cognition. The deficit is defined as mild ormajor—corresponding to VaD—based on objective domain measurements and on thedisability caused by the impairment. The diagnosis also requires evidence of signifi-cant cerebrovascular disease. Neuroimaging is emphasized in determining brain le-sions and to rule out other disorders. Imaging results are to be interpreted in light ofthe clinical presentation and the temporal development of symptoms, and the resultis a diagnosis of probable VCD. Sufficient evidence from a stroke incident or a clearfinding from a neurological examination are permitted as substitutes when imag-ing is not available thus warranting a diagnosis of possible VCD. In VCDs, cognitivesymptoms often present more severely in the acute phase and symptoms may be alle-viated later. A period of 3 months is set as a threshold value for persistent symptoms.The rate of progression and fluctuation of symptoms may vary due to the specificetiology, e.g. small vessel disease may present with fluctuating symptoms due to sev-eral successive events. Exclusion criteria include features such as memory deficit asthe early leading cognitive impairment as well as Parkinsonism.

It is common to find multiple processes impairing cognition at the same time.Alzheimer’s pathology is often accompanied with vascular changes, and vascularpathology decreases the threshold for clinical Alzheimer’s disease. The VASCOGguidelines recognize this clinical challenge and encourage choosing the most promi-

30

nent diagnosis, recognizing the uncertainty of the diagnosis, and acknowledgingother contributing pathologies.

2.3.2 Pathophysiology

Large vessel disease affects the larger arteries at the cortical level and is typically causedby an atherosclerotic plaque or a cardiac embolus. This leads to a relatively large sin-gle cortical infarction or to several smaller downstream infarctions. In many casesfocal neurological deficits are apparent alongside cognitive symptoms. Small vesseldisease refers to the stenosis of smaller perforating arteries in the brain parenchyma.Resulting ischemic changes take the form of lacunar infarcts, white matter lesions,perivascular space dilatation, microinfarcts, and microhemorrhages. Typically, smallvessel disease impairs executive functioning and the speed of processing. Depressionand gait disturbances can occur. Progression is typically more gradual and focal neu-rological symptoms are less frequent. Arterial wall defects lead to intracerebral orsubarachnoid hemorrhages. Prolonged hypoperfusion can lead to sclerosis, typicallyof the hippocampus, or take the form of laminar cortical sclerosis. (Erkinjuntti et al.,2015)

No biomarker measured in CSF is specific to VCI. High total tau is indicative ofneuronal damage that can be associated with VaD. Brain lesions typical to VaD canbe determined using MRI. Bleeds of different calibres are visible, as is thinning ofcortical grey matter in small vessel disease. White matter lesions seen on an MRI aretypically quantified using the Fazekas scale (Fazekas et al., 1987).

Other angiopathies known to affect arterial function as listed by Sachdev et al.(2014) include e.g. vasculitis, hereditary angiopathies, berry aneurysms, and CAA.In CAA accumulation of Aβ in the walls of small vessels often leads to microhemor-rhages or microinfarctions. An in-vivo diagnosis of CAA is based on the localizationof microhemorrhages in the cortical and subcortical regions on MRI. Amyloid-PETimaging does not differentiate between Aβ in brain tissue and in the arteries (Gore-lick et al., 2011). CAA is a very common vascular pathology found in AD patients(Smith and Greenberg, 2009) and the severities of AD and CAA pathologies are sig-nificantly correlated (Attems et al., 2005). Some studies indicate that a higher CAAload may impair cognition independent of other pathologies (Brenowitz et al., 2015).More specifically, CAA on neuroimaging has been associated with cognitive declinebefore the first clinically presenting intracerebral hemorrhage (Banerjee et al., 2018).

2.4 LEWY BODY DEMENTIAS

Lewy body dementias comprise of the disease identities dementia with Lewy bodies(DLB) and Parkinson’s disease dementia (PD-D), conditions that account for a significantportion of dementia cases in older age groups. Dementia prevalence increases withtime past the Parkinson’s disease diagnosis reaching 50% at 10 years post-diagnosis.The prevalence of DLB in patients with a dementia diagnosis is estimated to be upto 23%. It is estimated that DLB is an underdiagnosed condition, probably due to

31

difficulties in differentiating between it and AD. (Walker et al., 2015)According to the 2005 DLB Consortium criteria (McKeith et al., 2005), in addition

to dementia-level functional impairment, a DLB diagnosis requires the presence ofcore DLB features: fluctuating cognition, recurrent visual hallucinations, and spon-taneous parkinsonism. Supporting features include disturbances in sleep structure,changes on brain metabolism imaging, and specific changes on electroencephalog-raphy. These criteria have been proven to be specific but not very sensitive (Walkeret al., 2015). DSM-V defines major neurocognitive disorders with Lewy bodies similarly.PD-D is diagnosed according to criteria published in 2007 (Emre et al., 2007). Inaddition to established Parkinson’s disease the criteria require impairment in atten-tion, executive function, visuospatial function, or free recall; and supporting featuressuch as apathy, depression, and delusions are acknowledged. To differentiate DLBfrom PD-D, dementia should not present more than one year after the start of Parkin-sonism. In general, cognitive deficits in Lewy body dementias are characterized byimpaired executive function and visuospatial capabilities in contrast to AD-dementiaepisodic memory impairment. (Erkinjuntti et al., 2015)

α-synuclein, a protein functional in presynaptic terminals in the brain, is the pri-mary component of Lewy bodies. The pathological mechanism of the formation ofthese bodies is unclear similarly to the accumulation of Aβ in AD. No clear picturehas emerged on risk factors of α-synuclein accumulation. The bodies are associatedwith neuronal dysfunction in their vicinity, but whether the formation of these inclu-sions has a protective effect or if they represent upstream pathological processes isunknown. In PD-D, α-synuclein pathology is thought to be more strongly associatedwith dementia than in DLB, where mixed etiology with Aβ is thought to play a sig-nificant role. The severity of dementia in PD-D and DLB is associated with the levelof AD-type pathology present, whereas there is little evidence of concurrent vascularpathology having an effect. There are known autosomal dominant mutations thatlead to Lewy body dementias or Parkinson’s disease, and there is some evidence ofadditional familial clustering of DLB not explained by them. The APOE ε4 allele isassociated with elevated risk of Lewy body dementias but not to the extent of AD.(Walker et al., 2015)

Imaging of α-synuclein pathology is currently not possible, but single photonemission computer tomography and metabolic PET for hypoperfusion and hypo-metabolism have shown distinct occipital-lobe patterns in Lewy body dementias thatare not seen in AD (Minoshima et al., 2001). These tests are recognized as support-ive features in the 2005 DLB criteria, and they can aid in the differential diagnosisof α-synucleinopathies and AD. α-synuclein levels in cerebrospinal fluid have beenshown to some extent discriminate between dementia with Lewy bodies and AD,whereas CSF Aβ has been shown not to (Walker et al., 2015).

32

2.5 OTHER DEMENTIAS

Frontotemporal dementia (FTD) refers to a variety of syndromes which affect thefrontal and temporal neocortices. The clinical phenotype varies according to theexact region that is affected, and the type of neuropathology also varies. Gener-ally, the syndromes are characterized by behavioral changes, executive dysfunction,and difficulties in language. The clinical presentation can be similar to psychiatricconditions, and a differential diagnosis may be difficult. Clinical subtypes includebehavioural-variant FTD with a prefrontally and temporally dominated pathology, andprimary progressive aphasia of the non-fluent variant with a left-frontotemporal domi-nated pathology and of the semantic variant with temporally dominated pathology.As the disease progresses, the symptoms of the subtypes tend to converge as thepathology spreads. Neuropathologically, three types of pathology are recognized:30–50% of cases are tau dominated, in 50% of the cases TAR DNA-binding proteinwith a molecular weight 43 kDa (TDP-43) is found in the form of intracellular inclu-sions, and about 10% show fused-in-sarcoma protein inclusions. Structural MRI andmetabolic PET show regional cortical atrophy and metabolism, and amyloid PET isused to differentiate FTD and AD. FTD has a higher relative incidence in youngerage groups compared to other types of dementia, and a number of risk genes havebeen recognized. Most of the inherited disease cases are due to the genes C9orf72and GRN, resulting in TDP proteinopathy, and MAPT, resulting in tauopathy. (Banget al., 2015; Erkinjuntti et al., 2015)

Hippocampal sclerosis (HS) is a somewhat unspecific pathological finding thathas a relatively high prevalence in the very old. Historically, in cases where HS hadclearly dominated typical AD pathology and an amnestic impairment was present,the term hippocampal sclerosis dementia was used (Cykowski et al., 2017). More re-cently HS has been strongly associated with cortical TDP-43 accumulation, and hasbeen shown to also commonly present with AD and LBD pathology (Nag et al., 2015).HS without other neuropathology seems to be rare (Kero et al., 2018), and HS with-out TDP-43 pathology seems to not be associated with cognitive decline (Nag et al.,2015). Discussion is ongoing on how to conceptualize HS-related syndromes: HS de-mentia has been suggested to be an amnestic variant of frontotemporal degenerationdue to similarities in pathology (Onyike et al., 2013). To differentiate from other de-generative disorders, Nelson et al. (2016) suggest the term cerebral age-related TDP-43with sclerosis (CARTS) to be used independently for this type of pathology. Cykowskiet al. (2017) suggest further to differentiate between CARTS and AD or FTD withconcomitant TDP-43 pathology.

2.6 RISK FACTORS FOR DEMENTIA AND BRAIN PATHOLOGY

Epidemiological studies have revealed risk factors for cognitive decline and demen-tia, and for certain specific disease etiologies. In general, pathological processes lead-ing to dementia are still largely unclear, which is reflected in the difficulties of linkingrisk factors to a specific pathological process, or to dementia or declining cognition

33

more broadly. Dementia shares well-known risk factors with e.g. cardiovascular dis-ease, but the exact underlying mechanisms are not yet fully known.

For dementia, several risk factors have been recognized. Incidence of all-causedementia increases exponentially with age (Jorm and Jolley, 1998). The incidencepattern for AD is similar, but for VaD there is greater variability in different popula-tions. The incidence of AD in very old age is higher in women, and the VaD incidenceis higher in younger men (Jorm and Jolley, 1998). There is clear evidence of familialsusceptibility to dementia (Loy et al., 2014), partly due to the effect of specific riskgenes that have been identified.

Age and genes are immutable personal characteristics that cannot be influenced.Research into preventable risk factors has produced a large number of tentative riskfactors related to somatic and mental health, socioeconomic status, and lifestyle. Thedata are mostly observational, but randomized controlled trials do exists for some,like blood pressure and hypercholesterolemia. However, in many cases the resultsare mixed. Table 2 lists a number of potentially preventable risk factors for dementiaand gives an estimate made by the Alzheimer’s Association on the level of evidenceconcerning the association with dementia (Baumgart et al., 2015). In 2017, the Na-tional Academies of Sciences, Engineering, and Medicine and Health released a re-port outlining recommendations on interventions for some well-established risk fac-tors and also outlined research priorities for risk factors with insufficient data. Threeinterventions were indicated as promising based on the current status of evidence:cognitive training, blood pressure control in midlife, and increasing physical activity.The National Academies of Sciences assessments have also been included in Table 2.

2.6.1 Education

A person’s history of education and cognitive exertion seems to be associated withthe timing and rapidity of cognitive decline in advanced age. A low level of educationhas been linked to AD-dementia risk in a comprehensive meta-analysis (Caamano-Isorna et al., 2006). A later systematic review by Meng and D’Arcy (2012) confirmedthis for the risk of AD-dementia, VaD, and unspecified dementia. In that review, mostof the substudies reporting on brain pathology also found more severe pathologicalchanges in individuals with a higher education. These findings are in line with thecognitive reserve hypothesis stating that high education and other forms of cognitivetraining lead to higher resilience of the brain against pathological lesions and thatthis elevates the threshold level at which cognition is impaired.

There is some evidence to suggest that cognitive reserve may have a slowing effecton the accumulation of Aβ itself (Lo et al., 2013; Yasuno et al., 2015). Studies have alsoshown an association between altered brain structure and certain surrogate markersof cognitive reserve (Xu et al., 2015) and a higher cognitive reserve has also beenlinked to changes seen in functional MRI (Anthony and Lin, 2018).

34

Table 2: Dementia risk factors and protective factors. Alzheimer’s Association (AA) assess-ment of level of evidence on association (Baumgart et al., 2015) and National Academies ofSciences, Engineering, and Medicine and Health (NAS) recommendation on intervention (Na-tional Academies of Sciences, Engineering, and Medicine and Health, 2017).

Effect Factor Level of evidenceon association(AA assessment)

Recommendation forintervention(NAS assessment)

Riskfactors

Traumatic brain injury Strong —Midlife obesity Moderate —Midlife hypertension Moderate Intervention supported†

Current smoking Moderate —Diabetes Moderate Priority for research‡

History of depression Unclear Priority for research‡

Sleep disturbances Unclear Priority for research‡

Hyperlipidemia Unclear Priority for research‡

Vitamin B12 deficiency1 — Priority for research‡

Hearing loss2 — —Particulate air pollutants2 — —

Protectivefactors

Years of formal education Strong —Physical activity Moderate Intervention supported†

Mediterranean diet Lower Priority for research‡

Cognitive training Lower Intervention supported†

Moderate alcohol consumption Unclear —Social engagement Unclear Priority for research‡

†: Evidence of intervention to prevent Alzheimer’s-type cognitive decline is encouraging butinconclusive. Recommendation based on evidence, neurobiological plausibility, and benefitsto general health. ‡: Insufficient evidence to recommend intervention, additional researchneeded. —: No statement made on risk factor. For additional references see 1: (Ford andAlmeida, 2019), 2: (Livingston et al., 2017), and 3: (Baumgart et al., 2015).

35

2.6.2 Risk genes and causal mutations

Apolipoprotein E (APOE) gene polymorphism has been linked to the incidence of AD(Saunders et al., 1993). APOE in peripheral tissue takes part in the uptake of lipopro-teins such as high-density lipoprotein (HDL) and low-density lipoprotein (LDL) inthe liver. APOE does not cross the blood–brain barrier (BBB). APOE found in thebrain and the CSF is produced by the brain parenchymal cells, but its function isunclear. Three major alleles can be found in the population: the major type is ε3,while ε4 is the second most common, and ε2 is the most infrequent. The ε4 allele isthe risk allele. A heterozygous ε4 genotype increases the risk of AD threefold and ahomozygous genotype 8–12-fold, and ε4-carrying AD patients are typically youngerthan noncarriers (Alzheimer’s Association, 2016). ε2 may be AD-protective (Corderet al., 1994). Frequencies of the risk allele ε4 in populations around the world rangebetween 8% and 31% (Eichner et al., 2002). APOE is an important consideration indementia research, and the effect of its polymorphism is routinely taken into accountin the study of dementia risk factors. The other known risk genes are known to beassociated with the metabolism of APP and lipids, and with the immune system (Erk-injuntti et al., 2015). An AD-protective variant of the APP gene has also been found(Jonsson et al., 2012). The protective effect is mediated by altered β secretase cleav-age.

For non-sporadic AD, causal gene mutations have been found in the APP, PSEN1and PSEN2 genes. Changes in the APP amino acid sequence increase the proportionaloutput frequency of the Aβ42 subtype, and so do mutations in the γ secretase codingPSEN1 and PSEN2 genes. Down syndrome patients frequently show neuropathologysimilar to AD patients (Mann, 1988), possibly due to them having three alleles of APP.

2.6.3 Cardiovascular risk factors

There is an established link between the health of the cardiovascular system and therisk factors affecting it, and cognitive decline, dementia, and some of the specificunderlying diseases (Power et al., 2011).

Hypertension

The association between hypertension and dementia has long been studied in cross-sectional and longitudinal studies. Age-specific effects have been found, but the re-sults have to an extent been mixed. Some studies report on dementia overall, whilesome report separate results for AD and VaD, and the data is either measured or self-reported. A meta-analysis by Power et al. (2011) investigated populations with base-line mean ages 50-74 and found no association between the reported hypertensionhistory or current hypertension and AD. Three studies investigated mid-life (<65years) measured hypertension in association with AD and only one of those reporteda positive association between highly elevated mid-life systolic blood pressure (SBP)and incident AD (Kivipelto et al., 2001b). Another study included in the review foundsuggestive evidence for a nonlinear effect where both low and high SBP indicated

36

higher risk of AD (Launer et al., 2000). An adverse effect of increased mid-life dias-tolic hypertension on incident AD was suggested by the pooled analysis, althoughno single study reached statistical significance. In their study, Launer et al. (2000)verified the effect of high mid-life SBP on higher incidence of VaD, whereas with di-astolic blood pressure (DBP) there was no effect. A systematic review by Sharp et al.(2011) looking specifically at hypertension and VaD found a history of hypertensionand measured hypertension to be clearly associated with both a higher prevalenceand higher incidence. For all-cause dementia, a recent study found hypertension ata relatively low cut-off value of SBP>130 mmHg at 50 years of age to be associatedwith increased risk and this was independent of other cardiovascular diseases (Abellet al., 2018).

For late-life-measured BP, three studies with measured values in the meta-analysisby Power et al. (2011) reported a consistent but nonsignificant protective effect of highDBP. SBP measures indicated a similar effect but less consistently. A 9-year-follow-up of over-85-year-olds (Rastas et al., 2010) reported a protective effect of history ofhypertension on incident dementia, however measured baseline BP did not showthis association. A 3-year follow-up study of subjects over the age of 65 by Haydenet al. (2006) reported a positive association between hypertension and incident VaDin women and a negative association between hypertension and AD in both genders.More recently, Corrada et al. (2017) found the self-reported onset of hypertensiononly after the age of 80 and 90 to be associated with decreasing incidence rates ofdementia compared to controls with no hypertension.

Walker et al. (2019) found two mid-life–late-life BP profiles to be associated withhigh dementia risk: both sustained hypertension from mid-life to late-life and thedevelopment of mid-life to late-life hypotension indicated elevated risk.

Observational studies have shown the use of antihypertensive drugs to reducethe risk of both VaD and AD (Rouch et al., 2015). Interventions treating elevated BPhave shown a positive effect on cognitive decline, but not conclusively on demen-tia. Less than half of 11 randomized trials analysed by Rouch et al. (2015) found asignificant effect on cognitive decline or dementia with a maximal follow-up time of4.5 years within the trials. A large SPRINT MIND trial recently showed a benefit ofaggressive BP control over traditional BP targets in terms of incidence of MCI andincidence of either dementia or MCI, but was only suggestive for lower incidence ofdementia (SPRINT MIND Investigators for the SPRINT Research Group, 2019). Inthe same study, more aggressive treatment was associated with a lesser increase ofwhite matter (WM) lesion volume and greater a decrease in brain volume (SPRINTMIND Investigators for the SPRINT Research Group, 2019).

There is other supporting neuropathological evidence for the association of brainlesions with BP. A longitudinal study by Petrovitch et al. (2000) on the effect of midlifehypertension on brain pathology during a 36 year follow up provided insights intothe underlying brain changes: in those with elevated mid-life SBP, more Aβ plaqueswere found in the neocortex and the hippocampus, and the brain weight was alsolower. Elevated DBP was associated with increased counts of neurofibrillary tan-

37

gles in the hippocampus. A more recent cross-sectional study by Jeon et al. (2019)in a stratified analysis showed differences in the association of hypertension withbrain pathology in cognitively normal subjects and in those with AD-dementia: hy-pertension in cognitively normal APOE-ε4 noncarriers was associated with a lowercortical thickness in AD signature regions, but not Aβ accumulation, whereas ε4 car-riers with hypertension had a higher rate of Aβ accumulation. Among AD-dementiasubjects, hypertension was associated with lower Aβ deposition irrespective of theAPOE genotype.

Kennelly et al. (2009) summarize the mechanisms by which elevated BP—in com-bination with other cardiovascular risk factors—causes VaD-related pathology. Dam-age to the arterial wall in the form of reactive thickening of the media and devel-opment of atheromatous material predispose the vessel for local thrombi. Cardiacfailure and atrial fibrillation are cardiac outcomes of hypertension that may lead toembolus formation and infarctions.

Hypercholesterolemia

Changes in cholesterol metabolism have been associated with all-cause dementia andespecially with AD-dementia. High midlife total cholesterol has consistently beenassociated with a higher rate of incident all-cause dementia in later life in systematicreviews (Kivipelto and Solomon, 2006; Anstey et al., 2017). For late-life cholesterol,these reviews showed no association with incident dementia. Anstey et al. (2008)analysed studies that established normal cognition at the baseline and had availabledata on dementia etiology at the follow up. Although such studies consistently founda link between mid-life high cholesterol and AD-dementia specifically, no associationwith VaD was found. A later study did find an association with both AD and VaD ina 30-year follow up (Solomon et al., 2009). A decline in cholesterol levels from midlifeinto old age has been associated with higher AD rates (Anstey et al., 2017). Studieslooking at other dyslipidemias do not form a uniform body of evidence. Some studieshave reported not finding an association between cognitive decline/dementia or hightriglycerides and high density lipoprotein (Anstey et al., 2017).

Four trials targeting individuals with high cholesterol found no effect on cogni-tive performance or incident dementia in follow-ups ranging from 6 months to 5years (National Academies of Sciences, Engineering, and Medicine and Health, 2017).Combination therapy where statins are accompanied by drugs inhibiting gut choles-terol uptake did not produce better results. Geifman et al. (2017) found evidence forpotential intervention benefits in homozygous APOE ε4 carriers in subgroup analy-ses.

Cholesterol metabolism is essential for brain function, and brain cholesterol hasbeen linked to neurodegenerative diseases including AD (Bjorkhem, 2006). APOE isclosely related to cholesterol metabolism. Although the brain and peripheral choles-terol pools are separated by the blood-brain barrier, they can interact via metabolitessuch as oxysterols (Bjorkhem, 2006). Hypercholesterolemia may also increase therisk of dementia through the vascular pathway, i.e. increased risk of cardio- and cere-

38

brovascular disease.

Obesity

Obesity has a high comorbidity with other cardiovascular risk factors, most impor-tantly metabolic syndrome, type 2 diabetes, and hypertension. Meta-analyses andsystematic reviews have reported on the effect of body composition throughout thelife course. Midlife overweight (body mass index, BMI, in kg/m2 from 25.0–27.5 to30.0) and/or obesity (BMI >30) has been found to be associated with all-cause de-mentia (Anstey et al., 2011), AD (Beydoun et al., 2008; Profenno et al., 2010; Ansteyet al., 2011), and VaD (Anstey et al., 2011). A stable BMI into old age was not asso-ciated with dementia (Anstey et al., 2011), but weight gain seemed to be (Beydounet al., 2008). Beydoun et al. (2008) found associations with AD and VaD to be strongerwith longer follow-up times and younger baseline populations. Midlife underweighthas been associated with all-cause dementia (Beydoun et al., 2008) and AD (Ansteyet al., 2011). Old-age overweight has been found to be associated with a lower risk ofdementia (Baumgart et al., 2015). A large meta-analysis of BMI and incident demen-tia over different time periods demonstrated this reversion of association: BMI wasshown to be a risk factor over decades-long follow-up periods and protective overperiods of less than ten years (Kivimaki et al., 2018). The authors hypothesize thatover shorter periods weight loss may be caused by preclinical dementia showing apattern of reversed causality.

Trials investigating the effect of increased physical activity on dementia have beenpromising (National Academies of Sciences, Engineering, and Medicine and Health,2017), but the extent to which the effect is due to body weight is unclear. Other po-tential mechanisms include improved insulin sensitivity, reduction in hypertensionor high cholesterol, or neurological effects (Livingston et al., 2017).

Cardiovascular conditions

Specific cardiac conditions have been studied in association with cognitive declineand dementia. Atrial fibrillation (AF) is a risk factor for dementia not only throughits association with stroke (5-fold risk of stroke in AF), but also independent of priorstroke (Aldrugh et al., 2017). Non-stroke hypothesized causal explanations includecerebral hypoperfusion and possibly associated altered Aβ metabolism, vascular in-flammation, small vessel disease, and brain atrophy. Similar mechanisms may under-lie the association between dementia and heart failure, a condition often secondaryto a coronary conditions like coronary heart disease (CHD) or AF. Systematic reviewshave confirmed the positive association between heart failure and cognitive impair-ment (Cannon et al., 2017) and dementia (Wolters et al., 2018), and also that betweenCHD and dementia (Wolters et al., 2018).

39

Cardiorespiratory fitness

Overall cardiorespiratory fitness (CRF) as measured by the maximum oxygen con-sumption has been associated with better measured cognition in a cross-sectionalanalysis (Freudenberger et al., 2016) as well as in a longitudinal follow-up setting(Pentikainen et al., 2019). A study by Schultz et al. (2015) was able to link this effectto AD pathology in that they found better CRF to be protective of the harmful effectsof Aβ accumulation. CRF—an aggregate measure—has been associated with hyperc-holesterolemia, impaired fasting glucose, diabetes mellitus, hypertension, and a highBMI (Erez et al., 2015). The association between CRF and cardiac outcomes duringfollow-up was shown to be mediated by hypercholesterolemia, diabetes mellitus, andobesity.

2.6.4 Insulin resistance and diabetes

Diabetes mellitus (DM) is a growing problem in modern societies, especially typetwo diabetes (DM2) with its overall prevalence rising due to an aging populationand changing lifestyles. The pattern is similar to dementia in terms of the agingpopulation. Additionally, the conditions share etiological features. The APOE ε4allele is a known risk factor of AD, and the APOE gene is a regulator of glucoseand lipid metabolism (Cheng et al., 2012). Two meta-analyses have confirmed higherdementia incidence rates to be associated with DM (Cheng et al., 2012; Gudala et al.,2013). The effect was reported for all-type dementia, AD-dementia, and VaD, but therelative risk was clearly higher for VaD in both studies. Micro- and macrovasculardiseases are well-known complications of DM, and thus the association with VaDis understandable. The effect was not mediated by APOE status. The mechanismslinking DM and AD are still unclear. Hypotheses include vascular and metabolicprocesses including insulin resistance, but no definitive link to disease progressionor pathology has been made (Gudala et al., 2013). Ahtiluoto et al. (2010) found olderindividuals with DM to have higher dementia incidence rates, and in autopsy to havelower levels of Aβ and tau pathology and more vascular pathology. An analysis byMoran et al. (2015) across different diagnostic groups found DM2 to intensify taupathology but not Aβ. Roberts et al. (2014) linked DM diagnosis with AD-type brainhypometabolism patterns but found no association with AD-type Aβ accumulation.

Insulin resistance in peripheral tissue is a hallmark of DM2. In recent years re-search has been done on insulin resistance in the periphery and also in the centralnervous system in relation to neurodegeneration. Peripheral insulin resistance (IR)is typically quantified using the homeostasis model assessment for insulin resistance(HOMA-IR=[Insulin] · [Glucose] · constant) index value. For brain IR, a new tentativeblood biomarker has been suggested (Kapogiannis et al., 2015). In prediabetic anddiabetic subjects higher peripheral IR has been linked to similar brain hypometa-bolism patterns on PET as seen in AD patients (Baker et al., 2011), a result which isanalogous to findings for DM. No association has been found between late-life IRin cognitively healthy elderly and CSF Aβ (Laws et al., 2017) or Aβ on PET (Ekblad

40

et al., 2018). There does not seem to be an Aβ association in MCI or AD subjects either(Laws et al., 2017). Results in younger populations have been mixed (Willette et al.,2015; Westwood et al., 2017; Ekblad et al., 2018). An association between long-lastingIR and neurodegenerative changes on MRI—including hippocampal atrophy—hasbeen reported (Korf et al., 2006).

Association with brain insulin metabolism

Brain glucose metabolism is regulated in part by insulin-independent glucose trans-porters at the BBB and also by insulin-dependent transporters at the BBB and inplasma membranes of parenchymal brain cells. It is nowadays known that insulinplays a role in brain metabolism and signaling, whereas before the brain was thoughtto be indifferent to insulin signaling. Insulin presents itself manyfold in the centralnervous system: peripheral insulin is transported through the BBB, there are insulin-activated signaling pathways through the BBB, and there is endogenous insulin pro-duction in certain regions of the brain. Insulin receptors of the BBB are known todecrease in number with aging and long-term blood hyperinsulinemia. It is hypothe-sized, that constant peripheral IR and associated hyperinsulinemia are linked to a de-crease in brain insulin-dependent glucose intake. Furthermore, the pattern of insulintransporter types varies by brain region and some regions may be more dependenton insulin-dependent glucose intake. This may make these regions more sensitiveto other pathological insults such as those seen in AD. The reason for and mecha-nism of brain insulin production and uptake are still unclear, but it is hypothesizedthat insulin signaling might be linked to neuroprotective mechanisms. Brain insulinresistance may be linked to brain degradation through these mechanisms. Anotherhypothesis suggests brain IR to promote oxidative stress, possibly a catalyst of Aβ

and tau pathology. Oxidative stress is also linked to metabolic syndrome and obesity,which are upstream stages of the IR–DM progression. (Diehl et al., 2017)

Some insight into the interplay between insulin and Aβ has been gained in mice.In a healthy brain insulin has been shown to promote amyloid clearing. Aβ seemsto suppress insulin receptor levels as well as interfere with insulin receptor functionthus downregulating the effect of insulin in the brain and resulting in lower Aβ clear-ance. Furthermore, Aβ and insulin are both cleaved by the insulin-degrading enzyme(IDE). Hyperinsulinemia may thus lead to lower levels of Aβ cleavage. IDE also ap-parently only cleaves monomeric Aβ. It has also been confirmed that in APOE ε4positive AD patients hippocampal IDE levels are lower than in controls. (Diehl et al.,2017)

2.6.5 Lifestyle

Several lifestyle factors seem to be associated with dementia. Smoking in old age hasbeen linked to incident dementia according to several studies (Baumgart et al., 2015),and mid-life heavy smoking was a strong predictor of late-life dementia according toRusanen et al. (2011). Meta-analyses have found higher physical activity to be protec-tive against cognitive decline (Sofi et al., 2011) and also protective against dementia

41

and AD (Hamer and Chida, 2009). A more recent review found leisure time physi-cal activity to be more important than work-related activity in terms of AD incidence(Stephen et al., 2017a). There is evidence from a meta-analysis of light to moderate al-cohol consumption being protective against all-cause dementia, AD, and VaD, whencompared to no consumption (Anstey et al., 2009), but these results may partly bedue to selection bias and due to the fact that most studies do no differentiate betweenabstainers and persons who have quit drinking.

Some nutrients and food groups have been associated with dementia, althoughevidence is weaker than for some other risk factors modalities (Baumgart et al., 2015).In case of cognitive decline more broadly, a Mediterranean diet as a dietary patternand B vitamins, some antioxidants, vitamin D, and unsaturated fatty acids as specificnutrients have been associated with a protective effect on cognition in many studies(Scarmeas et al., 2018).

2.6.6 Psychosocial

Depression is a comorbid state related to dementia, and depression is associatedwith a two-fold prevalence of dementia in old age (Cherbuin et al., 2015). Studyof the causality of the two is difficult. However, depression in midlife has beenassociated with increased dementia incidence in late life supporting the view thatdepression might be a preventable risk factor (Byers and Yaffe, 2011). A feeling ofhopelessness—a very common symptom of depression—in midlife also had a simi-lar association (Hakansson et al., 2015). Depression may be a result of minor damagedue to cerebrovascular disease coinciding with cognitive impairment of the vasculartype. Depression and dementia are also linked through several risk factors, such asphysical inactivity, metabolic syndrome, and low-grade inflammation. As for AD-related pathology, depression is linked to elevated cortisol levels, and cortisol mayinduce atrophy of the hippocampus. Additionally, AD patients with depression havebeen reported to have higher Aβ accumulation in the hippocampus compared withnondepressed patients possibly due to increased cortisol. (Byers and Yaffe, 2011)

The CSF Aβ profile of older adults with depression resembled that of AD patientsin a meta-analysis (Nascimento et al., 2015). There is no good-quality data on theeffects of treatment of depression on the dementia incidence. Observational studieswith very short follow-up times have shown both improvement and impairment ofcognition (National Academies of Sciences, Engineering, and Medicine and Health,2017).

Low social participation, loneliness, and infrequent social contacts have been asso-ciated with impaired cognition in a meta-analysis (Kuiper et al., 2015). A higher levelof social activity has been suggested to be protective, but no prevention study datais available for the isolated effect of improved social engagement on cognition anddementia (National Academies of Sciences, Engineering, and Medicine and Health,2017).

42

Table 3: Examples of risk factor combinations in midlife and late life modulating the risk ofdementia.

Potentiating combinations Attenuating combinations

Midlifeeffects

– High alcohol consumption, smoking,low physical activity and saturated-fatintake have higher effect in APOE ε4 car-riers.

– Concurrent hypertension, obesity, highcholesterol, and low physical activity alladd to risk independently.

– Education somewhat mitigatesthe risk increase due to APOE ε4.

– Physical activity reduces the riskdue to APOE ε4.

– Risk due to low education is af-fected by complexity of occupa-tional activity.

Late-lifeeffects

– Chronic heart failure, low pulse pres-sure, and low DBP contribute to brainhypoperfusion and higher risk.

– High SBP, DM or prediabetes, and strokeindicate atherosclerosis/vascular dam-age and higher risk of dementia.

– Risk due to APOE ε4 is mitigatedby leisure time activities and lackof vascular risk factors.

Table adapted from Solomon et al. (2014a). Key: APOE apolipoprotein E, DBP and SBP dias-tolic and systolic blood pressure, DM diabetes mellitus.

2.7 A MULTIMODAL APPROACH TO DEMENTIA RISK MANAGE-MENT

The interactions between the aforementioned risk factors is an important field ofstudy, especially with future preventive interventions in mind. Prevention trials tar-geting a single risk factor have shown no clear benefits in terms of dementia as aprimary outcome, nevertheless, a positive effect on cognition has been seen in thecase of BP, for example. A multimodal approach to dementia risk management maybe needed in the future. Solomon et al. (2014a) list examples of observed combina-tion effects of risk factors. Table 3 shows how the effects of vascular risk factors varyalong the life course and APOE gene polymorphism interacts with life-style and car-diovascular risk factors.

The recommendation report by the National Academies of Sciences, Engineering,and Medicine and Health (2017) not only stated priorities for single-domain interven-tions (see Table 2), but also indicated that multidomain interventions are needed toinvestigate effective dementia prevention strategies. The aim would be to target mul-tiple risk factors concurrently and possibly affect several pathological disease pro-cesses. Several large-scale controlled trials with multidomain intervention strategiesare underway, or have already published results. The Prevention of Dementia by In-tensive Vascular Care (PreDIVA) randomized controlled trial (RCT) tested the efficacyof a nurse-led interventions targeting several cardiovascular, metabolic and lifestylerisk factors in an older age group in a primary health care setting, but the study failedto show an effect on dementia as the primary outcome during the 6 year follow up(Moll van Charante et al., 2016). The Finnish Geriatric Intervention Study to Pre-

43

vent Cognitive Impairment and Disability (FINGER) combined further domains inan RCT targeting at-risk individuals as discussed in more detail in section 4.3. Theintervention group was given cognitive training and provided with social activities,guidance on nutrition, an exercise program at the gym, and monitoring and manage-ment of cardiovascular and metabolic risk factors. The first results did show a statis-tically significant benefit to cognition as a primary outcome, and some subdomainsof cognition were also positively affected (Ngandu et al., 2015). In the MultidomainAlzheimer’s Prevention Trial (MAPT) cognitive training and increased physical train-ing were combined with nutritional guidance and an omega-3 fatty acid supplement,but in this older population with baseline subjective memory complaints there wereno significant differences in the primary cognitive outcome between any of the threeintervention groups and the placebo group (Andrieu et al., 2017). However, post-hoc analyses of high-risk groups defined in terms of elevated Cardiovascular RiskFactors, Aging and Dementia (CAIDE) dementia risk score (Chhetri et al., 2018) andbrain Aβ positivity (Delrieu et al., 2019) indicated positive effects. Another ongoingtrial is trying to reduce the cardiovascular risk and maintain cognitive function witha coach-supported interactive internet-based intervention for good diet, physical ac-tivity, and smoking cessation (Barbera et al., 2018).

Trials with positive findings have been able to show benefits to cognition as mea-sured by a global index, or in specific subdomains of cognition. No study has beenable to demonstrate an effect on incident dementia. Out of the large multidomainRCTs only PreDIVA was designed to do that within a 6-year follow-up time. Theyfound effects in at-risk subpopulations that were not evident in the general interven-tion population. These observations highlight the need for more efficient populationenrichment procedures. One future priority for intervention trials is to improve thesubject-selection methods by identifying those at increased risk of incident demen-tia or possibly those with subclinical disease pathology who are likely to benefit fromthe specific intervention (National Academies of Sciences, Engineering, and Medicineand Health, 2017). Disease-specific biomarkers of pathology may prove to be valu-able in subject selection, but they can also be helpful in monitoring intervention ef-fects. Further research is needed in linking biomarker-characterized pathology andclinical outcomes (National Academies of Sciences, Engineering, and Medicine andHealth, 2017).

2.8 DEMENTIA RISK MODELS AND SCORES

2.8.1 Definition of the prediction problem in a medical context

Risk modeling in medicine is multifaceted and has clinical applications for examplein diagnostics, patient selection and outcome prediction, primary prevention target-ing, and prediction of disease progression. Practical examples include the SystematicCoronary Risk Evaluation (SCORE; Perk et al., 2012) for prediction of cardiovascularfatality over 10 years, the Ottawa ankle rule for prediction of fracture and the needfor a radiograph in acute trauma (Stiell et al., 1992), and the quick Sequential [Sepsis-

44

related] Organ Failure Assessment (qSOFA; Singer et al., 2016) for prediction of highmortality risk in septic patients. The term prediction is used here for the deductionof an outcome (e.g. fracture yes/no, probability of survival at 5 years) based on asingle data point or by combining multifactorial data on the patient. The algorithmproducing this mapping between multifactorial data (predictors) and the outcome ishere defined as a prediction model. The outcome is often expressed as a binary result,but it should be noted that the context of the prediction model gives it a statistical in-terpretation in terms of the model’s sensitivity, specificity, positive predictive poweretc. The time perspectives of prediction models vary: models predicting the futureare called prognostic, and models in a cross-sectional setting are referred to as diagnos-tic (Collins et al., 2015). There are no methodological differences in the constructionof the two types of models, but the interpretation of the outcome measure determinesthe time frame (e.g. logistic regression). Other models, the Cox’s proportional haz-ards model for example, incorporate time explicitly to form prognostic predictions.

Inputs of the model ultimately determine the quality of prediction. There arebroadly two approaches to prediction model building. Predictors may be meaning-fully determined a priori based on knowledge on the biological process or on epi-demiological data on association with the outcome being modelled. In data-drivenmodel building, an algorithm determines the inclusion and weighting of candidatepredictors. In prognostic models the inclusion of a predictor naturally suggests acausative relationship with the outcome, but this does not necessarily have to be thecase. In diagnostic models the association between predictors and the outcome canbe mediated by causation, disease symptoms, or biological markers of the diseaseprocess, for example.

In the case of prognostic disease prediction models, predictors often include riskfactors that have been associated with the condition in epidemiological studies. In thebest case causality may have been established in an RCT. In the medical field, how-ever, patient data is usually difficult to obtain due to ethical considerations, costs,or data quality issues. This puts constraints on the selection of model inputs, andmodels differ in terms of breadth and complexity of data. Demographic and patient-record data can be obtained without physical contact, and self-reported data and ba-sic clinical measurements can be gathered by lower-skilled staff. More complex labo-ratory and imaging analyses requires higher-skilled medical staff. From the patient’spoint of view some examinations are more invasive and may bear risks in form ofcomplications or radiation dose. Indeed, some prediction models are designed to re-duce the need for additional analyses, and the Ottawa ankle rule is an example. Thecomplexity of a prediction model is determined by the set of predictors. Complexmultidomain models may incorporate predictors from multiple domains (e.g. de-mographic, imaging, and laboratory), whereas a simpler single-domain model couldinclude only disease genealogy, for instance.

Prediction models are constructed in a specific research setting with a specificsubject population and known method restrictions. The characteristics of the tar-get cohort in terms of age, demographic background, and risk factor profile are an

45

important part of the definition of the model and should be properly communicated.These, and other requirements are defined in the Transparent Reporting of a multi-variable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) Initiativestatement aimed at improving reporting standards in the field (Collins et al., 2015).

Prediction models are not only useful in a clinical setting, but also in research.For early phase drug trials, for example, the identification and enrolment of high-riskindividuals could increase power of the trial, reduce the number of participants re-quired, or reduce the intervention duration. A well-grounded prediction model canbe useful in enriching a study population beyond what can be achieved with the clas-sical approach of defining threshold values for selected risk factors. Solomon et al.(2019) outline two more scenarios where prediction models could be useful. Demen-tia prevention interventions could be fitted to match specific risk profiles instead ofa blanket intervention targeting a broader cohort. Additionally, utilizing predictionmodel risk estimates as trial outcomes could mitigate the need for long follow-uptimes. These estimates may prove be useful also in cases where the true outcome isvery rare.

2.8.2 Diagnostics of a prediction model

The quality of a prediction model is expressed using established statistical measuresincluding sensitivity, specificity, positive and negative predictive value, and accu-racy. These statistics should be used and reported together, as the choice of predictionmodel parameters and the tuning of the balance between sensitivity and specificityis to some extent arbitrary. The intended use of the model should guide the set-ting of parameters and threshold values. For example, in some cases false negativepredictions may be potentially life threatening and should be avoided at the cost ofspecificity. On the contrary, before executing a costly and laborious intervention avery specific model may be preferred for population enrichment.

When setting threshold values for dichotomous yes/no prediction outcomes isnot justified, a more general measure of model quality is used. The receiver operatingcharacteristic curve is a graphical presentation of model performance in the sensitiv-ity–specificity space used to describe the model’s ability to discriminate between in-dividuals. The area under the curve (AUC) quantifies the information in this graphin a single index value in the range 0.5–1, where a value of 0.5 equates to a randomprediction and 1 indicates perfect prediction. According to the statistical interpreta-tion, AUC represents the probability that the prediction model assigns a randomlychosen true positive case a higher risk estimate than a randomly chosen negativecase (Hanley and McNeil, 1982). Hosmer et al. (2013) suggest an experience-basedgeneral rule of thumb for AUC interpretation as follows: values less than 0.7 indi-cate “poor” discrimination, values between 0.7 and 0.8 “acceptable”, values between0.8 and 0.9 “excellent”, and values greater than 0.9 “outstanding”. The C-statistic(or concordance statistic) for a dichotomous outcome is an analogous measure origi-nally defined in terms of logistic regression (Hosmer et al., 2013). These measures arevalid for prediction of a binary outcome at a specific time point. Variations have been

46

developed for other applications such as survival analysis, for instance.In the building stage the model is estimated using a training population, which is

a group of individuals matching the intended cohort characteristics of the model. Asa general rule, the prediction performance of the estimated model will typically besuperior in the training population compared to what it would be in other data sets(Harrell et al., 1996). Additionally, many model types can be tuned to predict at anarbitrarily high performance level by increasing the model complexity. Hereby themodel fit increases, but not necessarily the model’s usefulness in a general setting. Themodel needs to be validated against an independent set of data. In internal validationthe study cohort itself is used by splitting it into a separate training population anda test population, against which the performance of the model is reported. For moregeneralizable and reliable assessment, external validation is performed by testing theperformance in another, independent, cohort. Studies analyzing the validation pro-cedures used in medical prediction studies have shown deficiencies, and the TRIPODguidelines also aim to standardize validation reporting (Collins et al., 2015).

2.8.3 Statistical methods underlying prediction models

Regression models are typical underlying statistical methods of prediction models.Logistic regression or Cox proportional hazards models are ways to quantify associationsbetween predictors and outcomes, and the resulting regression coefficients are goodcandidates for prediction model weights. In recent decades machine learning meth-ods have increasingly been applied in the medical field leading towards more data-driven models.

A support vector machine (SVM) is an example of a statistical method used to cat-egorize multidimensional data into prediction groups. The method relies on settingup a hyperplane in the multidimensional space defined by predictor variables in away that separates groups appropriately—that is, while avoiding overfitting. Thehyperplane is set up in reference to the closest data points which the algorithm de-fines using support vectors. In a simpler case a separation in space is obtained usinga linear hyperplane, but more complex hyperplanes can be set up by using so callednonlinear kernel functions. (Noble, 2006; Suthaharan, 2015)

A data-driven and somewhat more abstract machine learning subspecialty are al-gorithms consisting of nets or trees. An artificial neural network consists conceptuallyof net nodes and their connections— neurons and synapses. Data flows through thenet: model inputs enter the net from one side and an outcome exits the net on theother. Data is transformed at each node. Decision trees similarly facilitate process-ing of the model inputs at branching points starting at the trunk, and consecutivedecisions leading towards different outcome categories are represented by leaves.Each branching point represents an item of input data, and at each point a decisionis made about the following step. The decision is made based on a threshold value,which is a parameter of the model. The complexity of the models varies depend-ing on the number of layers of nodes/branches. For additional complexity, decisiontrees can for example be joined to form a random forest, in which parallel outcomes of

47

many trees are consolidated to a consensus outcome, or are vetted against each otherin a voting step. These models are fitted to training data. They produce typicallyblack-box-type models without the ability to carry out intuitive interpretations forcoefficients. (Graupe, 2007; Suthaharan, 2015)

Whatever the underlying technology of the model, internal validation is neededto assess the generalizability of the model. In cross-validation the study population isdivided randomly in a set proportion. One subpopulation is used to train the modeland the other, nominally independent subpopulation, is used to test the model. In10-fold cross-validation, for example, 9 out of 10 equally sized portions are used fortraining and the single leave-out portion for testing. Then, each of the 9 remainingportions are used as the testing population in sequence. This algorithm is repeated aset number of times resulting in 10×10 cross-validation, for example.

Principal components analysis (PCA) is a statistical method used to reduce the di-mensionality of high-dimension data. Conveniently, the resulting principal compo-nents (PCs) often have conceptual interpretations. A PCA on n-dimensional data re-sults in a set of n PCs. Mathematically, PCs are linear combinations (weighted sums)of original variables with the additional condition that PCs are uncorrelated to eachother. PCs are constructed in a way that the first PC attains values (PC scores) thatexplain a maximal amount of variance in the data. The next PCs are constructed sim-ilarly to maximize the coverage of the residual variance. In a typical case, a few of thefirst PCs together can explain most of the variance in a dataset. Being linear combina-tions of original variables, PCs often combine original features in a way that may bedriven by external—although not necessarily obvious—factors. Examination of PCloadings (weights of the linear combination) can help in assigning interpretations tothe PCs. (Dunteman, 1989)

2.8.4 Prognostic prediction of dementia

Prognostic prediction models have been constructed to estimate the risk of incidentcognitive impairment of varying severity and in different settings. Models have beendeveloped for predicting the conversion from MCI to AD (AUCs in the range 0.60–0.93), all-cause dementia based on late-life predictors and midlife predictors, specificdementias in late life, dementia in individuals with DM, and dementia in individ-uals from different educational backgrounds (Hou et al., 2019; Tang et al., 2015). Asystematic review by Tang et al. (2015) of models published in the preceding fiveyears (21 articles assessed) found an overwhelming majority of models to be builtaround a scoring system derived from logistic regression or Cox proportional haz-ards models. Two were constructed using a priori epidemiological evidence. The au-thors recognized nine distinct predictor modalities: demographic, subjective cogni-tive complaints, neuropsychological testing, health (symptoms, diagnoses, and mea-surements), lifestyle, diet, gene analytics, and MRI. Some models included predictorsoutside of these categories, for instance family history of dementia or cognitive activ-ity.

A recent review by Hou et al. (2019) identified 46 studies predicting incident de-

48

mentia in cognitively healthy individuals. Seven of these had been externally val-idated in terms of their discrimination performance. These seven models are sum-marized in Table 4. Most are intended for a general population, but two models arebuilt to predict dementia in individuals with DM2, a known risk factor of dementia.The Cardiovascular Risk Factors, Aging and Dementia (CAIDE; Kivipelto et al., 2006)risk score models the midlife risk of developing incident dementia in late life, whilemost of the other studies predict late-life dementia on a shorter time span of 3–10years. The Australian National University Alzheimer’s Disease Risk Index (ANU-ADRI; Anstey et al., 2013) has an age-adaptable design in that the scores given topredictors are age-dependent in cases where prior research supports this approach.For example, overweight and high cholesterol only increase risk at ages below 60.Additionally, age-related risk is stratified according to sex.

Most models have been built using a data-driven approach, where model predic-tors have been chosen from an available opportunistic set of variables using statisti-cal testing. Four studies use the Cox proportional hazards model which allows forconvenient treatment of attrition in the older cohorts. Two studies utilize a priorievidence on dementia risk factors and build models directly using predeterminedpredictors. ANU-ADRI is based on a systematic review of potential risk factors andmodel weights are determined from earlier published estimates. The complexityof the models varies a lot. The most focused prediction model consists of the free-recall score of the Free and Cued Selective Reminding Test (FCSRT-FR)—this modelis also among the best performing. Most models include age, but it should be notedthat the FCSRT-FR does not. An analysis by Mura et al. (2017) showed that com-bining age with free recall did not improve the results. Another well-performingmodel is the Taiwanese Health Improvement Network (THIN; Walters et al., 2016)registry-based model that identified about 930,000 patients for the training cohortand 260,000 patients for the test cohort, and analysed easily available demographic,life-style, prescription, and diagnosis data for effects. The performance was goodin the 60–79 age group, but a model trained with 80+ individuals had practicallyno predictive power. The other general-population late-life models—ANU-ADRIand the Dementia Screening Indicator (Barnes et al., 2014)—performed both onlymoderately despite ANU-ADRI’s multimodal extensive predictor set and evidence-based selection methodology. The models built for DM2 populations both includeddiabetes-related comorbidities, and the other additionally laboratory measurementsand medication information, and both had relatively long prediction horizons. Theyperformed equally at an acceptable performance level.

The CAIDE score combines demographic factors with cardiovascular health fac-tors for prediction over a longer time frame. The acceptable performance of the origi-nal study was replicated in a validation study with nearly double the follow-up time(Exalto et al., 2013a). The validation also showed the model to work well in Asian,black, and white cohorts (AUCs respectively 0.81, 0.75, and 0.74). However, the per-formance was clearly worse in older cohorts (Anstey et al., 2014), most likely due tothe varied effect of BMI and cholesterol in those age groups. The CAIDE score ver-

49

Tabl

e4:

Pro

gnos

ticde

men

tiapr

edic

tion

mod

els

with

exte

rnal

lyva

lidat

eddi

scrim

inat

ion

perfo

rman

ce.

Iden

tified

from

Hou

etal

.(20

19).

Mod

elBa

selin

eag

e,yr

.Po

pula

tion

crit

erio

nFo

llow

up,

mea

nyr

.Pr

edic

tors

incl

uded

Pred

icto

rse

lect

ion

AU

Cor

igin

alA

UC

val-

idat

ion

CA

IDE

(Kiv

ipel

toet

al.,

2006

;Ex

alto

etal

.,20

13b)

1M

idlif

eN

one

21(v

alid

a-ti

on36

)A

ge,e

duca

tion

,sex

,SBP

,BM

I,ch

oles

tero

l,ph

ysic

alin

acti

vity

(and

APO

E)D

ata-

driv

en(L

R)

0.77

(0.7

8w

.APO

E)0.

75

FCSR

T-FR

(Gro

ber

etal

.,20

10;D

erby

etal

.,20

13;M

ura

etal

.,20

17)

>65

Non

e3–

5Fr

eere

call

scor

eA

prio

riev

iden

ce—

0.81

–0.8

9

AN

U-A

DR

I(A

nste

yet

al.,

2013

,201

4)A

llag

esN

one

4–6

Age

,sex

,edu

cati

on,D

M,t

raum

atic

brai

nin

jury

,co

gnit

ive

acti

vity

,soc

iale

ngag

emen

t,sm

okin

g,al

coho

l,ph

ysic

alac

tivi

ty,fi

shin

take

,dep

ress

ive

sym

ptom

s

Apr

iori

evid

ence

—0.

65–0

.732

Dem

enti

aSc

reen

ing

Indi

cato

r3

(Bar

nes

etal

.,20

14)

>65

Non

e64

Age

,edu

cati

on,B

MI,

DM

,str

oke,

need

she

lpw

ith

mon

ey/m

edic

atio

n,de

pres

sive

sym

ptom

sD

ata-

driv

en(C

ox)

—0.

68–0

.78

TH

IN5

(Wal

ters

etal

.,20

16)

60–7

9N

one

5A

ge,s

ex,d

epri

vati

on,B

MI,

anti

hype

rten

sive

med

icat

ion,

smok

ing,

alco

hol,

DM

,dep

ress

ion,

stro

ke/T

IA,A

F,as

piri

nus

e

Dat

a-dr

iven

(Cox

)—

0.84

80–9

5N

one

5A

sab

ove,

bute

xclu

ding

depr

ivat

ion,

and

addi

ngSB

P,lip

idra

tio,

anxi

olyt

ics,

NSA

IDD

ata-

driv

en(C

ox)

—0.

56

DSD

RS

(Exa

lto

etal

.,20

13a)

>60

DM

210

Age

,edu

cati

on,c

ereb

rova

scul

aror

card

iova

scul

ardi

seas

e,de

pres

sion

,dia

beti

cco

mpl

icat

ions

Dat

a-dr

iven

(Cox

)0.

740.

75

ND

CM

P(L

ieta

l.,20

18)

>50

DM

28

Age

,sex

,DM

2du

rati

on,B

MI,

vari

atio

nof

fast

ing

gluc

ose

and

HbA

1c,s

trok

e,hy

pogl

ycem

ia,C

AD

,D

Mm

edic

atio

n

Dat

a-dr

iven

(Cox

)0.

760.

756

Key

:AF

atri

alfib

rilla

tion

,AU

Car

eaun

der

the

RO

Ccu

rve,

BMIb

ody

mas

sin

dex,

CA

Dco

rona

ryar

tery

dise

ase,

Cox

Cox

prop

orti

onal

haza

rds

mod

el,D

Mdi

abet

es,

HbA

1cgl

ycat

edhe

mog

lobi

n,LR

logi

stic

regr

essi

on,

NSA

IDno

nste

roid

alan

ti-i

nflam

mat

ory

drug

,SB

Psy

stol

icbl

ood

pres

sure

,TI

Atr

ansi

ent

isch

emic

atta

ck,

—no

tav

aila

ble.

Mod

els:

CA

IDE

Car

diov

ascu

lar

Ris

kFa

ctor

s,A

ging

and

Dem

enti

a;FC

SRT-

FRFr

eean

dC

ued

Sele

ctiv

eR

emin

ding

Test

,fr

eere

call;

AN

U-A

DR

IA

ustr

alia

nN

atio

nalU

nive

rsit

yA

DR

isk

Inde

x;TH

INTh

eH

ealt

hIm

prov

emen

tNet

wor

k;D

SDR

SD

iabe

tes

Spec

ific

Dem

enti

aR

isk

Scor

e;N

DC

MP

Nat

iona

lDia

bete

sC

are

Man

agem

ent

Prog

ram

.Fo

otno

tes:

1:

Val

idat

ion

inm

atch

ing

coho

rts

only

.2:

Som

epr

edic

tors

mis

sing

.3:

Sem

i-ex

tern

alva

lidat

ion,

pred

icto

rsan

dw

eigh

tsde

term

ined

bypo

olin

gov

erco

hort

s.4:A

com

puta

tion

alK

apla

n-M

eier

surv

ival

.5:R

etro

spec

tive

regi

stry

stud

y.6:C

ohor

tspl

ittr

aini

ng/t

esti

ng,1

0-yr

risk

esti

mat

e

50

sion with APOE ε4 status as a predictor performed marginally better than the basicversion.

The increasing prominence of brain pathology and biomarkers in dementia re-search has not yet penetrated into prognostic modeling. Models exist with MRIimaging predictors and genetic information (Tang et al., 2015), but amyloid or taumarkers, or markers of LBD pathology for that matter, have not been incorporated.As for the validation of these current and future advanced models, finding externalcohorts with the same expensive and possibly cumbersome biomarker analysis andlong follow-up times is difficult. Tang et al. (2015) point out that only few studiestake into consideration the costs associated with gathering predictor data, and thatthe problem of high costs is especially amplified in a population-based setting. Itwould be desirable to aim for a minimal predictor set while maintaining good predic-tion performance. The analysis of validated models (Table 4) showed top results for asimple score showing the impairment of free recall. The predictor was not supportedby any other predictor modality. Additionally, increasing the model complexity didnot always seem to improve prediction performance. The CAIDE score was not sig-nificantly improved by the inclusion of APOE ε4 status, nor did it show any higherperformance when additional midlife predictors were added (Exalto et al., 2013b).

Key methodological challenges of prognostic dementia prediction models wereidentified in a recent review (Goerdten et al., 2019). 33% of models were not validatedexternally or internally, and only 10% were validated externally. A large portion ofthe studies (44%) were built on ADNI data making the results less generalizable, es-pecially when external validation is not performed. The authors also commented thespecific problems with machine learning models, which is the most common modeltype with a 43% share of all models used. Although they are efficient and accurate,these data-driven models rely strongly on the selected data source. This may makethem difficult to apply in other settings. Indeed, only one externally validated modelidentified in Table 4 used a machine learning model. Typically the case frequencyis also higher in study populations than in a real-world setting. Models using re-gression were noted to frequently not check underlying data assumptions, such aslinearity.

It has been suggested that prediction efforts should in the future take into accountsubtle disease-induced changes in clinical testing, biomarker evidence of early dis-ease stages, and the changing nature of biomarkers during the life course (Ritchieand Muniz-Terrera, 2019). Furthermore, opportunities provided by modern statis-tical approaches such as machine learning algorithms should be investigated morethoroughly. Future prediction studies will show if incorporating new factors moreclosely linked to specific disease pathologies will allow for more precise results. Suchan approach narrows the gap between purely associative risk factors and diagnosticmarkers of disease, and such models would start resembling diagnostic models ormodels of disease progression. In prediction models for advanced age the diagnosticand prognostic models are easily intertwined, as the disease process is more likelyalready ongoing even if symptoms are not showing.

51

2.8.5 Prediction of brain amyloid

No prognostic prediction models for brain Aβ accumulation have so far been pub-lished. Diagnostic models do exist to predict Aβ pathology as confirmed by PETimaging or analysis of CSF, but no systematic review has been published on the re-sults. Terminology on the subject varies in the literature and the prediction problemhas, among other terms, been framed as “imputation” or “ascertainment”. A liter-ature search was performed on PubMed.gov on 9 May 2019. Study titles were re-viewed, and when necessary abstracts were investigated for relevance. The searchwas performed with the following search query:

(AD OR Alzheimer’s OR CSF OR PET)

AND (amyloid OR Aß OR beta-amyloid OR

amyloid-beta OR amyloidosis)

AND (prediction OR pre-screening OR

imputation OR ascertainment)

The search produced 407 results. 11 studies were identified, and two more basedon references in other reports. The results are presented in Table 5. Target populationsincluded cognitively normal (CN), MCI, and AD participants, but stratified resultswere not reported. The youngest cohort consisted of over-50-year-olds, althoughmost were older. Many studies used Alzheimer’s Disease Neuroimaging Initiative(ADNI) subcohorts that included at least 55-year-old participants. The Aβ status wasdetermined either by CSF analysis or PET. Some studies used a population with Aβ

CSF for training and a PET population for validation, or vice versa. Six of the modelsused external validation, others used either an internal validation only or no valida-tion at all. External validation was acknowledged only if the prediction model wascompletely estimated in the training set. It did not, for example, suffice that a data-driven MRI classifier was estimated using a separate cohort, but the weights of themultimodal model were estimated in the test cohort (Tosun et al., 2013). Some stud-ies saved a portion of their study population for external validation instead of usingan independent cohort. These cases have been highlighted in the table. All studieswere diagnostic, although some did use longitudinal data on cognition as a predictor.Three of the older studies used a logistic regression to build the models, whereas thenewer models mostly used machine learning techniques. Ansart et al. (2019) testeda random forest model, logistic regression, SVM, adaptive logistic regression, andan adaptive boosting model (AdaBoost) on different cohorts and found the randomforest to have the best overall performance. A random forest model was also usedby two other studies, and one study used a simpler decision tree model. An SVMwas used in three studies. Two studies did not report an AUC value but reportedaccuracy and positive prediction value instead.

The prediction performance varied according to the severity of the diagnoses in-cluded in the study population, as would be expected. AUCs were at the lowest levelsin the pure-CN populations (0.77 and 0.74, non-validated) and at the highest in mixedMCI/AD cohorts (0.87–0.88). Demographic information was included in most mod-

52

Tabl

e5:

Dia

gnos

ticpr

edic

tion

mod

els

forA

βpa

thol

ogy.

Mod

alit

ies

incl

uded

AU

C(o

ral

tern

ativ

e)

Stud

yA

gegr

oup

Pred

icti

onco

hort

Mod

elty

peD

emo.

Cog

n.A

POE

sMR

IBl

ood

Oth

erIn

tern

al/n

ova

lidat

ion

Exte

rnal

valid

atio

n

Mie

lke

etal

.(20

12)1

70–9

2C

NLR

xx

Mem

ory

com

plai

nt0.

70x

xx

Mem

ory

com

plai

nt0.

70Ba

har-

Fuch

set

al.(

2013

)>

60M

CI

LRx

0.77

–0.8

6To

sun

etal

.(20

13)

>55

MC

ILR

x0.

81x

0.70

xx

0.88

Burn

ham

etal

.(20

14)1

>60

CN

/MC

I/A

DR

Fx

x0.

810.

69x

xx

0.84

0.82

xx

xC

DR

0.88

0.85

Tosu

net

al.(

2014

)1>

55M

CI

LRx

xx

0.83

(CA

)x

xA

SL-M

RI

0.80

(CA

)H

aghi

ghie

tal.

(201

5)>

55C

N/M

CI

DT

x0.

762

x0.

742

xx

0.87

2

Apo

stol

ova

etal

.(20

15)

>55

MC

ISV

Mx

xx

xx

0.80

0.78

2

Inse

leta

l.(2

016)

1>

55C

NR

Fx

xx

Cog

niti

vech

ange

0.65

(PPV

)Le

eet

al.(

2018

)>

55M

CI/

AD

LRx

x0.

800.

72x

x0.

740.

70x

xx

Hyp

erte

nsio

n0.

870.

80W

estw

ood

etal

.(20

18)

64–7

13C

N/M

CI/

AD

SVM

x0.

67–0

.69

ten

Kat

eet

al.(

2018

)1>

50C

NSV

Mx

xx

x0.

74M

CI

xx

xx

0.81

Palm

qvis

teta

l.(2

019)

>60

CN

/MC

ILR

xx

x0.

81–0

.83

0.81

–0.8

2x

xx

x0.

83–0

.85

0.83

Ans

arte

tal.

(201

9)1

>55

–70

CN

/MC

IR

F4x

xx

x0.

67–0

.83

xx

xC

ogni

tive

chan

ge0.

72–0

.89

xx

x0.

61–0

.68

0.62

–0.6

6K

eyfo

rpr

edic

tion

coho

rt:C

Nco

gnit

ivel

yno

rmal

,MC

Imild

cogn

itiv

eim

pair

men

t,A

DA

lzhe

imer

’sdi

seas

e.Pr

edic

tion

mod

elty

pes:

LRlo

gist

icre

gres

-si

on,R

Fra

ndom

fore

stm

odel

,DT

deci

sion

tree

mod

el,S

VM

supp

ortv

ecto

rm

achi

ne.A

lter

nati

vepe

rfor

man

cem

easu

res:

CA

clas

sific

atio

nac

cura

cy,P

PVpo

siti

vepr

edic

tive

valu

e.O

ther

:A

SL-M

RI

arte

rial

spin

labe

lling

MR

I,A

POE

apol

ipop

rote

inE,

AU

Car

eaun

der

the

RO

Ccu

rve,

CD

RC

linic

alD

emen

tia

Rat

ing,

sMR

Ist

ruct

ural

MR

I.Fo

otno

tes:

1:

Som

em

odel

sw

ith

few

erpr

edic

tor

mod

alit

ies

omit

ted,

2:

Coh

ort

split

for

exte

rnal

valid

atio

n,3:

Ran

geof

popu

lati

onm

ean

ages

,4:R

Fbe

stam

ong

4ot

her

met

hods

test

ed.

53

els. The APOE genotype was an important predictor, especially in the cognitivelyimpaired. In a CN population APOE added alongside demographic information im-proved the AUC from 0.62 (not shown in table) to 0.70 (Mielke et al., 2012), and inan MCI cohort it improved a structural MRI predictor from AUC 0.70 to 0.88 (Tosunet al., 2013). In an MCI/AD cohort iterative reporting of performance measures withincreasing model complexity seemed to indicate APOE as a strong predictor (com-plete model AUC 0.87 non-validated, 0.80 validated; Lee et al., 2018). the objectivemeasurement of cognition had a predictive value in MCI subjects as a solitary predic-tor (Bahar-Fuchs et al., 2013). Adding a 24-month cognitive change could improve onthe cross-sectional measure somewhat in a CN population (not shown in table; Inselet al., 2016). Among older CN individuals objective cognitive scores were equal tosubjective memory complaints in terms of predictive power (Mielke et al., 2012). In acohort with MCI participants cognition and a blood assay had similar performanceson their own (AUC 0.74–0.76 validated), but the multimodal model achieved AUC0.87 (Haghighi et al., 2015). Structural MRI data was included in five models, and itsadded benefit to the model was demonstrated in two studies reporting performancefor parallel models (Tosun et al., 2013, 2014). Ansart et al. (2019) concluded that cog-nitive scores were superior to MRI as an alternative and that adding MRI with thecognitive scores did not improve results significantly—an important finding consid-ering costs and practicality of the model. Structural MRI was demonstrated to bemore effective in predicting than arterial spin labeling MRI measuring brain bloodflow (Tosun et al., 2014).

2.8.6 Prediction models in prevention trials

Three dementia prediction models have so far been used in intervention trials. TheCAIDE risk score was used to select at-risk individuals to take part in the FinnishGeriatric Intervention Study to Prevent Cognitive Impairment and Disability (FIN-GER; Ngandu et al., 2015, see section 4.3) RCT. The target cohort was designed to in-clude individuals with an increased risk of incident dementia based on preventablecardiovascular risk factors and with below-average cognitive performance, yet nosubstantial cognitive impairment. The CAIDE score threshold of ≥ 6 amounted to avery mild enrichment: 84% of the available population met this requirement (Nganduet al., 2014) with the lowest individual late-life dementia risk of 1.9% (95% confidenceinterval 0.2–3.5; Kivipelto et al., 2006) at the threshold level.

Another midlife life-style intervention trial, the Innovative Midlife Interventionfor Dementia Deterrence (In-MINDD), utilized a risk score of modifiable risk factorsconstructed based on a literature search (O’Donnell et al., 2015). In this study thescore was used as an educational tool to inform participants of their individual riskprofile. The personalized Lifestyle for Brain health (LIBRA) score takes into accountcoronary disease/hypertension and factors affecting those, obesity/diabetes and as-sociated life-style practices (physical activity and diet), renal disease, and alcoholconsumption. The individual risk profiles were used to motivate the proper man-agement of chronic diseases and to communicate in which areas lifestyles could be

54

improved.Recently, a genetic risk model was used to stratify subjects into low and high risk

groups in a delay-of-disease RCT (TOMMORROW) testing a DM2 medication on theincidence of MCI due to AD in the high-risk group (ClinicalTrials.gov, 2018). Therisk model combines the APOE genotype, translocase of outer mitochondrial mem-brane 40 homolog genotype, and age to produce the risk class prediction (Lutz et al.,2016), which in turn is used as a 5-year prognosis. The prediction model had previ-ously been externally validated for short term risk prediction. The trial was partlydesigned to validate the performance of the model by comparing a low-risk groupand a high-risk placebo group, but the trial was terminated prematurely following afutility analysis.

Prediction models in future dementia research

Ongoing research initiatives aim to build large well-managed and well-phenotypedcohorts with a variety of risk factor information. For example, the European Preven-tion of Alzheimer’s Dementia (EPAD) project aims to internationally improve the useof current cohorts and develop a longitudinal cohort for research of future interven-tions (Ritchie et al., 2016). Good quality and comprehensive risk factor coverage aswell as the inclusion of biomarker data will allow for novel prediction models, whichfurthermore aid in designing new interventions. Prediction models for pathologymay in the future be helpful in cost-effectively identifying target individuals withpathology for secondary prevention (i.e. diagnostic models) or for primary preven-tion (i.e. models more tuned for prognostic prediction).

55

3 AIMS OF THE STUDY

The general aim of this thesis was to develop prediction models for dementia andbrain pathology and to investigate associations between brain amyloid accumulationand diabetes-related markers. Prediction models may potentially be useful in iden-tifying at-risk individuals, targeting interventions, and finding optimal participantsto dementia research projects. Diabetes-related markers are particularly relevant inthis context given the increasing diabetes prevalence and potential mechanistic linksto dementia diseases. The specific aims were:

1 To predict incident dementia over a ten year period in a late-life cognitivelyhealthy population with multimodal predictors and a novel machine learningalgorithm (Study I).

2 To predict dementia and brain pathology in a population-based cohort of theoldest of old using multimodal predictors and a novel machine learning algo-rithm (Study II).

3 To predict the presence of in-vivo amyloid pathology in a cognitively healthyelderly population at risk of dementia with multimodal predictors and a novelmachine learning algorithm (Study III).

4 To study the associations of insulin resistance and other markers of type-twodiabetes with brain amyloid pathology in vivo in a cognitively healthy elderlypopulation at risk of dementia (Study IV).

57

4 SUBJECTS AND METHODS

Three separate study populations were used in this thesis project to build and val-idate prediction models. In the two observational studies CAIDE and Vantaa 85+,prognostic prediction models were built for dementia and brain pathology. Baselinedata from the FINGER intervention trial was used in diagnostic prediction of brainamyloid (FINGER-PET) and to assess associations between brain amyloid status andmetabolic markers of insulin resistance and diabetes (FINGER IR/DM). These studiesand the respective outcome measures are summarized in Table 1.

4.1 THE CAIDE STUDY OF YOUNGER OLD INDIVIDUALS

The longitudinal, observational, population-based Cardiovascular Risk Factors, Ag-ing and Dementia (CAIDE) study is an extension of cardiovascular surveys con-ducted in the 1972–1987 within the North Karelia Project and the Finnish part ofthe Monitoring Trends and Determinants in Cardiovascular Disease (FINMONICA)study (Puska et al., 1979, 1983; Vartiainen et al., 1994). These surveys were targetedat middle-aged persons with a mean age of 50.6 years at the initial visit. Later, for thepurposes of the CAIDE study (Kivipelto et al., 2001a,b), a random sample of 2,000individual participants aged 65–79 years were invited to a re-examination. The struc-ture of the study is described in more detail in Figure 1. 1,449 persons took part inthis first late-life re-examination in 1998. A second late-life follow-up was conductedin 2005–2008. This time 1,426 participants out of the initial 2,000 were still alive, and909 participated. Late-life visits were conducted at median ages 71.3 and 78.6 years.The CAIDE study was approved by the local ethics committee of Kuopio University

Table 1: Outcome measures in prognostic/diagnostic prediction by category in the three studycohorts of the thesis.

Prediction outcome by category CAIDEN=709&1,009(prognostic)

Vantaa 85+N=163&97

(prognostic)

FINGERN=48&41

(diagnostic)

Incident dementia + + -AD pathology Aβ plaques - Post mortem In vivo

Tau tangles - Post mortem -Vascular pathology Cerebral microinfarcts - Post mortem -

Cerebral macroinfarcts - Post mortem -Cortical macroinfacts - Post mortem -WM macroinfarcts - Post mortem -

Other pathology α-synuclein - Post mortem -CAA - Post mortem -Hippocampal sclerosis - Post mortem -TDP-43 protein - Post mortem -

Key: AD Alzheimer’s disease, CAA Cerebral amyloid angiopathy, WM White matter

59

Independent random samples from the North Karelia Project & FINMONICA study

Dementia N=61,MCI N=82.

Invited random sample N=2,000Age 65–79 y, location Kuopio & Joensuu

Midlife1972, -77, -82, and -87

1st late-lifere-examination1998

2nd late-lifere-examination2005–2008

Completed cognitive assessment N=1,409,Participants N=1,449

Invited eligible persons N=1,426Alive, living in region, known address

Dementia before end of 2000 N=13.

Completed cognitive assessment N=852,Participants N=909. New dementia N=62.

Cognitively healthy

Main study population N=709Dementia at 2nd re-examination N=39

Extended study population N=1,009Dementia N=151

National health registries

- Cognitively healthy at 1st re-examination- Participated in 2nd re-examination

- Cognitively healthy at 1st re-examination- No dementia in registries before end of 2000- May or may not have participated at

2nd re-examination- Nonparticipants who died without dementia

record excluded

Figure 1: CAIDE study design and formation of the study cohorts.

Hospital, and written informed consent was obtained from all participants.Study I of this thesis predicted dementia in the participants of the first late-life visit

in 1998 who were verified to be cognitively healthy—that is those with no MCI or de-mentia diagnosis. 709 of that cohort also participated in the 2005–2008 re-examinationafter a mean follow-up time of 8.3 years. This cohort formed the main study popula-tion of Study I. An extended study population was formed by augmenting this withhealth registry data. For an additional 300 individuals who did not participate in thelater re-examination register information on dementia diagnoses and mortality wasused. Any relevant record in the Hospital Discharge Register, Drug ReimbursementRegister, or Cause of Death Register before the end of 2008 led the individual to beclassified as having dementia. These registers have been found to have a good posi-tive predictive value, but lower sensitivity (Solomon et al., 2014b). Surviving nonpar-ticipants without a diagnosis were counted as not having dementia, and those thathad died without a diagnosis before 2008 were excluded. Additionally, individualswho had a recorded dementia diagnosis before the end of 2000 were excluded. Themean follow-up time in this extended population was 9.0 years.

60

Table 2: Baseline factors evaluated for model inclusion by category in the three study cohorts.

Population type and factors by category CAIDE Vantaa 85+ FINGER

N 709 & 1,009 245 & 163 48 & 41Age criterion Late-life > 85 60–77Population type General

populationGeneral

populationAt-risk

CNDemographic1 Education + + +

Social class − + −Cognition MMSE or SPMSQ + + −

Neuropsychological testing − − +Subjective complaints + + −Activities of daily living − + −

APOE genotype + + +Comorbidities Cardiovascular2 + + −

Diabetes mellitus + + +Stroke/TIA + + −

Vascular/DM Blood pressure + + +Blood pressure change + − −Lipids + + −Cholesterol change + − −Body mass index + + +Body mass index change + − −Waist-hip ratio + − −Smoking + + −Self-rated fitness + − −Physical activity + − −Insulin resistance − − +HbA1c − − +Blood assay of DM markers − − +

Psychosocial Depression + + −Hopelessness + − −

Structural MRI − − +Other Alcohol use + + −

Self-rated health + − −Key: APOE apolipoprotein E, CN cognitively normal, DM diabetes mellitus, HbA1c gly-cated hemoglobin, MMSE Mini mental state examination, SPMSQ Short portable mentalstatus questionnaire, TIA transient ischemic attack. Footnotes: 1: All studies include ageand sex, 2: Includes angina pectoris, atrial fibrillation, coronary heart disease, heart failure,hypertension, myocardial infarction, and arteriosclerosis obliterans.

61

The CAIDE late-life visits utilized a three-step procedure to assess cognition. Dur-ing the screening phase the participants were interviewed, and a set of tests were con-ducted to assess different cognitive domains: general cognitive screening with theMini-Mental State Examination (MMSE; Folstein et al., 1975), episodic memory withthe immediate word recall test (Nyberg et al., 1997; Heun et al., 1998), semantic mem-ory with the category fluency test (Borkowski et al., 1967), psychomotor speed with thebimanual Purdue Peg Board test (Tiffin, 1968) and the letter digit substitution test(Wechsler, 1944), executive function with the Stroop test (Stroop, 1935), and prospectivememory with a prospective memory task (Einstein et al., 1997).

An MMSE score of ≤24 indicated a referral to the clinical assessment phase, and in2005–2008 this was also indicated by a decrease of ≥3 points, a delayed recall wordlist score ≤70% of the Finnish CERAD, or an informant claim of cognitive decline(2005–2008 criteria were sensitized to identify MCI better). A review board assessedthe results from detailed somatic and neuropsychological testing, and when neces-sary used blood analysis, imaging, and in some cases a CSF analysis in the differ-ential diagnosis phase. Dementia was diagnosed according to DSM-IV and specificdementias were identified according to established criteria. Dementia at the secondCAIDE re-examination was the prediction target in Study I. The prediction targets inall studies of this thesis are summarized in Table 1.

Extensive data on health and behavior related factors were collected at each late-life visit in addition to the cognitive assessments. Self-administered questionnaireson sociodemographic characteristics, medical history, and health related behaviorwere used. Depression was assessed using the Beck Depression Inventory (BDI;Beck et al., 1961) and self-rated memory was assessed by administering the Subjec-tive Memory Questionnaire (SMQ; Powell, 1980). The APOE genotype was assessedfrom leukocytes using a polymerase chain reaction and HhaI digestion (Tsukamotoet al., 1993). Table 2 lists all the available factor modalities and factors from the 1998visit that were considered as potential predictors in Study I.

4.2 THE VANTAA 85+ STUDY OF OLDEST OLD INDIVIDUALS

The Vantaa 85+ study is a longitudinal observational study of cognition and postmortem neuropathology in the oldest of the old (Polvikoski et al., 1995; Rastas et al.,2010; Ahtiluoto et al., 2010). Residents of Vantaa—a city in southern Finland—aged≥85 years were invited to participate in the study 1991. The study structure is out-lined in Figure 2. The participation rate was very high at 98%. The baseline clinicalexamination was successfully completed for 553 persons. A cohort of 339 individualscompleting the baseline examination who were assessed not to have dementia consti-tuted the cohort from which the two study populations of Study 2 were derived. Allparticipants gave their written informed consent to participate in the baseline exami-nation, and nearest relatives of the deceased signed written consent for the autopsies.The study was approved by the ethics committee of the Health Centre of the City ofVantaa.

62

All Vantaa residents aged ≥85 in 1991N=601 Did not attend baseline

examination:⁻ Died N=36⁻ Not reached N=1⁻ Refused N=11

Baseline examination N=553 Exclusion: dementia N=214

Pathology prediction cohortN=163

Dementia prediction cohortN=245, dementia N=97

Autopsy not conductedN=176

Exclusion: death within 2 yearsN=94

No dementia at baselineN=339

Figure 2: Vantaa 85+ study design and formation of the study cohorts.

The dementia prediction cohort consisted of those taking part in re-examinationsin 1996, 1999, and 2001 to assess dementia. In addition, diagnoses were recorded for101 participants prior to death. 94 participants who died within a two year windowfrom the baseline visit were excluded in order to limit differences in the time to deathfor dementia/non-dementia participants. The dementia prediction cohort consistedof 245 individuals.

The pathology prediction cohort of 163 persons consisted of those who did nothave dementia at the baseline and had autopsy data available. Within the Vantaa85+ study altogether 304 autopsies were performed, and 16 out of these were onindividuals who died before the baseline visit.

Dementia was diagnosed by a two-party consensus based on somatic, cognitive,and functioning assessments during visits and available health records. Data wasgathered at the baseline visit by a physician and a trained nurse. The baseline fac-tors assessed for eligibility as a predictor are listed in Table 2. The MMSE and theShort Portable Mental Status Questionnaire (SPMSQ; Pfeiffer, 1975) were used for acognitive assessment, and functioning was assessed with the activities of daily liv-ing questionnaire and with the Instrumental Activities of Daily Living Scale (ADLand IADL; Katz et al., 1963; Lawton and Brody, 1969). Competence in daily activitieswas quantified on a self-rated scale of 1–6 (from independent to needs help in all ac-tivities). Subjective memory complaint was assessed as no, a little, or yes. Depressionwas assessed using the Zung Self-Rating Depression Scale (Zung et al., 1965). For sur-veyed comorbidities, the category noted as Cardiovascular in Table 2 included anginapectoris, heart infarction, atrial fibrillation, heart failure, arteriosclerosis obliterans,and hypertension. HDL and LDL were determined from blood samples using enzy-matic methods (Rastas et al., 2010). The APOE genotype was determined using DNA

63

minisequencing and amplification through a polymerase chain reaction followed byrestriction enzyme digestion with HhaI (Hixson and Vernier, 1990; Syvanen et al.,1993). DSM-III criteria were used for dementia, and appropriate established criteriafor specific dementias.

Several pathological features were identified and classified in brain autopsies.These features are grouped in Table 1 by the pathology type. Aβ pathology (Polvikoskiet al., 1995) was classified using the CERAD protocol, and tau pathology (Myllykan-gas et al., 1999) was classified by using Braak staging. Macroscopic and microscopicinfarcts were identified as previously described (Tanskanen et al., 2012). Addition-ally, cerebral amyloid angiopathy (Tanskanen et al., 2012), FTD-related α-synucleinpathology (Oinas et al., 2009), hippocampal sclerosis (Kero et al., 2018), and TDP-43accumulation (Kero et al., 2018) were assessed.

4.3 THE FINGER TRIAL

The Finnish Geriatric Intervention Study to Prevent Cognitive Impairment and Dis-ability (FINGER) was a blinded, randomized controlled trial with the aim to test amultidomain life-style intervention in the prevention of cognitive decline and de-mentia (Kivipelto et al., 2013; Ngandu et al., 2014, 2015). The study had a multicentredesign and included a population-based sample of elderly persons who were at riskof cognitive decline. The sample originated from Finnish health surveys from 1972–2007 as shown in Figure 3. Along with an age criterion, participants were requiredto have a CAIDE risk score (Kivipelto et al., 2006) greater than or equal to six to beinvited to a screening visit. In more detailed testing, candidate participants had tomeet CERAD criteria that demonstrated cognitive performance at a mean level orsomewhat lower than the Finnish general population (Hanninen et al., 2010). Thespecific criteria were word list learning task of ten times three words score less thanor equal to 19, word list recall less than or equal to 75%, or an MMSE score less thanor equal to 26. Exclusion criteria included dementia, substantial cognitive decline,MMSE less than 20, and conditions preventing safe engagement in intervention ac-tivities (Kivipelto et al., 2013). The subsequent multidomain intervention includeddiet guidance, exercise, cognitive training, and vascular monitoring over a two yearperiod. Results have been published showing a benefit on overall cognition (Nganduet al., 2015), and extended follow-ups of the study participants are still ongoing. Thestudy was approved by the coordinating ethics committee of the hospital District ofHelsinki and Uusimaa. Participants gave written informed consent at the screeningand baseline visits.

A subset of the participants in the Turku area—a city in south-western part ofFinland—was invited to take part in an amyloid-PET/MRI substudy. In total 48 in-dividuals underwent PET imaging using 11C-Pittsburgh compound B (PIB) after thebaseline visit. Details on the imaging are presented by Kemppainen et al. (2017). TheFINGER-PET participants were somewhat older (mean age 70.8 vs. 69.3) than theparent cohort due to the later initiation of the recruitment process in Turku. No other

64

FINRISK surveys 1972–2007Random population-based samples

FIN-D2D surveys 2004 & 2007Random population-based samples

FINGER eligibility screening visit2009–2011 N=2,654(CERAD criterion)

FINGER prescreening 2009 N=5,496(Age 60–77, CAIDE score ≥6)

FINGER baseline visit and randomization for intervention

N=1,260

FINGER-PET substudy N=48(baseline PIB-PET & MRI)

FINGER intervention trial

FINGER PET & IR/DM cohortN=41 (baseline PIB-PET & DM markers)

Figure 3: FINGER study design and formation of the FINGER-PET study population and thePET & IR/DM cohort.

significant differences were noted. PIB images were analysed by two experiencedreaders and a consensus visual assessment of amyloid positivity (Aβ+) was made.Aβ+ individuals typically showed cortical retention predominantly in AD-typical re-gions, and Aβ- persons displayed nonspecific accumulation in white matter. Thiscohort constituted the prediction cohort of Study III with amyloid positivity on PIB-PET as a prediction outcome. See Table 1 for comparison of cohorts.

Participant health data was gathered at the baseline/randomization visit by astudy physician and nurse. Cognition was measured using a modified version ofthe Neuropsychological Test Battery (mNTB; Harrison et al., 2007). Subscores wereused for the executive functioning, memory, and processing speed cognitive do-mains. Scores of individual cognitive tests were transformed into standardized Zscores and then the sum scores for the NTB total and sub-domains were calculated(Kivipelto et al., 2013). The APOE genotype was determined by polymerase chain re-action using TaqMan genotyping assays for 2 single-nucleotide polymorphisms andan allelic discrimination method (De la Vega et al., 2005).

All participants of the FINGER-PET cohort in Turku underwent at baseline a brain3T MRI with T1-weighted sagittal sequences and FLAIR coronal sequences (Kemp-painen et al., 2017). The cortical thickness by region and brain region volumes wereattained using the Freesurfer image analysis suite (version 5.0.3). A measure of AD-type cortical thinning was calculated as an average of the entorhinal, inferior tem-poral, middle temporal, and fusiform regions (Jack et al., 2015a). Medial temporallobe atrophy (MTA) was assessed on the Scheltens scale (Scheltens et al., 1992) byone blinded specialist from T1-weighted images (Stephen et al., 2017b).

65

41 of the FINGER-PET participants had data available on IR- and DM-relatedblood markers. Fasting blood glucose, insulin, HbA1c, and a 12-item Bio-Plex ProHuman Diabetes assays were analysed. The assays included adiponectin, adispin,C-peptide, ghrelin, GIP, GLP-1, glucagon, insulin, leptin, PAI-1, resistin, and visfatin.The HOMA-IR was calculated based on insulin and glucose measures. These indi-viduals constituted the study cohort of Study IV (see Figure 3).

4.4 DISEASE STATE INDEX

The Disease State Index (DSI) is a machine learning algorithm designed to discrim-inate populations in terms of a condition. The discrimination is based on an indexvalue that—as the name implies—aims to represent the state of an underlying dis-ease based on a body of patient data. The DSI value is a continuous value allowinga more precise assessment of the disease state than a dichotomous algorithm would.The version of DSI used in this thesis classifies a subject as having a disease versus nothaving it, but newer versions allow classification into more than two categories. Thealgorithm was originally developed at the state-run VTT Technical Research Centreof Finland as a back end to a clinical decision support system, which allows a clin-ician to graphically examine operation of the algorithm in terms of different modelpredictors. The system has been further developed by Combinostics Ltd as part ofthe EU-funded PredictAD and PredictND tools, clinical decision support systems forAD diagnosis and differential diagnosis of dementia, respectively.

The DSI has previously been successfully used to discriminate between AD andCN (Mattila et al., 2011) and FTD and CN/MCI (Munoz-Ruiz et al., 2013), predictMCI–dementia conversion (Mattila et al., 2011, 2012b; Hall et al., 2015; Rhodius-Meester et al., 2016), and classify dementias based on structural MRI (Koikkalainenet al., 2016) and multimodally (Tolonen et al., 2018). The model has been previouslydescribed in detail by Mattila et al. (2011, 2012a). The algorithm is trained on a set ofindividuals with empirical predictor value distributions and binary outcomes. Fig-ure 4 shows example distributions of a predictor for positive and negative outcomecases. In respect to this pair of distributions, a fitness function is defined:

fitness(a) =LP (a)

LP (a) +RC(a)=

FN(a)

FN(a) + FP (a)

Here LP (a) is the left integral of the positive outcome distribution at a and RC(a)

is the right integral of the negative outcome distribution. These correspond to thecases with false negative and false positive predictions, respectively. The functionis monotonic with increasing values of a being assigned increasing values and amax

being assigned the maximal value of 1. Using the figure as a visual aid, it is intuitivelyeasy to see how the ratio of the red shaded area at the left side will grow in proportionto the sum of the shaded areas when a moves to the right. Each predictor is assigneda function fitnessi.

The predictors’ ability to discriminate between the outcomes varies and is re-

66

aFactor Value

Freq

uenc

y

LP(a)RC(a)NegativesPositives

Figure 4: Derivation of the fitness function from the empirical outcome distributions. LP (a)represents the left integral of the positive outcome distribution at a (false negative predictionat threshold level a) and RC(a) represents the right integral of the negative outcome distribu-tion (false positive prediction).

flected in the empirical distributions. The relevance of a predictor for the predictiontask is defined as

relevance(b) = max (0, LC(b) +RP (b)− 1)

= max (0, specificity(b) + sensitivity(b)− 1) ,

where b is the decision threshold for the factor, LC(b) is the left integral of the negativeoutcome distribution, andRP (b) is the right integral of the positive outcome distribu-tion at b. The integrals can readily be interpreted as the specificity and the sensitivityof the classifier, respectively. The decision threshold b denotes the value of the factorat which the fitness function reaches 0.5. Relevance assumes values in the range 0–1which is similar to the fitness function. For categorial variables the relevance is calcu-lated similarly, but the comparison of groups is limited to the individuals who sharethe same category value.

Predictor-specific fitness function values and relevance values are combined in themodel by weighting the function values with relevance values. The composite DSIvalue for an individual with its set of predictor values is defined as a weighted sumover each predictor i:

DSI =

∑i relevancei × fitnessi∑

i relevancei

Being an average of fitness values, DSI assumes values in the interval 0–1.The DSI can be calculated for the complete model as described above, but also for

a smaller subset of predictors or individual predictors. Conceptually linked predic-tors can be grouped together to assess their combined effect. Cardiovascular health,for example, can be modelled by combining blood pressure measurements, lipid val-ues, and smoking habits under one category.

67

4.5 DATA ANALYSIS AND PREDICTION MODELS

Differences between prediction outcome groups—that is, incidence of dementia andascertainment of pathology—were tested for statistical significance using the χ2 testfor categorical factors and the Mann-Whitney U test for continuous and ordered fac-tors. In the two FINGER studies (Studies III and IV) the Mann-Whitney U test wasused for all factors.

In the CAIDE model (Study I), the APOE genotype was modelled both in terms ofthe ε4 carrier status and as a variable describing genotype risk order ε2ε2 < ε2ε3 <

(ε2ε4 = ε3ε3) < ε3ε4 < ε4ε4 (Corder et al., 1993, 1994). In Vantaa 85+ (Study II), theAPOE genotype was modelled in four parallel ways by including ε2 and ε4 carrier-ships as binary factors, ε3 homozygousness as a binary factor, and all genotypes asa categorical factor. In the FINGER PET and IR/DM studies (Studies III and IV) theAPOE genotype was modelled simply as dichotomous ε4 carrier status.

In the FINGER-PET study (Study III), volumetric MRI measures were expressedin relation to the intracranial volume and bilateral measures were consolidated intoan average.

In order to analyse the relationship of different pathologies and dementia in theVantaa 85+ study (Study II), a dimension-reduction step was performed. Principalcomponents (PCs) were estimated for all dementia prediction cohort individuals,and for dementia and no-dementia individuals separately. PC loadings were usedto identify interrelationships between pathology types. The PC analysis was done onMATLAB R2015b.

4.5.1 Prediction and validation

Model predictors were identified from a group of candidate predictors by analyzingthe group mean value differences. This step reduces noise and generally improves theDSI prediction results. Given that the empirical predictor value distributions are closeto continuous and have approximately the same variance, no significant predictorshould be excluded based on this criterion. A p-value threshold of 0.05 for statisticalsignificance was set in the CAIDE and Vantaa 85+ studies. Additionally, the choice ofmodel building parameters was investigated using a spectrum of p-value thresholdvalues used to filter factors according to their significance. In the smaller-populationFINGER-PET prediction model all candidate factors were used.

Predictor grouping was utilized in the CAIDE and FINGER DSI models to formbroad groups for socio-demographic features, cardiovascular health, cognition, self-rated health measures, and MRI findings. A somewhat more granular approach wastaken in the Vantaa 85+ DSI model, in which groups were formed to gather all plasmalipid types together, for instance, and also to group parallel APOE genotype catego-rizations together. The grouping was held constant in all the Vantaa 85+ pathologyprediction models, but the predictor set varied according to the p-values in regardto the specific pathology. In the reduced-dimensionality pathology prediction modelthe unmodified principal component scores were used as singular predictors without

68

using the DSI.Prediction models built with the DSI were all internally validated. The CAIDE

model was trained and tested using cross-validation with the data divided into 50×5-folds, the Vantaa 85+ models using 10×10 cross-validation for both dementia andpathology, and the FINGER model using 100×5 cross-validation. Prediction perfor-mance against the binary outcome in each case was reported as the AUC. The AUCsare reported as mean values from the cross-validation folds, and dispersion valueswere also reported. The CAIDE DSI model was also validated against a linear-kernelsupport vector machine using the same data set.

4.5.2 Association analysis in FINGER IR/DM cohort

Logistic regression models were built to investigate the association of IR/DM mark-ers and Aβ positivity. The DM and APOE ε4 status were included as confounders,and blood marker concentrations were log-transformed. Statistical significance wasdetermined with correction for multiple comparisons using the false discovery ratemethod (Benjamini and Hochberg, 1995). All analyses were run on MATLAB R2017b,and function mnrfit was used for a logistic regression.

69

5 RESULTS

5.1 PREDICTING INCIDENT DEMENTIA

Both the CAIDE late-life cohorts and the Vantaa 85+ dementia prediction cohort un-derwent a prescreening for potential factors. Prediction models were built using theDSI and the prediction performance was assessed in a similar manner in both studies.

All factors crossing the 5% significance level threshold were included as predic-tors. All predictors selected for use in the CAIDE main model and the Vantaa 85+dementia prediction model are listed in Table 2 for contrast. The table lists AUC val-ues for predictors, predictor groups, and the complete model. Predictor-level data inthe extended population model are shown in detail in the original publication (StudyI).

5.1.1 Population characteristics, and predictors

A few key characteristics of the late-life CAIDE populations are presented in Table 1.More detailed characteristics are presented in the Study I original publication. Inboth the main and extended study populations, in the statistical testing, individu-als who developed dementia were significantly older, did worse on most subdomaincognitive tests, had poorer scores on the SMQ, and had a higher frequency of car-diovascular comorbidities and the APOE ε4 allele. In the main study population,individuals who developed dementia had also significantly a lower SBP and DBP,and had lower scores on three more SMQ questions. In the extended population,differences in cognitive testing results were more pronounced: MMSE aggregate andverbal expression subdomain scores were lower in individuals who developed de-mentia. As for the midlife–late-life changes, the BMI had on average increased by 1.6

Table 1: General characteristics of populations at baseline and frequency of outcome mea-sures.

CAIDE Vantaa 85+ FINGER

Maincohort

Extendedcohort

Dementiaprediction

Pathologyprediction

Aβ pre-diction

IR/DMcohort

N 709 1,009 245 163 48 41Baseline mean age 70.1 yr. 70.5 yr. 88.4 yr. 88.7 yr. 71.4 yr. 71.1 yr.Mean follow-up 8.3 yr. 9.0 yr. 5.6 yr. 4.1 yr. — —APOE ε4 carrier 32% 34% 21% 20% 30% 30%Diabetes mellitus 2% 3% 23% 28% 15% 15%Incident dementia 6% 15% 40% 36% — —Aβ positive share — — — 77% 42% 39%Key: Aβ amyloid beta protein, APOE apolipoprotein E, IR/DM insulin resistance and dia-betes mellitus

71

Table 2: Prediction results for incident dementia in the younger-old-age CAIDE populationand in the oldest-old Vantaa 85+ population.

AUC of dementia prediction

CAIDEmain studypopulation1

N=709

Vantaa 85+dementia

prediction cohortN=245

Complete model 0.79 0.73Age 0.67 —Education NA 0.60Cognitive testing 0.73 0.72

Executive functioning 0.68 NAEpisodic memory 0.64 NAProspective memory 0.62 NAPsychomotor speed 0.62 NAVerbal expression — NAMMSE Total — 0.71MMSE Calculation NA 0.60MMSE Orientation NA 0.64MMSE Other tasks NA 0.65MMSE Wordlist NA 0.68SPMSQ NA 0.71

Subjective Memory Questionnaire 0.64 NATotal score 0.62 NAForgetting phone numbers 0.61 NAForgetting names of actors 0.60 NAForgetting clothing size 0.59 NAForgetting midsentence 0.58 NA

Competence in Daily Activities NA 0.61Vascular factors 0.65 —

Systolic BP 0.63 —Diastolic BP 0.64 —Presence of comorbidity2 0.56 —

APOE genotype 0.59 0.58All genotypes modelled 0.60 0.58ε4 carrier 0.57 —ε2 carrier NA 0.56ε3ε3 genotype NA 0.57

“NA” indicates that the factor was not available and “—” indicates thatit was not accepted for the model after significance testing. Key: APOEapolipoprotein E, AUC area under the ROC curve, BP blood pressure,MMSE Mini-mental state examination, SPMSQ Short Portable Mental Sta-tus Questionnaire, — was not available as candidate predictor. Footnotes:1: The extended model additionally included MMSE total score and ver-bal expression score, 2: For differences in comorbid conditions see Table 2

72

kg/m2 in the no-dementia group in both populations and remained more or less onmidlife levels in the incident dementia group, i.e. −0.1 and 0.2 kg/m2 in the respec-tive main and extended populations. These differences were statistically significant.The SBP and DBP had both decreased significantly in the main population, as did theDBP in the extended population. The total cholesterol was lower in late-life in theextended population.

The Vantaa 85+ dementia prediction cohort has been characterized in detail in theoriginal publication (Study II). A brief overview is presented in Table 1. The groupdeveloping dementia during the follow up was less educated, scored lower on theMMSE and all its derivative subscores, made more errors on the SPMSQ, and hada different distribution of APOE genotypes. No differences were observed in age,cardiovascular factors, depression, or BMI.

5.1.2 Dementia prediction in the younger old (CAIDE)

The prediction performance in the late-life CAIDE cohort as measured by AUC was0.79 in the main population and 0.75 in the extended population in cross-validation.Receiver operating characteristics (ROC) curves are shown in Figure 1. Comparisonof group level AUCs in the two models are shown in Table 3, where also results froma separate analysis with midlife–late-life changes in vascular factors are presented.Noncross-validated complete-model AUCs were 0.84 and 0.76, respectively. Cogni-tive testing as a category was the best predictor in both models at a respective 0.73and 0.69, while not reaching the performance of the complete model. Other predic-tor types did improve the model beyond that achieved for cognition. Age was thesecond best performing predictor. Subjective memory assessments performed worsethan objective cognitive testing. Vascular factors did have some predictive power inthe main population, but practically none in the extended population. The APOEgenotype had poor predictive power. Changes in vascular parameters from midlifeperformed somewhat better than cross-sectional values.

The DSI models were also investigated in binary prediction using different indexcut-off values. As an example from the more comprehensive table in the originalpublication, a DSI threshold of 0.5 for positive prediction resulted in 0.74 accuracy,0.73 sensitivity, and 0.74 specificity in the main population, and 0.67, 0.69, and 0.67,respectively, in the extended population. Results for these statistics were better inalmost every case in the main population.

The CAIDE model results were validated in terms of the method used. A paral-lel SVM was set up using the same population data and cross-validation principles.Using the MATLAB fitcsvm function, parameters were set empirically for the bestperformance. The SVM achieved an AUC score of 0.77 in the main population and0.74 in the extended population.

Furthermore, the choice of model building parameters was investigated. Table4 of the original publication (Study I) lists a spectrum of p-value threshold valuesused to filter factors according to their significance. The results demonstrated that alaxer requirement and a larger predictor set resulted in lower performance, as did a

73

Table 3: Performance in the prognostic prediction of dementia in the CAIDE younger-old pop-ulations.

AUC (95% confidence interval)

Main studypopulation

Extended studypopulation

Complete model 0.79 (0.79-0.80) 0.75 (0.74-0.75)Age 0.67 (0.65-0.68) 0.66 (0.66-0.67)Cognitive testing† 0.73 (0.73-0.74) 0.69 (0.69-0.70)Subjective Memory Questionnaire† 0.64 (0.63-0.66) 0.58 (0.57-0.58)Vascular factors† 0.65 (0.64-0.66) 0.53 (0.52-0.53)APOE genotype† 0.59 (0.58-0.60) 0.60 (0.59-0.61)

Complete model with vascular changes‡ 0.80 (0.79-0.81) 0.78 (0.77-0.79)Vascular changes 0.68 (0.66-0.69) 0.65 (0.64-0.66)

Change in systolic BP 0.65 (0.63-0.66) —Change in diastolic BP 0.61 (0.59-0.62) 0.61 (0.59-0.62)Change in BMI 0.68 (0.67-0.69) 0.68 (0.67-0.69)Change in total cholesterol — 0.55 (0.54-0.57)

Key: APOE apolipoprotein E, AUC area under the ROC curve, BMI body massindex, BP blood pressure, MMSE Mini-mental state examination, — not in-cluded after significance filtering. Footnotes: †: Only group-level result shown,for individual factors in the main model see Table 2; ‡: The complete modelalso includes all predictors used in the upper panel of the table in the respec-tive populations.

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

1 − Specificity (False Positive Rate)

Sens

itivi

ty (T

rue

Posi

tive

Rat

e)

Main population (AUC 0.79)Extended population (AUC 0.75)

Figure 1: Receiver operating curves for prediction of incident dementia in the CAIDE popula-tions.

74

very small high-relevance predictor model. No requirement for significance led to anAUC score of 0.74 in the main population. However, the predefined threshold of 5%was not the optimal value in either population. Exploratory analyses were conductedto test nonlinear tail effects for BMI, blood pressure, and cholesterol. Dichotomousvariables for crossing extreme distribution tail values were added to the model, butnone of the tested cut-off values showed an effect on the overall performance.

5.1.3 Dementia prediction in the older old (Vantaa 85+)

Prediction results for dementia in the Vantaa 85+ study are shown in Table 2. In thisconsiderably older population the dementia incidence was much higher, but the pre-diction performance was lower (AUC 0.73 vs. CAIDE 0.79). Here, too, an objectiveassessment of cognition was the best predictor, practically on a par with the perfor-mance in the CAIDE (AUC 0.72 vs 0.73 in CAIDE main model). MMSE and SPMSQwere equally good predictors. All other predictor modality groups performed in therange 0.58–0.61 or were not selected for the model in the first place. As for the parallelAPOE predictor representations, ε2 had a higher prevalence in the incident-dementiagroup and indicated increased risk, and ε3ε3 was enriched in the no-dementia groupand indicated a protective effect. However, both had a poor predictive performance.

5.1.4 Dementia and neuropathology at death

In addition to the prognostic dementia prediction model, an exploratory diagnosticprediction analysis was performed using pathology findings to predict dementia atthe time of death. Table 4 presents the characteristics of pathology in terms of havingdementia versus not having dementia. AD-pathology was significantly more preva-lent in the dementia group, as was CAA, HS, TDP-43 protein, and cortical macroin-farcts, but not macroinfarcts elsewhere or microinfarcts. α-synuclein pathology didnot show any significant differences.

The results from a PCA on neuropathology are presented in Table 5. The three firstPCs are shown for the entire population and subpopulations without dementia andwith dementia at death. The three PCs explained together 56–59% of the variance inthe data in each population. The first PC had strong loadings concerning both ADneuropathological findings and CAA. In the dementia group there were also strongnegative loadings for all/cortical macroinfarcts. The first PC could be interpreted as“AD-type pathology”.

The second PC had a different loadings profile for the dementia and no-dementiagroups. In the entire population and in the no-dementia group PC2 had a strongpositive loading for all/cortical macroinfarcts and weaker positive loadings for WMmacroinfarcts and α-synuclein. In this subpopulation the second PC could be in-terpreted as “Vascular pathology”. In the dementia group, PC2 had large positiveloadings for most AD-type and vascular pathologies and negative loadings for HSand TDP-43. This PC could be interpreted simply as reflecting age, as both HS andTDP-43 occur predominantly in the very old whereas the other pathologies do occuralso in earlier old age.

75

Table 4: Neuropathology characteristics at autopsy according to the dementia status for par-ticipants without dementia at baseline in Vantaa 85+.

No dementia(N=104)

Dementia(N=59)

p-value

β-amyloid load 74 (71%) 52 (88%) 0.020Tangle count 55 (53%) 44 (75%) 0.008Neuropathological AD† 39 (38%) 38 (64%) 0.001CAA‡ 59 (58%) 44 (76%) 0.040Cerebral macroinfarcts 47 (45%) 33 (56%) 0.200Cortical macroinfarcts 23 (22%) 24 (41%) 0.020WM macroinfarcts 14 (14%) 9 (15%) 0.800Cerebral microinfarcts‡ 16 (16%) 11 (19%) 0.700α-synuclein 26 (25%) 22 (37%) 0.100Hippocampal sclerosis 2 (2%) 9 (15%) 0.002TDP-43 8 (8%) 14 (24%) 0.007Values are shown as absolute numbers (percentages). The p-valueis calculated with the Fisher’s exact test. Footnotes: †: Definedbased on the National Institute on Aging–Alzheimer’s Associationcriteria (Hyman et al., 2012) using the combination of Braak andCERAD scores, and dichotomized as present (intermediate or highlikelihood of AD) vs. absent (low likelihood of AD); ‡: 4 partici-pants missing data

The third PC in the dementia group had combined high positive loadings for HSand TDP-43 and a large negative loading for α-synuclein. In the no-dementia groupPC3 was driven by the tangle count and had lower negative and positive loadings forother types of pathology.

PC1 had moderate predictive power for the diagnostic prediction of dementiawith an AUC of 0.71. PC2 and PC3 had AUCs of 0.60 and 0.54, respectively.

5.2 PREDICTING BRAIN PATHOLOGY

5.2.1 Longitudinal prediction of pathology (Vantaa 85+)

An overview of the Vantaa 85+ pathology prediction cohort is given in Table 1 (p. 71),and more details are included in the original publication (Study II). The cohort wason average 88.7 years old, 19% were male, and the mean education duration was4.1 years. The predictors for each type of pathology crossing the 5% significancethreshold are listed in Table 3 of the original publication.

Amyloid and tau related pathology

The APOE genotype was included as a predictor for all amyloid and tau relatedpathologies, namely the Aβ load, tau tangle count, CAA, and neuropathologicalAD which is defined here as an intermediate or high likelihood of AD based on theNIA-AA criteria (Hyman et al., 2012). Additionally, impairment in daily activities

76

Table 5: Principal components of pathology and their performance in the prediction of demen-tia in Vantaa 85+.

Groups by outcome

Prediction cohort No dementia Dementia

PC1 PC2 PC3 PC1 PC2 PC3 PC1 PC2 PC3

Explained variance 25% 20% 11% 26% 19% 12% 24% 21% 14%AUC to predict dementia 0.71 0.60 0.54 — — — — — —

β-amyloid load 41 3 -31 47 6 -43 16 25 24Tangle count 48 0 38 44 15 68 37 28 8Neuropathological AD† 59 -7 -4 54 6 5 51 33 20CAA 47 -3 -18 51 -2 -33 28 37 -9All cerebral macroinfarcts -1 73 -13 -11 77 -11 -48 47 6Cortical macroinfarcts 7 59 -15 -2 51 -19 -43 49 8WM macroinfarcts -8 24 1 -12 25 2 -19 9 8Cerebral microinfarcts 11 10 16 6 7 31 2 27 17α-synuclein 5 20 81 2 23 31 -10 9 -53Hippocampal sclerosis 0 4 -2 4 4 0 -17 -22 46TDP-43 protein 6 1 2 4 -1 3 -7 -14 59PC loadings expressed in percentages. Key: AUC area under the ROC curve, CAA cerebralamyloid angiopathy, WM white matter. Footnotes: †: Defined based on the National Insti-tute on Aging–Alzheimer’s Association criteria (Hyman et al., 2012) using a combinationof Braak and CERAD scores, and dichotomized as present (intermediate or high likelihoodof AD) vs. absent (low likelihood of AD)

predicted a higher Aβ load; a higher total cholesterol and LDL predicted a highertangle count; a subjective memory decline predicted a higher tangle count and neu-ropathological AD; a lower social class predicted neuropathological AD; and havingno cardiovascular comorbidity and male sex predicted the presence of CAA.

Prediction AUCs for the four pathologies were in the range 0.64–0.68 with neu-ropathological AD having the highest value. The APOE genotype was modelled inall cases using multiple parallel presentations, and the effects of the alleles varied ac-cording to the pathology. The APOE category AUCs were in the range 0.60–0.65. Theε4 allele was predictive of all pathology types, the ε2 allele was protective against anAβ load and neuropathological AD, and the ε3ε3 genotype was protective againsta high tau tangle count and CAA. All other predictors had a poor predictive perfor-mance with AUCs below 0.62.

Vascular pathology

More predictors for cerebral macroinfarcts were identified than for microinfarcts, andthe prediction results were better. Macroinfarcts overall were predicted by the pres-ence of a cerebrovascular comorbidity, lower MMSE total score and wordlist sub-score, higher BMI, and impairment in daily activities. The predictors varied some-what by region, but cerebrovascular comorbidities were predictive in every case.Cortical macroinfarcts were predicted by the APOE ε4 allele, and the ε3ε3 genotype

77

was protective. The presence of WM macroinfarcts was predicted by a ε2 carriershipand both low HDL and LDL. Cerebral microinfarcts were predicted only by a lowerduration of education.

The prediction performance for WM macroinfarcts at AUC 0.76 was better than forany other vascular pathology. The AUC was 0.71 for cortical macroinfarcts and 0.72for macroinfarcts overall. Cholesterol was a strong predictor for WM macroinfarcts(group level AUC 0.72). The APOE was a somewhat weaker predictor for vascularpathology than for amyloid and tau related pathologies (at the group level 0.60–0.61vs. 0.60–0.65). Other predictors had AUCs in the range 0.59–0.64.

Other pathology

For α-synuclein pathology no significant predictors were found. HS was predictedby a lower MMSE total score, wordlist and other task subscores, and by being a cur-rent smoker. These predictors had good predictive performance at AUC 0.78, andcognition was the stronger modality (group level AUC 0.75). The deposition of TDP-43 was predicted by having fewer depressive symptoms, and the performance wasmoderate (AUC 0.69).

5.2.2 Diagnostic prediction of brain amyloid (FINGER)

Key characteristics of the FINGER-PET study are summarized in Table 1 (p. 71) forcomparison with the other studies. Results from diagnostic prediction of in vivoamyloid positivity (Aβ+) are presented in Table 6. Twenty individuals (42%) were as-sessed Aβ+ at baseline imaging. Aβ+ individuals had statistically significantly (95%confidence level, p-values not corrected for multiple comparisons) higher frequencyof the APOE ε4 allele, a lower executive functioning score, and more neurodegen-erative changes on MRI. Volumes were significantly lower in the cortex and greymatter overall, as well as for the cerebellar cortex, thalamus, putamen, hippocampus,amygdala, accumbens area, and ventral diencephalon. The MTA on the Scheltensscale was more pronounced in the Aβ+ group. Sociodemographic factors, vascularfactors, overall cognition, or any of the cognitive subdomains showed no significantdifferences.

All factors were included in the model. The prediction AUC of the completemodel was 0.78 in cross validation and 0.88 without cross validation. Single-predictorAUCs were in the range 0.45–0.75. MRI was the best performing predictor category atAUC 0.75. Volumetric FreeSurfer estimates as a group (AUC 0.72) and a visual MTA(0.71) performed equally at a moderate level, and the AD-specific cortical thicknessmeasure performed worse (0.65). The APOE and the executive functioning subdo-main score also had some predictive power at 0.69 and 0.69 each. Cognition as acategory did worse than the executive functioning score on its own. The BMI wasa stronger predictor than hypertension in the cardiovascular category, which per-formed poorly at an AUC of 0.60. Sociodemographic factors also lacked predictivepower.

78

Table 6: FINGER-PET diagnostic prediction results for Aβ and outcome-group mean values.

Group mean

AUC (95% CI) Aβ- (N=28) Aβ+ (N=20) p-value

Complete model 0.78 (0.65–0.91)Sociodemographic 0.54 (0.37–0.70)

Sex (female) 0.48 (0.35–0.60) 14 (50%) 8 (40%) 0.505Age 0.45 (0.28–0.61) 70.2 71.6 0.310Education (years) 0.59 (0.43–0.75) 9.7 8.9 0.320

Cardiovascular 0.60 (0.46–0.75)Body mass index 0.65 (0.50–0.79) 28.9 26.2 0.088Hypertension 0.49 (0.37–0.61) 10 (36%) 9 (45%) 0.529

APOE ε4 carrier† 0.69 (0.56–0.82) 4 (14%) 10 (53%) 0.005Cognition 0.65 (0.49–0.81)

Total score 0.55 (0.38–0.72) 0.04 -0.09 0.421Memory 0.54 (0.38–0.70) -0.11 0.04 0.385Processing speed 0.57 (0.41–0.73) 0.16 -0.10 0.184Executive function 0.69 (0.53–0.84) 0.16 -0.22 0.026

Magnetic resonance imaging 0.75 (0.61–0.89)Volumes (% of ICV) 0.72 (0.57–0.88)

Total cortex 0.73 (0.59–0.88) 0.29 0.27 0.007Total grey matter 0.72 (0.57–0.88) 0.39 0.36 0.009Cerebellum cortex 0.69 (0.54–0.84) 0.063 0.059 0.027Thalamus proper 0.70 (0.55–0.85) 9.3E-3 8.4E-3 0.022Caudate 0.65 (0.49–0.81) 4.9E-3 4.5E-3 0.070Putamen 0.71 (0.56–0.87) 7.0E-3 6.1E-3 0.014Pallidum 0.61 (0.45–0.77) 1.9E-3 1.8E-3 0.198Brain Stem 0.61 (0.45–0.77) 0.014 0.014 0.229Hippocampus 0.70 (0.54–0.86) 5.2E-3 4.6E-3 0.019Amygdala 0.69 (0.53–0.85) 2.3E-3 2.0E-3 0.030Accumbens area 0.75 (0.62–0.89) 6.6E-4 5.6E-4 0.004Ventral diencephalon 0.68 (0.53–0.83) 5.0E-3 4.7E-3 0.037Cerebrospinal fluid 0.61 (0.44–0.78) 8.8E-4 8.1E-4 0.171Optic chiasm 0.60 (0.41–0.78) 1.4E-4 1.2E-4 0.164Total corpus callosum 0.62 (0.45–0.79) 2.0E-3 1.7E-3 0.058

Visual MTA (Scheltens) 0.71 (0.59–0.84) 1.0 1.6 0.007AD cortical thickness (mm) 0.65 (0.48–0.82) 2.8 2.8 0.084

The Wilcoxon rank sum test was used to calculate p-values for all variables. AUC values fromcross validation. Key: Aβ amyloid beeta protein, APOE apolipoprotein E, AUC area under theROC curve, BP blood pressure, CI confidence interval, ICV intracranial volume, MTA medialtemporal lobe atrophy. Footnotes: †: One Aβ+ person was missing data

79

A pragmatic analysis on the value of the different predictor modalities was alsoconducted and is shown in detail in Table 4 of the Study III manuscript. A combina-tion of modalities that require no specialized equipment—demographic information,cardiovascular data, and cognitive measures—predicted amyloid positivity at AUC0.62. The APOE genotype or MRI could each improve this to AUC 0.71–0.72. Asimple model using only the APOE genotype and a visual MTA assessment jointlypredicted amyloid at AUC 0.81, a result superior to that of the complete model.

5.2.3 Associations between biomarkers of DM and brain amyloid (FINGER)

The FINGER IR/DM cohort of 41 participants were on average 71.1 years old, 39%were Aβ+, and 15% had DM (Table 1, p. 71). The frequency of DM or BMI did not dif-fer in the outcome groups, but APOE ε4 allele was more frequent in the Aβ+ group(56% vs. 12%). Table 7 presents the mean concentrations of the biomarkers in theAβ- and Aβ+ groups in the left panel. The insulin plasma concentration was statisti-cally significantly lower in Aβ+ individuals at the 95% confidence level before correc-tion for multiple comparisons. Differences in the insulin-related measures C-peptideconcentration—cleaved during insulin production—and HOMA-IR—a derivative in-dex value–were significant only at the 90% confidence level. The plasminogen ac-tivator inhibitor-1 (PAI-1) concentration was lower in Aβ+ individuals at the 95%confidence level. Other biomarkers showed no significant differences.

Logistic regression models were built iteratively for the metabolic markers us-ing different sets of potential confounders. The final models were estimated usingthe DM status and APOE ε4 carrier status as confounders. The coefficient of theAPOE genotype was significant in all models, and that of the DM status was not inany model. Model coefficients for all markers are presented in Table 7 in the rightpanel. The linear regression model equation is included in the table legend. Beforecorrection for multiple comparisons, coefficients of C-peptide, insulin, PAI-1, andHOMA-IR were significant. The coefficients indicated higher IR and elevated PAI-1to be associated with lowered odds of Aβ+. After correction these four markers weresignificant only at the 90% confidence level. Models with either BMI, age, or sex asadditional confounders showed a similar pattern, and no differences in significanceafter correction were observed.

80

Table 7: Population characteristics and logistic regression coefficients in the FINGER IR/DMpopulation.

Meanconcentrationa

Logistic regression model

Metabolic marker Aβ- Aβ+ Bb (95% CI) p-value

C-peptide (103 pg/ml) 1.31 ∗ 0.95 -5.7 (-10.4 – -1.1) 0.016†

Ghrelin (103 pg/ml) 1.57 1.55 0.1 (-6.1 – 6.3) 0.972GIP (103 pg/ml) 0.29 0.29 -1.5 (-5.4 – 2.3) 0.436GLP-1 (103 pg/ml) 0.59 0.58 0.0 (-8.8 – 8.8) 0.998Glucagon (103 pg/ml) 1.07 1.00 -2.1 (-11.3 – 7.0) 0.646Insulin (103 pg/ml) 0.27 ∗∗ 0.17 -4.5 (-8.3 – -0.8) 0.017†

Leptin (103 pg/ml) 7.55 6.06 -1.6 (-4.1 – 0.8) 0.191PAI-1 (103 pg/ml) 5.31 ∗∗ 4.16 -13.3 (-24.0 – -2.6) 0.015†

Resistin (103 pg/ml) 2.22 2.03 -3.7 (-10.1 – 2.8) 0.266Visfatin (103 pg/ml) 4.83 4.43 -2.0 (-6.8 – 2.7) 0.401Adiponectin (106 pg/ml) 5.45 6.03 -0.3 (-2.3 – 1.8) 0.808Adipsin (106 pg/ml) 1.21 1.45 1.1 (-2.1 – 4.2) 0.500fP-Glucose (mmol/l) 5.92 6.30 4.8 (-9.3 – 18.9) 0.505B-HbA1c (mmol/mol) 36.72 37.25 15.0 (-10.9 – 40.9) 0.258HOMA-IR (mmol·mU/l2) 2.06 ∗ 1.33 -4.5 (-8.3 – -0.7) 0.019†

Regression: ln(YAβ+/YAβ−) = C+BDMXDM+BAPOEXAPOE+BX log(X)Key: Aβ amyloid beta protein, GIP Gastric inhibitory polypeptide, GLP-1Glucagon-like peptide-1, PAI-1 Plasminogen activator inhibitor-1, HbA1cGlycated hemoglobin, HOMA-IR Homeostatic Model Assessment for In-sulin Resistance. Footnotes: a: Group differences tested for significanceusing the Mann-Whitney U test, ∗ for significance at 10% confidence level,∗∗ for 5% ; b: Coefficient of log-transformed value.

81

6 DISCUSSION

6.1 PREDICTION OF INCIDENT DEMENTIA IN THE YOUNGEROLD

Dementia prediction up to ten years into the future in the CAIDE general popu-lation samples of cognitively healthy individuals succeeded well (AUC 0.75–0.79).The mean ages were 70.1 and 70.5 years in the main and extended populations, re-spectively. The internally validated performance was on par or slightly better thanthe published values for the multimodal models identified in section 2.8.4 (p. 48).The externally validated performance is in general lower, and that would be ex-pected for the CAIDE DSI models also. In age-matched general population cohorts(60 yr.<age<80 yr.), the externally validated performance for previously publishedprediction models was in the range of AUC 0.68–0.89 in follow-up studies of 3–6years. For longer follow-up times of 8–10 years in DM populations and 21 years in ageneral population the AUC was 0.75 in all cases. None of the externally validatedlate-life prediction models included APOE genotype data, and no factors describinglongitudinal change in any modality were included apart from fasting glucose varia-tion in a DM cohort.

Interestingly, the best performing model used a free recall score as a solitary pre-dictor for incident dementia over a 3–5-year period. The score achieved AUC 0.89over a four-year period in a 70-year-or-older cohort with subjective memory com-plaints (Derby et al., 2013). Another validation study had similar results in a popu-lation with no memory complaint requirement (Mura et al., 2017). In those studies,adding sociodemographic predictors or APOE did not significantly improve the re-sults. Results from the CAIDE populations mirror this in that cognitive testing resultswere the best predictors. However, adding other modalities did improve predictionresults. Age was also a relatively strong predictor, as would be expected given itswell-established status as a risk factor.

Another noteworthy model with published results used only data from UK healthregistries with very good results (AUC 0.84; Walters et al., 2016). The model re-lied on recorded data only, and the authors suspected dementia diagnoses to beunder-recorded, possibly lowering performance. One would also assume that healthrecords would accrue more rapidly for individuals with health problems and de-mentia risk factors, possibly leading to an enrichment of the dementia risk in thestudy population. The reported AUC is in any case surprisingly high compared toother similar prediction models. The CAIDE extended population also utilized publichealth records, but not to infer predictor data, only to establish dementia diagnoses.The sensitivity of the Finnish hospital and drug prescription registries is in the range62–71% for AD and/or dementia, showing an underrepresentation of dementia ashypothesized by Walters et al. (2016). The prediction results were overall poorer in

83

the extended model, but the relative importance of predictor modalities remained thesame. Only the vascular measurements category did markedly worse in comparison.The extended population may include individuals with poorer general health, whowere not able to take part in the second late-life visit. They may have had a higherdementia incidence due to risk factors, or they may have died at a younger age beforedementia onset. The dementia incidence was higher in the extended population (15%vs. 6%). Individuals who died without a dementia diagnosis in health registries hadto be excluded from the extended model because the DSI cannot account for disease-free survival time as a variable but can only take into account dementia status at theend point.

The APOE genotype, in predicting dementia, performed worse than any othermodality. For the midlife CAIDE risk score, the APOE offered a small improvementin predictive power. Some prior models have included APOE information as it hasdemonstrated a benefit in prediction, whereas other newer genetic markers have of-fered little additional benefit (Tang et al., 2015). The APOE ε4 prevalences of 32%(main population) and 34% (extended population) were roughly in accordance withpreviously published prevalence estimates of around 33–42% for North Europeanmiddle-aged subjects and 17% for centenarians (Norberg et al., 2011). The FINGERpopulation had a similar ε4 prevalence.

Parameters of vascular health did not feature in the CAIDE late-life models asprominently as they did in the midlife models. BP measurements were includedin the main model but not in the extended model. Both populations, however, in-cluded predictors that quantify the change from midlife to the late-life predictionbaseline. Change in the BMI—that is, less weight gain in the dementia group—was abetter predictor of incident dementia than change in BP, or any cross-sectional vascu-lar measure. The presence of cardiovascular comorbidities did not predict dementiawell, which may in part be due to the fact that only conditions severe enough to berecorded in the Hospital Discharge Register were included. This can also explain therelatively low recorded prevalence of DM in the CAIDE populations—2% and 3% inthe respective main and extended populations.

These results in dementia prediction in the younger old highlight the potential foridentifying individuals who are most at risk of developing dementia. These are theindividuals who would benefit the most from targeted interventions. The model wasinternally validated as per current guidelines (Collins et al., 2015), and the use of ageneral population sample will support good generalizability in external populationsin the future. The DSI prediction model also shows which risk factors are important ata population level. The tool also allows for an analysis of risk profiles of individuals.Such a feature could be useful in a clinical setting when highlighting or targeting anindividual’s most relevant risk factors.

84

6.2 PREDICTION OF INCIDENT DEMENTIA IN THE OLDEST OLD

The Vantaa 85+ dementia prediction cohort was almost twenty years older than theCAIDE late-life populations, and the sample was highly representative of the localage cohort. 98% of all eligible residents participated initially. Despite the high base-line mean age of 88 years, a mean follow-up time of 5.6 years was achieved. TheAPOE ε4 allele prevalence was in line with published Finnish population estimates.The Dementia incidence was high compared to the CAIDE populations at 40% pera mean of 6 years follow up versus 6–15% per a mean of 8–9 years, which is to beexpected in this age cohort (Gardner et al., 2013).

The dementia prediction performance overall (AUC 0.73) was weaker than for thetwo younger CAIDE populations (AUC 0.75–0.79), but better than for the only age-matched validated prediction model identified in section 2.8.4 (p. 48). The health-registry-based model by Walters et al. (2016) had practically no predictive power inan 80+ population (AUC 0.56), although that model lacked cognitive measures. Theauthors attributed the poor performance partly to the lack of routine health check-ups and resulting lack of registry entries in that age cohort. The results for youngercohorts were generally better than in the Vantaa 85+ study.

Measures of cognition were the best predictors of incident dementia for the Van-taa 85+ cohort—which is analogous to the CAIDE models. However, other modali-ties added little to the performance of cognitive questionnaires (cognition group AUC0.72). A lower duration of education and low competence in daily activities were pre-dictive of dementia, but had a clearly lower level of performance. Importantly, agewas not a predictor of dementia in this age group. This is contrary to what would beexpected, as the incidence in this age group is high and even relatively small base-line age differences could potentially be reflected in differences in the incidence rates.The FINGER eligibility criteria may have affected this. Additionally, vascular healthwas not predictive. This contrasts with midlife prediction models, and also with theCAIDE models, in which cardiovascular measurements and especially changes inthose measurements were predictive. A large autopsy study by Jellinger and Attems(2010) may help explain this finding. The relative prevalence of a pure form of VaDat death was shown to decrease with increasing age from age 60 to 90+, and the rel-ative prevalence of AD and mixed-AD pathology was shown to increase. Moreover,the prevalence of pure AD was found to decrease after the age of 90. This findingsupports the pattern seen in the Vantaa 85+ study. The mechanisms leading to vas-cular brain pathology may be relatively less important than in younger age groups.The clinical phenotype may perhaps be dominated by AD-type pathology and itscombined effect with other brain pathologies.

The APOE ε2 allele was predictive of incident dementia. In younger populationsε2 is thought to be protective. The status of the ε4 allele as a poor predictor wasexpected based on previous studies in the very old (Juva et al., 2000; Corrada et al.,2013), but the outright negative effect of ε2 has not been shown before. The ε2 allelehas been shown not to be protective of dementia in the oldest of the old in shorterfollow-up studies (Skoog et al., 1998; Juva et al., 2000; Qiu et al., 2004), and one study

85

demonstrated a deleterious effect on the VaD incidence (Skoog et al., 1998). In theVantaa 85+ cohort, homozygous ε3 genotype was protective of dementia. The nega-tive effect of the ε2 allele may be explained by findings in the Vantaa 85+ pathologyprediction cohort as discussed in section 6.4.1.

6.3 PREDICTION OF BRAIN AMYLOID AND AD-TYPE PATHOL-OGY

6.3.1 Frequency of amyloid beta, APOE ε4, and dementia

Two studies had available data for predicting brain Aβ accumulation. The Vantaa85+ study focused on a general population sample with a mean age of 89 years,whereas the FINGER-PET population consisted of at-risk individuals from a gen-eral population of around 71 years of age. Nominally the FINGER prediction modelwas diagnostic and the Vantaa 85+ model prognostic, but the mean follow-up timeof 4 years was short in the context of AD-type pathology. Aβ and tau pathology de-velop during a period of up to decades (Jack et al., 2013), and most of the AD-typepathology observed at the end of the Vantaa 85+ follow up was probably present atthe baseline. For this type of pathology, the model more likely represented a mixeddiagnostic/prognostic model. In-vivo PET imaging of amyloid has had good con-cordance with neuropathologically determined Aβ positivity. The sensitivity of thevisual determination of amyloid positivity in PET imaging is 92–98%, and specificity98–100% when using pathology as a gold standard (Clark et al., 2012; Sabri et al.,2015). The dementia incidence of 36% in the Vantaa 85+ pathology prediction pop-ulation is in line with previously published estimates of 18–38% in this age cohort(Gardner et al., 2013).

The prevalence of Aβ pathology at the time of autopsy was more similar to anAD-dementia population than an old-age CN population. The Vantaa 85+ cohortwas on average 93 years old at autopsy and 77% were Aβ+. For the 80–90-year-oldcohort, the in-vivo Aβ prevalence estimate is 33–59% for CN individuals, 60–71% forindividuals with MCI, 79–84% for individuals with AD-dementia, and 36–50% fora VaD cohort (Jansen et al., 2015; Ossenkoppele et al., 2015). The Aβ incidence in-creases steeply after approximately the age of 70 especially in ε4 carriers (Jack et al.,2015b). Many participants in the Vantaa cohort were older than the range of ages forwhich these prevalence estimates have been published. Additionally, a significantportion of the nondementia group may have had MCI, for which the participantswere not tested. These facts, and methodological differences between the Vantaa 85+post-mortem assessment and the in-vivo ascertainment of Aβ+ used in the referencestudies may explain the observed difference in Aβ prevalence. In the FINGER cohortwith a mean age of 71 years, the Aβ prevalence was 42%, which is high in compari-son to the previously published estimate of 16–33% for CN 60–80-year-olds (10–28%noncarriers and 29–68% carriers; Jansen et al., 2015). The prevalence range estimatefor individuals with subjective cognitive impairment is 17–35%, and for MCI 37–60%(Jansen et al., 2015). The Aβ prevalence of the FINGER-PET population was closer to

86

that of an MCI population. The effect is probably due to the FINGER recruitment cri-teria, which may have enriched the participant population with subclinical AD-typepathology.

The APOE ε4 prevalence of 21% in the Vantaa 85+ cohort at the baseline is roughlyin agreement with previously published data on APOE genotype frequencies in North-ern Europe (Norberg et al., 2011). A noncarrier survival effect has been observed inFinnish centenarians, among which the APOE ε4 frequency was 8% and the ε2 fre-quency was enriched to 7% (Louhija et al., 1994). In the Vantaa 85+ cohort, 16% wereε2 carriers. The surviving APOE ε4 carriers in this age group were not subjected tothe same elevated risk of incident dementia as younger carriers, although the preva-lence among carriers remains high (Corrada et al., 2013; Gardner et al., 2013). TheAPOE ε4 prevalence in the FINGER cohort at 30% was in line with previously pub-lished estimates (Norberg et al., 2011).

6.3.2 Amyloid beta prediction

The overall prognostic amyloid prediction performance was 0.66 in the Vantaa 85+study, rated “poor” using terminology by Hosmer et al. (2013). Previously publisheddiagnostic models indicate better prediction performance in younger cohorts. Theone study identified in section 2.8.5 with an older cohort used age, sex, family historyof dementia, subjective memory complaint, APOE, and a global cognitive score aspredictors and achieved an AUC value of 0.70 (Mielke et al., 2012). Considering thatno report of cross-validation being used was found, the performance of that modelis probably on par with the Vantaa 85+ model. Models with fewer modalities hadpoorer performance. The predictors chosen for these models consist of risk factorsthat have been studied in younger age groups, and they may not be equally relevantin older cohorts. A similar effect has been demonstrated in the case of dementia riskscores, which have not performed as well outside their assigned age cohorts.

The prediction of Aβ+ on PET in a younger cohort produced better results. Per-formance of the complete FINGER-PET model including structural MRI achieved anAUC value of 0.78, and 0.71 without MRI. Both could be considered “acceptable” asper the criteria by Hosmer et al. (2013). Two models identified in section 2.8.5 re-ported AUCs on models in a CN population: these were a model by Mielke et al.(2012), and a model in a somewhat younger population (>50 yr.) that included MRI(AUC 0.74 cross-validated; ten Kate et al., 2018). Performance in FINGER-PET studywas somewhat better, although the population was much smaller (48 vs. 483 for theformer and 337 for the latter) possibly leading to overfitting and impairing its gen-eralizability. The AUCs of models involving MCI populations were typically greaterthan 0.80. These results in the younger old cohort show the potential for such modelsin identifying Aβ+ individuals even at a pre-MCI stage. A prediction model like thiswould facilitate the identification of populations with a considerably higher preva-lence of Aβ+, thus reducing the number of invasive, time-consuming, and costlyassessments during the screening process of a clinical trial, for example.

The APOE genotype is one of the two identified predictors of Aβ pathology in the

87

oldest old, competence in daily activities being the other with a similar AUC value.Prediction performance of the ε4 allele was better in the younger, late-life FINGERpopulation (0.69 vs. 0.60). The ε2 allele was protective against Aβ accumulation inthe Vantaa 85+ study, a finding which was analogous to another study of the oldestold (Berlau et al., 2013). Aβ positivity has been observed to rapidly increase after theage of 70 in ε4 carriers while maintaining more or less the same rate of increase innoncarriers (Jack et al., 2015b), which would imply that the prediction performanceof the APOE genotype should improve with advancing age. Findings in the twostudies of this thesis contradict this, possibly due a ε4 noncarrier survival effect inthe population of the oldest old.

Cognition was not predictive of Aβ in the 85+ population, and only the executivefunction subdomain was predictive in the FINGER study. Lower cognitive scoreshave previously been linked to Aβ+ in CN individuals (Bennett et al., 2006; Petersenet al., 2016), although not in all studies (Rowe et al., 2010; Oh et al., 2012; Wirth et al.,2013). The FINGER population was mildly enriched for lower cognitive performance,and the population thus lacked one tail of the cognitive score distribution.

Cardiovascular factors were not included in the 85+ model at all, and in the FIN-GER population low BMI had modest predictive power while hypertension had noneat all. A low BMI at younger-old ages has previously been associated with an Aβ load(Ewers et al., 2012; Toledo et al., 2012), although these studies also included individ-uals with MCI and dementia at the baseline.

No sociodemographic factors—including age—were predictive. Although Aβ

pathology becomes more prevalent with increasing age, age was not a useful pre-dictor. This may be due to a saturation effect in the Vantaa 85+ cohort with Aβ

prevalence of 77% at death, but the relatively wide age spectrum of FINGER-PETparticipants (60–77 yr.) should have powered age as a predictor in that population,especially given the previously noted accelerated increase in the Aβ prevalence after70 years of age in ε4 carriers (Jack et al., 2015b).

Structural MRI measurements were the strongest predictors in the late-life popula-tion, which was to be expected. Decreased brain volumes and a high MTA rating areindicative of neurodegeneration, which is more likely to be present in the Aβ+ group.In AD, Aβ pathology is accompanied by neurodegenerative processes, partly asso-ciated to the tau pathology (Jack et al., 2013). These associations were also evidentin the PCA conducted in the Vantaa 85+ population. The first principal component“AD-type pathology” strongly linked amyloid pathology and tau pathology in theno-dementia group, and this PC also explained most of the variance in the pathol-ogy findings. In the dementia group this effect was weaker, and presence of AD-pathology indicated a lower macroinfarct load as expressed by the opposing signs ofthe scores. However, the first PC indicated that both groups showed most variance inrelation to AD-type of pathology. That is, AD-type pathology was the most importantdeterminant of the pathological profile even in non-demented individuals.

Few studies of CN participants have reported on the added value of MRI in amy-loid prediction, but prediction in cohorts with cognitively impaired individuals indi-

88

cated added value of this modality. Tosun et al. (2013) showed an added benefit ofadding structural MRI to prediction using the APOE genotype, which by itself was astrong predictor in the MCI study population (AUC from 0.81 to 0.88). Analogous fig-ures in the FINGER-PET study showed an improvement from 0.69 to 0.81–0.82. Sim-ilar results were found in other studies by adding MRI to the demographic data andAPOE (from 0.69 to 0.83; Tosun et al., 2014). On the other hand, change in cognitionwas found to be a superior substitute to structural MRI in a multimodal predictionmodel (Ansart et al., 2019).

Volumetric estimates of brain regions were the best predictors in a leave-one-outanalysis of MRI factor modalities in the FINGER-PET cohort. As part of the multi-modal model, visual Scheltens scores of MTA were almost as predictive as the setof volumetric measures. Such a visual rating is easier to obtain and would be moreuseful in a clinical setting and most research settings as well.

6.3.3 Associations of metabolic markers of diabetes and amyloid beta

DM has been reported to have a positive association with dementia of the AD and theVaD type, but the effect seemed to be stronger for VaD than AD (Cheng et al., 2012;Gudala et al., 2013). Studies have shown pathology underlying an elevated demen-tia incident rate in elderly DM patients to skew towards vascular pathology insteadof tau and Aβ pathology (Ahtiluoto et al., 2010). The DM–Aβ association was notsignificant in other previous studies (Moran et al., 2015; Roberts et al., 2014). Thus,the mechanisms of the DM–AD association are unclear, as is the role of Aβ. IR is ahallmark of DM2, and central nervous system IR has been suggested to be linked toAβ pathology through neuroinflammatory pathways or through competitive cleav-age of insulin and Aβ by the same enzyme (de la Monte, 2017). The epidemiologicalevidence discussed in section 2.6.4 showed this association to be rather weak, withno association in the elderly, and evidence being mixed in younger age groups. Thesuggestive finding in this thesis of lower IR in Aβ+ elderly without dementia or sub-stantial impairment adds to these prior studies. However, the findings in Study IVwere not significant after correction for multiple comparisons.

The prevalence of DM was 15%, and it did not differ significantly between theoutcome groups. This prevalence was in good agreement with the estimated Finnishprevalence of DM2 (65–74 years old cohort, previously diagnosed 10%, previouslyundiagnosed 12%; Peltonen et al., 2006).

To the author’s knowledge, no data has been published on the association of pe-ripheral blood PAI-1 levels and in-vivo Aβ markers. In prior studies, PAI-1 in CSF hasbeen reported to have no association with AD status (Martorana et al., 2012; Leunget al., 2015). Study IV suggested higher levels of PAI-1 to be protective against Aβ,although not significantly after correction for multiple comparisons. PAI-1 down-regulates the activity of the protein-cleaving plasmin system, and it is considered tobe a risk factor for atherosclerosis in the periphery due to its prothrombotic effect.In the central nervous system, however, PAI-1 and the plasmin system may interactwith Aβ fibrils and possibly affect plaque formation (Bi Oh et al., 2015), or be directly

89

neuroprotective (Cho et al., 2013).Both the HOMA-IR and PAI-1 findings suggested a higher potential load of vas-

cular pathology in Aβ negative individuals, at least in the periphery. It is unclearwhether this association with PAI-1 is mediated through effects in the central ner-vous system, or if the Aβ load is reduced due to effects in the cardiovascular system.The study population was also prefiltered—albeit fairly mildly—by cardiovascularrisk and cognition. The FINGER participants with low IR represent those who havemaintained adequate insulin sensitivity despite an elevated cardiovascular risk. Inthis specific subpopulation the mechanisms linking IR and Aβ accumulation may bedifferent.

The FINGER-PET population and the derivative IR/DM population was smallconsisting of only 41 subjects. This limited the use of confounders in the regressionanalysis and limited power.

The exploratory study did not find any other suggestive associations between thetested markers and Aβ positivity. To the authors knowledge, no data has been pub-lished on a complete assay of IR/DM markers previously. FINGER is an ongoinglongitudinal study and may in the future allow for the analysis of these markers overtime.

6.3.4 Prediction of other AD and amyloid related pathology

Tau tangles were predicted in the Vantaa 85+ population at a slightly lower perfor-mance than for Aβ. Predictive models for the neuropathological AD status—with anintermediate or high likelihood of AD based on a combination of Braak and CERADscores (Hyman et al., 2012)—had a better performance than models predicting eitherAβ or tau by themselves. The role of APOE was similar as for Aβ: ε4 allele waspredictive of pathology, and ε2 was protective. For tau, ε3 homozygousness wasprotective rather than ε2 carriership. This difference may be linked to a finding byBerlau et al. (2009) on elevated CERAD scores at autopsy in both ε2 and ε4 carriers,although ε2 did not raise the odds of developing dementia in that study like ε4 did.

The tangle count was predicted, in contrast to Aβ pathology, by the total choles-terol and LDL. Whether there is a vascular effect on tau pathology specifically isunclear, or whether other mechanisms underlying the previously observed risk in-crease associated with midlife vascular risk factors can explain the observation. Theneuropathological aggregate for AD did not have lipids or any other vascular fac-tors as predictors. However, subjective memory decline was a predictor, althougha rather weak one. This may indicate that the aggregate measure of AD pathologycorresponds better to the clinical phenotype than Aβ or tau separately.

Cerebral amyloid angiopathy was predicted by APOE alleles in the same way asAβ. Interestingly, cardiovascular comorbidities were predictive of lower CAA. Arte-rial amyloid deposition has been associated with intracerebral hemorrhages (ICHs),and the incidence of dementia after an ICH is high (Wermer and Greenberg, 2018;Banerjee et al., 2018). Persons with CAA without an ICH have been reported to scorelower on cognitive testing, possibly because CAA pathology of the arteries and corti-

90

cal arterioles also predisposes the patient to cortical microinfarctions (Banerjee et al.,2018). The effect seen in the Vantaa 85+ study may be partly due to a selection effect,where at the baseline individuals with both a high cardiovascular risk load (hyper-tension etc.) and CAA may have been excluded due to dementia or they may nothave survived to old age. The 85+ CAA population may thus represent persons withhigh cardiovascular resiliency to ICH due to CAA. Cerebrovascular comorbidities,however, did not show an effect. CAA was also predicted by male gender.

In the PC analysis of the no-dementia group, CAA contributed to the first PC—“AD-type pathology”—with an equal contribution to neuropathological AD. In thedementia group, CAA contributed with a somewhat smaller coefficient and vascularpathology contributed with an opposite sign. This may point to a differential roleof CAA in amyloid accumulation in healthy individuals with subclinical amounts ofpathology and in the brain of a dementia patient where vascular insults modify theclinical phenotype.

6.4 PREDICTION OF OTHER BRAIN PATHOLOGY

6.4.1 Vascular pathology

The prediction performance for vascular pathology was better than for other typesof pathology. The performance for WM macroinfarcts was the best, with blood lipidsbeing stand-out predictors. Lower HDL and lower LDL were predictive of the pres-ence of WM macroinfarcts. This result is counterintuitive, as lower HDL could beexpected to have a negative vascular effect, and likewise a lower LDL could be ex-pected to have a positive effect. Additionally, cerebrovascular comorbidities werepredictive of the same pathology. Perhaps cerebrovascular insults were of a moredetrimental magnitude in the high-LDL population and the results show a survivaleffect. It is also possible that these lipids and their potential changes over time priorto the baseline study visit have a completely different significance concerning the onbrain health of an 85+ population than for the vascular health of younger popula-tions. APOE ε2 carriership was predictive of WM macroinfarcts, a finding whichalso points to a survival effect. In younger populations both the ε2 and ε4 alleleshave been linked to more pronounced findings of cerebrovascular disease on MRI(Schilling et al., 2013). The findings indicating an increased risk of WM infarcts dueto ε2 may be a possible explanation as to why the ε2 allele was predictive of dementiain the Vantaa 85+ study.

Macroinfarcts in the cortex were predicted by the APOE genotype in the samepattern as they were for tau. The aggregate measure of all macroinfarcts was pre-dicted by factors better representative of the clinical state, namely poorer cognitionand competence in daily activities. It is notable, that objective measures of cognitionwere predictive of vascular pathology, but not AD-type pathology.

The presence of cerebral microinfarcts was only predicted by short duration ofeducation. Microinfarcts are a hallmark of CAA (Wermer and Greenberg, 2018), andthese pathologies could be expected to share predictors. No meaningful link was

91

seen on the PCA either.The PCA showed a differential pattern of vascular pathology in the no-dementia

and dementia groups. The second PC “Vascular pathology” was the second most im-portant determinant of variance in the no-dementia group, but did not show a clearpattern in the dementia group. It would be tempting to interpret the no-dementiacohort temporally as a pre-dementia cohort with more variability in the level of vas-cular pathology. However, the difference seen in PC2 may also be driven by otherdeterminants that differentiate the no-dementia and dementia groups. A survival ef-fect may also have an impact, as discussed above in relation to the APOE genotypes.

6.4.2 Hippocampal sclerosis, TDP-43 protein, and α-synuclein

The prevalence of HS and TDP-43 accumulation was 7% and 13%, respectively, inthe Vantaa 85+ autopsy population, as would be expected based on previously pub-lished estimates. HS is predominantly present in the oldest of the old, and priorresearch indicates a prevalence of 5–30% in 90–100-year-olds (Nelson et al., 2013).The prevalence of the strongly HS-associated TDP-43 in cognitively healthy elderlyis estimated at 24% (13–34% at a 95% confidence interval) worldwide (Nascimentoet al., 2018), and at 14% (9–20%) in Europe.

Low cognitive measures predicted HS well, and the overall performance wasbetter than for other pathologies. Hippocampus-associated memory tasks—such aswordlist task, which shows deficiencies in the HS+ autopsy population—have beenpreviously shown to be associated with HS, while cortex-dependent tasks such asverbal fluency remain relatively unaffected (Nelson et al., 2013). TDP-43 was pre-dicted only by depressive symptoms. Disease progression in TDP-43 accumulationis varied (Nascimento et al., 2018), and it is difficult to assess whether the effect ofdepressive symptoms seen here is generalizable.

No predictors were found for α-synuclein pathology, which is in line with priorresearch indicating no clear pattern of risk factors.

6.5 PREDICTION OF DEMENTIA VERSUS BRAIN PATHOLOGY

The prediction of incident dementia and different brain pathologies showed differ-ences in their overall performance and the relative importance of predictor modali-ties, which additionally varied in importance during the life course. Figure 1 visuallysummarizes the prediction results of this thesis. The horizontal axis represents thelate-life years of the life course, and in the context of this thesis this is further dividedinto younger and older old age groups. Predictors are represented by spheres, whosesizes correspond to the observed prediction performance. As evident in the upperpanel, cognition was the dominant predictor modality for dementia throughout oldage. Age was still relevant at the beginning, but not towards the end, when functionalmeasures gained relevance instead. Vascular predictors could be useful at the begin-ning of the late-life period, although not to the same degree as in midlife. The APOEgenotype was not a very significant determinant of incident dementia in comparison.

92

Key: Aβ amyloid beta protein, APOE apolipoprotein E, BMI body mass index, CAA cerebralamyloid angiopathy, CardioVD cardiovascular disease, CerebroVD cerebrovascular disease,HS hippocampal sclerosis, sMRI structural magnetic resonance imaging

Figure 1: Predictors of dementia and neuropathology in younger and older old age. Spheresize corresponds to predictor AUC.

93

For brain pathology, predictors showed a different pattern of importance. TheAPOE genotype was more predictive than cognition was of Aβ pathology through-out old age, and of neuropathological AD and CAA in the oldest of the old (also of taupathology, which is not shown in the figure). This underlines a disconnect betweenthe clinical phenotype and underlying AD and amyloid pathology. In general, ADpathology can be found in elderly persons without cognitive impairment (Bennettet al., 2006). Additionally, APOE ε4 has been shown to modulate the link betweenAβ load and cognition (Kantarci et al., 2012). Specifically, ε4 carriers experience moredecline in cognition with an Aβ load. An APOE ε4-stratified analysis would be in-teresting to perform in the future to test this in the pathology populations. Cognitionwas, however, a strong predictor of hippocampal sclerosis in the oldest of the oldand a weak predictor of cortical macroinfarcts, for which APOE played no role (thepattern was somewhat different for macroinfarcts in the cortex and in WM).

Whereas a low level of education was predictive of dementia, this finding did notseem to have a counterpart in pathology. The only effect was seen on microinfarcts(not shown in figure), but this had a low performance. According to a definition byStern et al. (2018), cognitive reserve is the ability of cognitive processes to adapt tochanges and insults of the brain. Findings in the Vantaa 85+ study agree with thenotion that the cognitive reserve due to higher education is a purely functional en-tity with no correlate in brain pathology. Cognitive reserve is thought to arise fromeither neural reserve—implying resistance of certain brain regions to insults—or neu-ral compensation, where unaffected brain regions compensate for the dysfunction ofaffected regions. There is functional MRI evidence for both mechanisms, with neuralcompensation becoming more important in more developed cognitive impairment(Anthony and Lin, 2018). Among other factors, physical activity has been consideredas a contributor to cognitive reserve (in addition to the cardiovascular benefits), andstudies have shown an increase in physical activity to be associated with structuralbrain changes (Rovio et al., 2010; Xu et al., 2015). None of the pathologies studiedin the Vantaa 85+ study reflect these associations. It should be noted that the meanlength of education was low at 4.3 years, and many other determinants of cognitivereserve may have had an effect along the life course.

The negligible role of age in dementia prediction in the oldest of the old was repli-cated in the lower pathology panel in Figure 1, where no pathology was predicted byage apart from the FINGER-PET model in which the AUC was 0.45. Gender was alsoomitted in almost all models for both dementia and pathology. However, CAA waspredicted by male gender. Male predominance has been shown at least in one study,in which 88% of the studied AD patients had CAA, and male participants had sig-nificantly higher CAA scores (Shinohara et al., 2016). CAA was strongly associatedwith AD and was also predicted by the APOE ε4 allele. Dementia prediction modelsincluded APOE as a predictor, but the effect of gender seen on CAA pathology didnot extend to the clinical dementia models.

Structural MRI was the dominant predictor of amyloid pathology in younger oldage. This is to be expected given the association of brain amyloid and neurodegener-

94

ation and atrophy, and the elevated a priori probability of pathology due to preselec-tion. Volumetric brain measures are typically also used as measures of brain reserve(Stern et al., 2018). The FINGER study inclusion criteria may have led to an overrep-resentation of low brain reserve i.e. lower brain volume on MRI possibly affecting thegeneralizability of these results. Whether amyloid positive individuals with higherbrain capacity have better cognitive performance should be further investigated inlarger cohorts.

Even without MRI data, in-vivo prediction of prevalent Aβ in the younger oldwas more effective than prediction in the older old group, where additional numberof years of follow up were allowed for the pathologies to develop. This, given thesimilar pattern in dementia prediction, calls for earlier intervention during the lifecourse, where the models indicate a higher potential for intervention in causal pro-cesses underlying the predictors, as well as better possibilities to find less advancedtargetable pathological changes.

95

7 CONCLUSIONS

The findings from the studies of this thesis support the following conclusions:

– The ten-year risk of late-life dementia was predicted well by multimodal pre-dictors. Cognition was the most important predictor, but genetic, vascular, andlife-style predictors added value.

– Changes in the parameters of vascular health may predict incident late-life de-mentia more accurately than cross-sectional measurements.

– Cognition was the dominant predictor of incident dementia in the oldest of theold. Other predictors recognized in younger populations lost value in this agegroup.

– In the oldest of the old, the prediction of vascular pathology succeeded moder-ately well, whereas the prediction of AD-type pathology was poor.

– Brain amyloid positivity in the cognitively healthy elderly could be predictedusing multimodal predictors including APOE genotype and structural MRI atan moderate-to-excellent level—which would be well-suited for amyloid preva-lence enrichment in populations.

– Two conclusions can be drawn on the role of APOE polymorphism:

– The ε4 allele predicted brain pathology better than the clinical outcome ofdementia in late life.

– The ε2 allele was not protective of all pathology in the oldest of the old; ε2carriership may predict white matter macroinfarcts over four years.

– High insulin resistance and high levels of PAI-1 in individuals with a lowerbrain amyloid burden may indicate brain resilience to higher a cardiovascularload.

These results demonstrate how the use of machine learning systems and multi-modal data can predict dementia and brain pathology years before incidence. Thiswill allow for a wider window of opportunity for prevention. The studies pointedto novel differences in dementia and pathology prediction, which impact the choiceof predictors for different applications of the models. Additionally, prediction inyounger old individuals was more effective and may also provide more time andopportunities for intervention. Prediction of the amyloid status proved accurate to alevel that may substantially aid in the design of intervention trials targeting amyloid,for example. The observed link between high insulin resistance and a high level ofPAI-1, and a low amyloid burden may prove useful in early detection of disease oroffer insight into preventive strategies.

97

This machine learning-based approach to multimodal prediction of both dementiaand brain pathology is relatively new in the dementia prevention field, and should befurther validated in other cohorts, including also other new and existing biomarkersthat were not tested as part of this thesis.

98

8 FUTURE PERSPECTIVES

Specific key recommendations have been made to advance future dementia risk mod-els (Ritchie and Muniz-Terrera, 2019). Risk models would benefit from biomarkerpredictors that would increase their performance at an early disease stage where tra-ditional predictors are insufficient. New markers of disease, such as tau-PET imaging,may prove potent predictors of disease, but are not necessarily practical on a largescale. Inclusion of a broad set of clinical predictor modalities and more precise test-ing of cognition may help indicate earlier changes associated with disease. The DSImachine learning algorithm with its multimodal technical design is a step towardsthis goal. In dementia prediction, cognition was a strong predictor, and a modelcombining even more accurate measurement of cognition and other clinical predic-tors could be useful worldwide in under-resourced clinics with limited possibilitiesfor biomarker testing. For example, such a model could be used for targeting morecostly and complex biomarker testing towards smaller, better defined risk groups.Additionally, to increase the usefulness of prediction models, modifiable risk factorsshould be included for the purposes of preventive and disease-modifying efforts.Such predictors were well represented in this thesis due to the design of the originalstudy cohorts. However, few were found useful in the final models, especially forthe oldest of the old. Recent guidelines on risk reduction of cognitive decline and de-mentia by the World Health Organization (2019) recommend intervention on severalmodifiable risk factors despite, in many cases, low or moderate levels of evidence, be-cause the benefits overweigh the risks both for cognition and health generally. Earlyintervention already in midlife has also been recommended, since several modifiablerisk factors in midlife show a clearer association with dementia risk compared toolder ages.

Personalized risk modeling and intervention design may be a part of the future.Factors such as cognitive reserve or genetic susceptibility modulate a person’s base-line risk, and other factors may affect their response to intervention. Such effects havebeen observed in carriers of different APOE polymorphisms, for example (Jensenet al., 2019). Tools such as clinical decision support systems with integration intohealth care systems and interfaces for laying out personalized risk and interventionprofiles (Mattila et al., 2012a) could prove applicable and useful even in primarycare. For dementia prevention, a personalized medicine approach may be needed.Systems that are especially economical in terms of health care system resources alsoshow promise, as the dementia-risk population is set to grow in light of current de-mographics and longevity gains. For example, a recent Internet-based interventionprogramme (Barbera et al., 2018) may offer a light-touch measure for dementia pre-vention requiring only small investment on an individual-by-individual basis. Theremay be interesting possibilities in combining such systems with health-register-basedprediction models that have shown good results in finding at-risk individuals (Wal-ters et al., 2016).

99

The steps taken in this thesis have contributed to the previously stated goals ofidentifying at-risk populations and including biomarkers as intermediate outcomesin prevention trials (National Academies of Sciences, Engineering, and Medicine andHealth, 2017). Pathology prediction models can identify affected individuals for ei-ther trial participation or possibly be used as surrogate outcomes themselves in somesettings. In future, these models need to be deployed in intervention trials. Newand improved prediction model generations will probably benefit from utilizing lon-gitudinal patient information from multiple sources, a task similar to those alreadyundertaken in many other fields in the big-data era.

100

REFERENCES

Abell, J. G., Kivimaki, M., Dugravot, A., Tabak, A. G., Fayosse, A., Shipley, M. et al.2018. Association between systolic blood pressure and dementia in the WhitehallII cohort study: role of age, duration, and threshold used to define hypertension.Eur Heart J, 39(33):3119–3125.

Ahtiluoto, S., Polvikoski, T., Peltonen, M., Solomon, A., Tuomilehto, J., Winblad, B.et al. 2010. Diabetes, Alzheimer disease, and vascular dementia: a population-based neuropathologic study. Neurology, 75(13):1195–202.

Aldrugh, S., Sardana, M., Henninger, N., Saczynski, J. S. and McManus, D. D. 2017.Atrial fibrillation, cognition and dementia: a review. J Cardiovasc Electrophysiol, 28(8):958–965.

Alzheimer’s Association. 2016. 2016 Alzheimer’s disease facts and figures.Alzheimer’s & Dementia, 12(4):459–509.

American Psychiatric Association. 1994. Diagnostic and statistical manual of mentaldisorders. Washington, DC, 4th edition.

American Psychiatric Association. 2000. Diagnostic and statistical manual of mentaldisorders. Washington, DC, 4th, text rev. edition.

American Psychiatric Association. 2013. Diagnostic and statistical manual of mentaldisorders. Washington, DC, 5th edition.

Andrieu, S., Guyonnet, S., Coley, N., Cantet, C., Bonnefoy, M., Bordes, S. et al. 2017.Effect of long-term omega 3 polyunsaturated fatty acid supplementation with orwithout multidomain intervention on cognitive function in elderly adults withmemory complaints (MAPT): a randomised, placebo-controlled trial. Lancet Neurol,16(5):377–389.

Ansart, M., Epelbaum, S., Gagliardi, G., Colliot, O., Dormont, D., Dubois, B. et al.2019. Reduction of recruitment costs in preclinical AD trials: validation of auto-matic pre-screening algorithm for brain amyloidosis. Stat Methods Med Res, Epubahead of print.

Anstey, K. J., Cherbuin, N., Budge, M. and Young, J. 2011. Body mass index in midlifeand late-life as a risk factor for dementia: a meta-analysis of prospective studies.Obes Rev, 12(5):e426–37.

Anstey, K. J., Lipnicki, D. M. and Low, L.-F. 2008. Cholesterol as a risk factor fordementia and cognitive decline: a systematic review of prospective studies withmeta-analysis. Am J Geriatr Psychiatry, 16(5):343–54.

Anstey, K. J., Mack, H. A. and Cherbuin, N. 2009. Alcohol consumption as a riskfactor for dementia and cognitive decline: meta-analysis of prospective studies.Am J Geriatr Psychiatry, 17(7):542–55.

Anstey, K. J., Cherbuin, N. and Herath, P. M. 2013. Development of a new method

101

for assessing global risk of Alzheimer’s disease for use in population health ap-proaches to prevention. Prevention science, 14(4):411–21.

Anstey, K. J., Cherbuin, N., Herath, P. M., Qiu, C., Kuller, L. H., Lopez, O. L. et al. 2014.A self-report risk index to predict occurrence of dementia in three independentcohorts of older adults: The ANU-ADRI. PLoS One, 9(1):e86141.

Anstey, K. J., Ashby-Mitchell, K. and Peters, R. 2017. Updating the evidence on theassociation between serum cholesterol and risk of late-life dementia: Review andmeta-analysis. J Alzheimers Dis, 56(1):215–228.

Anthony, M. and Lin, F. 2018. A systematic review for functional neuroimaging stud-ies of cognitive reserve across the cognitive aging spectrum. Arch Clin Neuropsychol,33(8):937–948.

Apostolova, L. G., Hwang, K. S., Avila, D., Elashoff, D., Kohannim, O., Teng, E. et al.2015. Brain amyloidosis ascertainment from cognitive, imaging, and peripheralblood protein measures. Neurology, 84(7):729–37.

Attems, J., Jellinger, K. A. and Lintner, F. 2005. Alzheimer’s disease pathology influ-ences severity and topographical distribution of cerebral amyloid angiopathy. ActaNeuropathol, 110(3):222–31.

Bahar-Fuchs, A., Villemagne, V., Ong, K., Chetelat, G., Lamb, F., Reininger, C. B. et al.2013. Prediction of amyloid-β pathology in amnestic mild cognitive impairmentwith neuropsychological tests. J Alzheimers Dis, 33(2):451–62.

Baker, L. D., Cross, D. J., Minoshima, S., Belongia, D., Watson, G. S. and Craft, S.2011. Insulin resistance and Alzheimer-like reductions in regional cerebral glucosemetabolism for cognitively normal adults with prediabetes or early type 2 diabetes.Arch Neurol, 68(1):51–7.

Banerjee, G., Wilson, D., Ambler, G., Osei-Bonsu Appiah, K., Shakeshaft, C., Lu-nawat, S. et al. 2018. Cognitive impairment before intracerebral hemorrhage isassociated with cerebral amyloid angiopathy. Stroke, 49(1):40–45.

Bang, J., Spina, S. and Miller, B. L. 2015. Frontotemporal dementia. Lancet, 386(10004):1672–82.

Barbera, M., Mangialasche, F., Jongstra, S., Guillemont, J., Ngandu, T., Beishuizen,C. et al. 2018. Designing an Internet-based multidomain intervention for the pre-vention of cardiovascular disease and cognitive impairment in older adults: TheHATICE trial. J Alzheimers Dis, 62(2):649–663.

Barnes, D. E., Beiser, A. S., Lee, A., Langa, K. M., Koyama, A., Preis, S. R. et al. 2014.Development and validation of a brief dementia screening indicator for primarycare. Alzheimers Dement, 10(6):656–665.e1.

Baumgart, M., Snyder, H. M., Carrillo, M. C., Fazio, S., Kim, H. and Johns, H. 2015.Summary of the evidence on modifiable risk factors for cognitive decline and de-mentia: A population-based perspective. Alzheimers Dement, 11(6):718–26.

Beach, T. G., Monsell, S. E., Phillips, L. E. and Kukull, W. 2012. Accuracy of theclinical diagnosis of Alzheimer disease at National Institute on Aging Alzheimer

102

Disease Centers, 2005-2010. J Neuropathol Exp Neurol, 71(4):266–73.

Beck, A. T., Ward, C. H., Mendelson, M., Mock, J. and Erbaugh, J. 1961. An inventoryfor measuring depression. Arch Gen Psychiatry, 4:561–71.

Benjamini, Y. and Hochberg, Y. 1995. Controlling the false discovery rate: A practicaland powerful approach to multiple testing. J R Stat Soc Series B Stat Methodol, 57(1):289–300.

Bennett, D. A., Schneider, J. A., Arvanitakis, Z., Kelly, J. F., Aggarwal, N. T., Shah,R. C. et al. 2006. Neuropathology of older persons without cognitive impairmentfrom two community-based studies. Neurology, 66(12):1837–44.

Berg, L. 1988. Clinical dementia rating (CDR). Psychopharmacol Bull, 24(4):637–9.

Berlau, D. J., Corrada, M. M., Head, E. and Kawas, C. H. 2009. APOE ε2 is asso-ciated with intact cognition but increased Alzheimer pathology in the oldest old.Neurology, 72(9):829–34.

Berlau, D. J., Corrada, M. M., Robinson, J. L., Geser, F., Arnold, S. E., Lee, V. M.-Y.et al. 2013. Neocortical β-amyloid area is associated with dementia and APOE inthe oldest-old. Alzheimers Dement, 9(6):699–705.

Beydoun, M. A., Beydoun, H. A. and Wang, Y. 2008. Obesity and central obesity asrisk factors for incident dementia and its subtypes: a systematic review and meta-analysis. Obes Rev, 9(3):204–18.

Bi Oh, S., Suh, N., Kim, I. and Lee, J.-Y. 2015. Impacts of aging and amyloid-βdeposition on plasminogen activators and plasminogen activator inhibitor-1 in theTg2576 mouse model of Alzheimer’s disease. Brain Res, 1597:159–67.

Bjorkhem, I. 2006. Crossing the barrier: oxysterols as cholesterol transporters andmetabolic modulators in the brain. J Intern Med, 260(6):493–508.

Bondi, M. W., Edmonds, E. C. and Salmon, D. P. 2017. Alzheimer’s disease: Past,present, and future. J Int Neuropsychol Soc, 23(9-10):818–831.

Borkowski, J., Benton, A. and Spreen, O. 1967. Word fluency and brain damage.Neuropsychologia, 5:135–140.

Boyle, P. A., Yu, L., Leurgans, S. E., Wilson, R. S., Brookmeyer, R., Schneider, J. A.et al. 2019. Attributable risk of alzheimer’s dementia attributed to age-relatedneuropathologies. Ann Neurol, 85(1):114–124.

Braak, H. and Braak, E. 1991. Neuropathological stageing of Alzheimer-relatedchanges. Acta Neuropathol, 82(4):239–59.

Brenowitz, W. D., Nelson, P. T., Besser, L. M., Heller, K. B. and Kukull, W. A. 2015.Cerebral amyloid angiopathy and its co-occurrence with Alzheimer’s disease andother cerebrovascular neuropathologic changes. Neurobiol Aging, 36(10):2702–8.

Burnham, S. C., Faux, N. G., Wilson, W., Laws, S. M., Ames, D., Bedo, J. et al. 2014.A blood-based predictor for neocortical aβ burden in Alzheimer’s disease: resultsfrom the AIBL study. Mol Psychiatry, 19(4):519–26.

Byers, A. L. and Yaffe, K. 2011. Depression and risk of developing dementia. Nat Rev

103

Neurol, 7(6):323–31.

Caamano-Isorna, F., Corral, M., Montes-Martınez, A. and Takkouche, B. 2006. Edu-cation and dementia: a meta-analytic study. Neuroepidemiology, 26(4):226–32.

Cannon, J. A., Moffitt, P., Perez-Moreno, A. C., Walters, M. R., Broomfield, N. M.,McMurray, J. J. V. et al. 2017. Cognitive impairment and heart failure: Systematicreview and meta-analysis. J Card Fail, 23(6):464–475.

Cheng, G., Huang, C., Deng, H. and Wang, H. 2012. Diabetes as a risk factor fordementia and mild cognitive impairment: a meta-analysis of longitudinal studies.Intern Med J, 42(5):484–91.

Cherbuin, N., Kim, S. and Anstey, K. J. 2015. Dementia risk estimates associated withmeasures of depression: a systematic review and meta-analysis. BMJ Open, 5(12):e008853.

Chhetri, J. K., de Souto Barreto, P., Cantet, C., Pothier, K., Cesari, M., Andrieu, S.et al. 2018. Effects of a 3-year multi-domain intervention with or without omega-3 supplementation on cognitive functions in older subjects with increased CAIDEdementia scores. J Alzheimers Dis, 64(1):71–78.

Cho, H., Joo, Y., Kim, S., Woo, R.-S., Lee, S. H. and Kim, H.-S. 2013. Plasminogen ac-tivator inhibitor-1 promotes synaptogenesis and protects against Aβ1−42-inducedneurotoxicity in primary cultured hippocampal neurons. Int J Neurosci, 123(1):42–9.

Chui, H. C., Victoroff, J. I., Margolin, D., Jagust, W., Shankle, R. and Katzman, R. 1992.Criteria for the diagnosis of ischemic vascular dementia proposed by the State ofCalifornia Alzheimer’s Disease Diagnostic and Treatment Centers. Neurology, 42(3Pt 1):473–80.

Clark, C. M., Pontecorvo, M. J., Beach, T. G., Bedell, B. J., Coleman, R. E., Doraiswamy,P. M. et al. 2012. Cerebral PET with florbetapir compared with neuropathology atautopsy for detection of neuritic amyloid-β plaques: a prospective cohort study.Lancet Neurol, 11(8):669–78.

ClinicalTrials.gov. Identifier NCT01931566, Biomarker qualification for risk ofmild cognitive impairment (MCI) due to Alzheimer’s disease (AD) and safetyand efficacy evaluation of pioglitazone in delaying its onset, 2018. URLclinicaltrials.gov/ct2/show/NCT01931566.

Collins, G. S., Reitsma, J. B., Altman, D. G. and Moons, K. G. M. 2015. Transparentreporting of a multivariable prediction model for individual prognosis or diagnosis(TRIPOD): the TRIPOD statement. BMC Med, 13:1.

Corder, E. H., Saunders, A. M., Strittmatter, W. J., Schmechel, D. E., Gaskell, P. C.,Small, G. W. et al. 1993. Gene dose of apolipoprotein E type 4 allele and the risk ofAlzheimer’s disease in late onset families. Science, 261(5123):921–3.

Corder, E. H., Saunders, A. M., Risch, N. J., Strittmatter, W. J., Schmechel, D. E.,Gaskell, P. C., Jr et al. 1994. Protective effect of apolipoprotein E type 2 allelefor late onset Alzheimer disease. Nature genetics, 7(2):180–4.

Corrada, M. M., Paganini-Hill, A., Berlau, D. J. and Kawas, C. H. 2013. Apolipopro-

104

clinicaltrials.gov/ct2/show/NCT01931566

tein E genotype, dementia, and mortality in the oldest old: the 90+ study.Alzheimers Dement, 9(1):12–8.

Corrada, M. M., Hayden, K. M., Paganini-Hill, A., Bullain, S. S., DeMoss, J., Aguirre,C. et al. 2017. Age of onset of hypertension and risk of dementia in the oldest-old:The 90+ study. Alzheimers Dement, 13(2):103–110.

Cykowski, M. D., Powell, S. Z., Schulz, P. E., Takei, H., Rivera, A. L., Jackson, R. E.et al. 2017. Hippocampal sclerosis in older patients: Practical examples and guid-ance with a focus on cerebral age-related TDP-43 with sclerosis. Arch Pathol LabMed, 141(8):1113–1126.

de la Monte, S. M. 2017. Insulin resistance and neurodegeneration: Progress towardsthe development of new therapeutics for Alzheimer’s disease. Drugs, 77(1):47–65.

De la Vega, F. M., Lazaruk, K. D., Rhodes, M. D. and Wenz, M. H. 2005. Assessment oftwo flexible and compatible SNP genotyping platforms: TaqMan SNP genotypingassays and the SNPlex genotyping system. Mutat Res, 573(1-2):111–35.

Delrieu, J., Payoux, P., Carrie, I., Cantet, C., Weiner, M., Vellas, B. et al. 2019. Multido-main intervention and/or omega-3 in nondemented elderly subjects according toamyloid status. Alzheimers Dement, 15(11):1392–1401.

Derby, C. A., Burns, L. C., Wang, C., Katz, M. J., Zimmerman, M. E., L’italien, G. et al.2013. Screening for predementia AD: time-dependent operating characteristics ofepisodic memory tests. Neurology, 80(14):1307–14.

Diehl, T., Mullins, R. and Kapogiannis, D. 2017. Insulin resistance in Alzheimer’sdisease. Transl Res, 183:26–40.

Dubois, B. 2000. ‘Prodromal Alzheimer’s disease’: a more useful concept than mildcognitive impairment? Curr Opin Neurol, 13(4):367–9.

Dubois, B. and Albert, M. L. 2004. Amnestic MCI or prodromal Alzheimer’s disease?Lancet Neurol, 3(4):246–248.

Dubois, B., Feldman, H. H., Jacova, C., Dekosky, S. T., Barberger-Gateau, P., Cum-mings, J. et al. 2007. Research criteria for the diagnosis of Alzheimer’s disease:revising the NINCDS-ADRDA criteria. Lancet Neurol, 6(8):734–46.

Dubois, B., Feldman, H. H., Jacova, C., Cummings, J. L., Dekosky, S. T., Barberger-Gateau, P. et al. 2010. Revising the definition of Alzheimer’s disease: a new lexicon.Lancet Neurol, 9(11):1118–27.

Dubois, B., Feldman, H. H., Jacova, C., Hampel, H., Molinuevo, J. L., Blennow, K.et al. 2014. Advancing research diagnostic criteria for Alzheimer’s disease: theIWG-2 criteria. Lancet Neurol, 13(6):614–29.

Dubois, B., Hampel, H., Feldman, H. H., Scheltens, P., Aisen, P., Andrieu, S. et al.2016. Preclinical Alzheimer’s disease: Definition, natural history, and diagnosticcriteria. Alzheimers Dement, 12(3):292–323.

Dubois, B., Epelbaum, S., Nyasse, F., Bakardjian, H., Gagliardi, G., Uspenskaya, O.et al. 2018. Cognitive and neuroimaging features and brain β-amyloidosis in indi-

105

viduals at risk of Alzheimer’s disease (INSIGHT-preAD): a longitudinal observa-tional study. Lancet Neurol, 17(4):335–346.

Dunteman, G. 1989. Principal Components Analysis. Quantitative Applications in theSocial Sciences. Sage Publications, Newbury Park, California.

Eichner, J. E., Dunn, S. T., Perveen, G., Thompson, D. M., Stewart, K. E. and Stroehla,B. C. 2002. Apolipoprotein E polymorphism and cardiovascular disease: a HuGEreview. Am J Epidemiol, 155(6):487–95.

Einstein, G. O., Smith, R. E., McDaniel, M. A. and Shaw, P. 1997. Aging and prospec-tive memory: The influence of increased task demands at encoding and retrieval.Psychology and Aging, 12:479–488.

Ekblad, L. L., Johansson, J., Helin, S., Viitanen, M., Laine, H., Puukka, P. et al. 2018.Midlife insulin resistance, APOE genotype, and late-life brain amyloid accumula-tion. Neurology, 90(13):e1150–e1157.

Emre, M., Aarsland, D., Brown, R., Burn, D. J., Duyckaerts, C., Mizuno, Y. et al. 2007.Clinical diagnostic criteria for dementia associated with Parkinson’s disease. MovDisord, 22(12):1689–707.

Engelhardt, E. and Grinberg, L. T. 2015. Alois Alzheimer and vascular brain disease:Arteriosclerotic atrophy of the brain. Dement Neuropsychol, 9(1):81–84.

Erez, A., Kivity, S., Berkovitch, A., Milwidsky, A., Klempfner, R., Segev, S. et al. 2015.The association between cardiorespiratory fitness and cardiovascular risk may bemodulated by known cardiovascular risk factors. Am Heart J, 169(6):916–923.e1.

Erkinjuntti, T. and Gauthier, S. 2009. The concept of vascular cognitive impairment.Front Neurol Neurosci, 24:79–85.

Erkinjuntti, T., Remes, A. and Rinne, J. 2015. Muistisairaudet. Kustannus OyDuodecim, 2nd edition.

Ewers, M., Schmitz, S., Hansson, O., Walsh, C., Fitzpatrick, A., Bennett, D. et al. 2012.Body mass index is associated with biological CSF markers of core brain pathologyof Alzheimer’s disease. Neurobiol Aging, 33(8):1599–608.

Exalto, L. G., Biessels, G. J., Karter, A. J., Huang, E. S., Katon, W. J., Minkoff, J. R. et al.2013a. Risk score for prediction of 10 year dementia risk in individuals with type 2diabetes: a cohort study. Lancet Diabetes Endocrinol, 1(3):183–90.

Exalto, L. G., Quesenberry, C. P., Barnes, D., Kivipelto, M., Biessels, G. J. and Whitmer,R. A. 2013b. Midlife risk score for the prediction of dementia four decades later.Alzheimer’s & dementia, 10(5):562–70.

Fazekas, F., Chawluk, J. B., Alavi, A., Hurtig, H. I. and Zimmerman, R. A. 1987. MRsignal abnormalities at 1.5 T in Alzheimer’s dementia and normal aging. AJR Am JRoentgenol, 149(2):351–6.

Folstein, M. F., Folstein, S. E. and McHugh, P. R. 1975. ”Mini-mental state”. a practicalmethod for grading the cognitive state of patients for the clinician. J Psychiatr Res,12(3):189–98.

106

Ford, A. H. and Almeida, O. P. 2019. Effect of vitamin B supplementation on cognitivefunction in the elderly: A systematic review and meta-analysis. Drugs Aging, 36(5):419–434.

Freudenberger, P., Petrovic, K., Sen, A., Toglhofer, A. M., Fixa, A., Hofer, E. et al.2016. Fitness and cognition in the elderly: The Austrian Stroke Prevention Study.Neurology, 86(5):418–24.

Gardner, R. C., Valcour, V. and Yaffe, K. 2013. Dementia in the oldest old: a multi-factorial and growing public health issue. Alzheimers Res Ther, 5(4):27.

Geifman, N., Brinton, R. D., Kennedy, R. E., Schneider, L. S. and Butte, A. J. 2017.Evidence for benefit of statins to modify cognitive decline and risk in Alzheimer’sdisease. Alzheimers Res Ther, 9(1):10.

Goerdten, J., Cukic, I., Danso, S. O., Carriere, I. and Muniz-Terrera, G. 2019. Statisticalmethods for dementia risk prediction and recommendations for future work: Asystematic review. Alzheimers Dement (N Y), 5:563–569.

Gorelick, P. B., Scuteri, A., Black, S. E., Decarli, C., Greenberg, S. M., Iadecola, C.et al. 2011. Vascular contributions to cognitive impairment and dementia: a state-ment for healthcare professionals from the American Heart Association/AmericanStroke Association. Stroke, 42(9):2672–713.

Graupe, D. 2007. Principles of Artificial Neural Networks. Advanced Series in Circuitsand Systems Ser. World Scientific Publishing, 2nd edition.

Grober, E., Sanders, A. E., Hall, C. and Lipton, R. B. 2010. Free and cued selective re-minding identifies very mild dementia in primary care. Alzheimer Dis Assoc Disord,24(3):284–90.

Gudala, K., Bansal, D., Schifano, F. and Bhansali, A. 2013. Diabetes mellitus andrisk of dementia: A meta-analysis of prospective observational studies. J DiabetesInvestig, 4(6):640–50.

Haghighi, M., Smith, A., Morgan, D., Small, B. and Huang, S. 2015. Identifyingcost-effective predictive rules of amyloid-β level by integrating neuropsychologicaltests and plasma-based markers. J Alzheimers Dis, 43(4):1261–70.

Hakansson, K., Soininen, H., Winblad, B. and Kivipelto, M. 2015. Feelings of hope-lessness in midlife and cognitive health in later life: A prospective population-based cohort study. PLoS One, 10(10):e0140261.

Hall, A., Munoz-Ruiz, M., Mattila, J., Koikkalainen, J., Tsolaki, M., Mecocci, P. et al.2015. Generalizability of the Disease State Index prediction model for identify-ing patients progressing from mild cognitive impairment to Alzheimer’s disease.J Alzheimers Dis, 44(1):79–92.

Hamer, M. and Chida, Y. 2009. Physical activity and risk of neurodegenerative dis-ease: a systematic review of prospective evidence. Psychol Med, 39(1):3–11.

Hanley, J. A. and McNeil, B. J. 1982. The meaning and use of the area under a receiveroperating characteristic (ROC) curve. Radiology, 143(1):29–36.

107

Hanninen, T., Pulliainen, V., Sotaniemi, M., Hokkanen, L., Salo, J., Hietanen, M. et al.2010. Early detection of cognitive changes in memory diseases: new cut-off scoresfor the Finnish version of CERAD neuropsychological battery. Duodecim, 126(17):2013–21.

Harrell, F. E., Jr, Lee, K. L. and Mark, D. B. 1996. Multivariable prognostic models: is-sues in developing models, evaluating assumptions and adequacy, and measuringand reducing errors. Stat Med, 15(4):361–87.

Harrison, J., Minassian, S. L., Jenkins, L., Black, R. S., Koller, M. and Grundman, M.2007. A neuropsychological test battery for use in Alzheimer disease clinical trials.Arch Neurol, 64(9):1323–9.

Hayden, K. M., Zandi, P. P., Lyketsos, C. G., Khachaturian, A. S., Bastian, L. A., Cha-roonruk, G. et al. 2006. Vascular risk factors for incident Alzheimer disease andvascular dementia: the Cache County study. Alzheimer Dis Assoc Disord, 20(2):93–100.

Heun, R., Burkart, M., Wolf, C. and Benkert, O. 1998. Effect of presentation rate onword list learning in patients with dementia of the Alzheimer type. Dement GeriatrCogn Disord, 9(4):214–8.

Hixson, J. E. and Vernier, D. T. 1990. Restriction isotyping of human apolipoproteinE by gene amplification and cleavage with HhaI. J Lipid Res, 31(3):545–8.

Hosmer, D. W., Lemeshow, S. and Sturdivant, R. X. Applied Logistic Regression, chapterAssessing the Fit of the Model. John Wiley & Sons, 3rd edition, 2013.

Hou, X.-H., Feng, L., Zhang, C., Cao, X.-P., Tan, L. and Yu, J.-T. 2019. Models forpredicting risk of dementia: a systematic review. J Neurol Neurosurg Psychiatry, 90(4):373–379.

Hyman, B. T., Phelps, C. H., Beach, T. G., Bigio, E. H., Cairns, N. J., Carrillo, M. C.et al. 2012. National Institute on Aging-Alzheimer’s Association guidelines for theneuropathologic assessment of Alzheimer’s disease. Alzheimers Dement, 8(1):1–13.

Insel, P. S., Palmqvist, S., Mackin, R. S., Nosheny, R. L., Hansson, O., Weiner, M. W.et al. 2016. Assessing risk for preclinical β-amyloid pathology with APOE, cogni-tive, and demographic information. Alzheimers Dement (Amst), 4:76–84.

Jack, C. R., Jr and Vemuri, P. 2018. Amyloid-β—a reflection of risk or a preclinicalmarker? Nat Rev Neurol, 14(6):319–320.

Jack, C. R., Jr, Knopman, D. S., Jagust, W. J., Petersen, R. C., Weiner, M. W., Aisen,P. S. et al. 2013. Tracking pathophysiological processes in Alzheimer’s disease: anupdated hypothetical model of dynamic biomarkers. Lancet Neurol, 12(2):207–16.

Jack, C. R., Jr, Wiste, H. J., Weigand, S. D., Knopman, D. S., Mielke, M. M., Vemuri,P. et al. 2015a. Different definitions of neurodegeneration produce similar amy-loid/neurodegeneration biomarker group findings. Brain, 138(Pt 12):3747–59.

Jack, C. R., Jr, Wiste, H. J., Weigand, S. D., Knopman, D. S., Vemuri, P., Mielke, M. M.et al. 2015b. Age, sex, and APOE ε4 effects on memory, brain structure, and β-amyloid across the adult life span. JAMA Neurol, 72(5):511–9.

108

Jack, C. R., Jr, Bennett, D. A., Blennow, K., Carrillo, M. C., Feldman, H. H., Frisoni,G. B. et al. 2016. A/T/N: An unbiased descriptive classification scheme forAlzheimer disease biomarkers. Neurology, 87(5):539–47.

Jack, C. R., Jr, Bennett, D. A., Blennow, K., Carrillo, M. C., Dunn, B., Haeberlein,S. B. et al. 2018. NIA-AA research framework: Toward a biological definition ofAlzheimer’s disease. Alzheimers Dement, 14(4):535–562.

Jansen, W. J., Ossenkoppele, R., Knol, D. L., Tijms, B. M., Scheltens, P., Verhey, F. R. J.et al. 2015. Prevalence of cerebral amyloid pathology in persons without dementia:a meta-analysis. JAMA, 313(19):1924–38.

Jellinger, K. A. and Attems, J. 2010. Prevalence of dementia disorders in the oldest-old: an autopsy study. Acta Neuropathol, 119(4):421–33.

Jensen, C. S., Simonsen, A. H., Siersma, V., Beyer, N., Frederiksen, K. S., Gottrup, H.et al. 2019. Patients with Alzheimer’s disease who carry the APOE ε4 allele benefitmore from physical exercise. Alzheimers Dement (N Y), 5:99–106.

Jeon, S. Y., Byun, M. S., Yi, D., Lee, J. H., Choe, Y. M., Ko, K. et al. 2019. Influenceof hypertension on brain amyloid deposition and Alzheimer’s disease signatureneurodegeneration. Neurobiol Aging, 75:62–70.

Jonsson, T., Atwal, J. K., Steinberg, S., Snaedal, J., Jonsson, P. V., Bjornsson, S. et al.2012. A mutation in APP protects against Alzheimer’s disease and age-relatedcognitive decline. Nature, 488(7409):96–9.

Jorm, A. F. and Jolley, D. 1998. The incidence of dementia: a meta-analysis. Neurology,51(3):728–33.

Juva, K., Verkkoniemi, A., Viramo, P., Polvikoski, T., Kainulainen, K., Kontula, K.et al. 2000. APOE epsilon4 does not predict mortality, cognitive decline, or demen-tia in the oldest old. Neurology, 54(2):412–5.

Kantarci, K., Lowe, V., Przybelski, S. A., Weigand, S. D., Senjem, M. L., Ivnik, R. J.et al. 2012. APOE modifies the association between Aβ load and cognition incognitively normal older adults. Neurology, 78(4):232–40.

Kapogiannis, D., Boxer, A., Schwartz, J. B., Abner, E. L., Biragyn, A., Masharani, U.et al. 2015. Dysfunctionally phosphorylated type 1 insulin receptor substrate inneural-derived blood exosomes of preclinical Alzheimer’s disease. FASEB J, 29(2):589–96.

Katz, S., Ford, A. B., Moskowitz, R. W., Jackson, B. A. and Jaffe, M. W. 1963. Studiesof illness in the aged. The index of ADL: A standardized measure of biological andpsychosocial function. JAMA, 185:914–9.

Kemppainen, N., Johansson, J., Teuho, J., Parkkola, R., Joutsa, J., Ngandu, T. et al.2017. Brain amyloid load and its associations with cognition and vascular riskfactors in FINGER study. Neurology, 90(3):e206–e213.

Kennelly, S. P., Lawlor, B. A. and Kenny, R. A. 2009. Blood pressure and the risk fordementia: a double edged sword. Ageing Res Rev, 8(2):61–70.

109

Kero, M., Raunio, A., Polvikoski, T., Tienari, P. J., Paetau, A. and Myllykangas, L.2018. Hippocampal sclerosis in the oldest old: A Finnish population-based study.J Alzheimers Dis, 63(1):263–272.

Kivimaki, M., Luukkonen, R., Batty, G. D., Ferrie, J. E., Pentti, J., Nyberg, S. T. et al.2018. Body mass index and risk of dementia: Analysis of individual-level datafrom 1.3 million individuals. Alzheimers Dement, 14(5):601–609.

Kivipelto, M. and Solomon, A. 2006. Cholesterol as a risk factor for alzheimer’sdisease—epidemiological evidence. Acta Neurol Scand Suppl, 185:50–7.

Kivipelto, M., Helkala, E. L., Hanninen, T., Laakso, M. P., Hallikainen, M., Alhainen,K. et al. 2001a. Midlife vascular risk factors and late-life mild cognitive impairment:A population-based study. Neurology, 56(12):1683–9.

Kivipelto, M., Helkala, E. L., Laakso, M. P., Hanninen, T., Hallikainen, M., Alhainen,K. et al. 2001b. Midlife vascular risk factors and Alzheimer’s disease in later life:longitudinal, population based study. BMJ, 322(7300):1447–51.

Kivipelto, M., Ngandu, T., Laatikainen, T., Winblad, B., Soininen, H. and Tuomilehto,J. 2006. Risk score for the prediction of dementia risk in 20 years among middleaged people: a longitudinal, population-based study. Lancet Neurol, 5(9):735–41.

Kivipelto, M., Solomon, A., Ahtiluoto, S., Ngandu, T., Lehtisalo, J., Antikainen, R.et al. 2013. The Finnish geriatric intervention study to prevent cognitive impair-ment and disability (FINGER): study design and progress. Alzheimers Dement, 9(6):657–65.

Kivipelto, M., Mangialasche, F. and Ngandu, T. 2018. Lifestyle interventions to pre-vent cognitive impairment, dementia and Alzheimer disease. Nat Rev Neurol, 14(11):653–666.

Koikkalainen, J., Rhodius-Meester, H., Tolonen, A., Barkhof, F., Tijms, B., Lemstra,A. W. et al. 2016. Differential diagnosis of neurodegenerative diseases using struc-tural MRI data. Neuroimage Clin, 11:435–449.

Korf, E. S. C., White, L. R., Scheltens, P. and Launer, L. J. 2006. Brain aging in veryold men with type 2 diabetes: the Honolulu-Asia aging study. Diabetes Care, 29(10):2268–74.

Kuiper, J. S., Zuidersma, M., Oude Voshaar, R. C., Zuidema, S. U., van den Heuvel,E. R., Stolk, R. P. et al. 2015. Social relationships and risk of dementia: A systematicreview and meta-analysis of longitudinal cohort studies. Ageing Res Rev, 22:39–57.

Launer, L. J., Ross, G. W., Petrovitch, H., Masaki, K., Foley, D., White, L. R. et al. 2000.Midlife blood pressure and dementia: the honolulu-asia aging study. NeurobiolAging, 21(1):49–55.

Laws, S. M., Gaskin, S., Woodfield, A., Srikanth, V., Bruce, D., Fraser, P. E. et al. 2017.Insulin resistance is associated with reductions in specific cognitive domains andincreases in CSF tau in cognitively normal adults. Sci Rep, 7(1):9766.

Lawton, M. P. and Brody, E. M. 1969. Assessment of older people: self-maintainingand instrumental activities of daily living. Gerontologist, 9(3):179–86.

110

Lee, J. H., Byun, M. S., Yi, D., Sohn, B. K., Jeon, S. Y., Lee, Y. et al. 2018. Prediction ofcerebral amyloid with common information obtained from memory clinic practice.Front Aging Neurosci, 10:309.

Leung, Y. Y., Toledo, J. B., Nefedov, A., Polikar, R., Raghavan, N., Xie, S. X. et al.2015. Identifying amyloid pathology-related cerebrospinal fluid biomarkers forAlzheimer’s disease in a multicohort study. Alzheimers Dement (Amst), 1(3):339–348.

Li, C.-I., Li, T.-C., Liu, C.-S., Liao, L.-N., Lin, W.-Y., Lin, C.-H. et al. 2018. Risk scoreprediction model for dementia in patients with type 2 diabetes. Eur J Neurol, 25(7):976–983.

Livingston, G., Sommerlad, A., Orgeta, V., Costafreda, S. G., Huntley, J., Ames, D.et al. 2017. Dementia prevention, intervention, and care. Lancet, 390(10113):2673–2734.

Lo, R. Y., Jagust, W. J. and Alzheimer’s Disease Neuroimaging Initiative. 2013. Effectof cognitive reserve markers on Alzheimer pathologic progression. Alzheimer DisAssoc Disord, 27(4):343–50.

Louhija, J., Miettinen, H. E., Kontula, K., Tikkanen, M. J., Miettinen, T. A. and Tilvis,R. S. 1994. Aging and genetic variation of plasma apolipoproteins. Relative loss ofthe apolipoprotein E4 phenotype in centenarians. Arterioscler Thromb, 14(7):1084–9.

Loy, C. T., Schofield, P. R., Turner, A. M. and Kwok, J. B. J. 2014. Genetics of dementia.Lancet, 383(9919):828–40.

Lutz, M. W., Sundseth, S. S., Burns, D. K., Saunders, A. M., Hayden, K. M., Burke,J. R. et al. 2016. A genetics-based biomarker risk algorithm for predicting risk ofAlzheimer’s disease. Alzheimers Dement (N Y), 2(1):30–44.

Mann, D. M. 1988. Alzheimer’s disease and Down’s syndrome. Histopathology, 13(2):125–37.

Martorana, A., Sancesario, G. M., Esposito, Z., Nuccetelli, M., Sorge, R., Formosa, A.et al. 2012. Plasmin system of Alzheimer’s disease patients: CSF analysis. J NeuralTransm (Vienna), 119(7):763–9.

Mattila, J., Koikkalainen, J., Virkki, A., Simonsen, A., van Gils, M., Waldemar, G.et al. 2011. A Disease State Fingerprint for evaluation of Alzheimer’s disease. JAlzheimers Dis, 27(1):163–76.

Mattila, J., Koikkalainen, J., Virkki, A., van Gils, M. and Lotjonen, J. 2012a. Designand application of a generic clinical decision support system for multiscale data.IEEE transactions on bio-medical engineering, 59(1):234–40.

Mattila, J., Soininen, H., Koikkalainen, J., Rueckert, D., Wolz, R., Waldemar, G. et al.2012b. Optimizing the diagnosis of early Alzheimer’s disease in mild cognitiveimpairment subjects. J Alzheimers Dis, 32(4):969–79.

McKeith, I. G., Dickson, D. W., Lowe, J., Emre, M., O’Brien, J. T., Feldman, H. et al.2005. Diagnosis and management of dementia with Lewy bodies: third report ofthe DLB consortium. Neurology, 65(12):1863–72.

111

McKhann, G., Drachman, D., Folstein, M., Katzman, R., Price, D. and Stadlan, E. M.1984. Clinical diagnosis of Alzheimer’s disease: report of the NINCDS-ADRDAwork group under the auspices of Department of Health and Human Services TaskForce on Alzheimer’s Disease. Neurology, 34(7):939–44.

McKhann, G. M., Knopman, D. S., Chertkow, H., Hyman, B. T., Jack, C. R., Jr, Kawas,C. H. et al. 2011. The diagnosis of dementia due to Alzheimer’s disease: recommen-dations from the National Institute on Aging-Alzheimer’s Association workgroupson diagnostic guidelines for Alzheimer’s disease. Alzheimers Dement, 7(3):263–9.

Meng, X. and D’Arcy, C. 2012. Education and dementia in the context of the cog-nitive reserve hypothesis: a systematic review with meta-analyses and qualitativeanalyses. PLoS One, 7(6):e38268.

Mielke, M. M., Wiste, H. J., Weigand, S. D., Knopman, D. S., Lowe, V. J., Roberts,R. O. et al. 2012. Indicators of amyloid burden in a population-based study ofcognitively normal elderly. Neurology, 79(15):1570–7.

Minoshima, S., Foster, N. L., Sima, A. A., Frey, K. A., Albin, R. L. and Kuhl, D. E.2001. Alzheimer’s disease versus dementia with Lewy bodies: cerebral metabolicdistinction with autopsy confirmation. Ann Neurol, 50(3):358–65.

Mirra, S. S., Heyman, A., McKeel, D., Sumi, S. M., Crain, B. J., Brownlee, L. M. et al.1991. The consortium to establish a registry for Alzheimer’s disease (CERAD).Part II. Standardization of the neuropathologic assessment of Alzheimer’s disease.Neurology, 41(4):479–86.

Moll van Charante, E. P., Richard, E., Eurelings, L. S., van Dalen, J.-W., Ligthart, S. A.,van Bussel, E. F. et al. 2016. Effectiveness of a 6-year multidomain vascular careintervention to prevent dementia (preDIVA): a cluster-randomised controlled trial.Lancet, 388(10046):797–805.

Moran, C., Beare, R., Phan, T. G., Bruce, D. G., Callisaya, M. L., Srikanth, V. et al. 2015.Type 2 diabetes mellitus and biomarkers of neurodegeneration. Neurology, 85(13):1123–30.

Morris, J. C. 1993. The clinical dementia rating (CDR): current version and scoringrules. Neurology, 43(11):2412–4.

Morris, J. C., Heyman, A., Mohs, R. C., Hughes, J. P., van Belle, G., Fillenbaum, G.et al. 1989. The consortium to establish a registry for Alzheimer’s disease (CERAD).Part I. Clinical and neuropsychological assessment of Alzheimer’s disease. Neurol-ogy, 39(9):1159–65.

Munoz-Ruiz, M. A., Hartikainen, P., Hall, A., Mattila, J., Koikkalainen, J., Herukka,S.-K. et al. 2013. Disease State Fingerprint in frontotemporal degeneration withreference to Alzheimer’s disease and mild cognitive impairment. J Alzheimers Dis,35(4):727–39.

Mura, T., Baramova, M., Gabelle, A., Artero, S., Dartigues, J.-F., Amieva, H. et al.2017. Predicting dementia using socio-demographic characteristics and the freeand cued selective reminding test in the general population. Alzheimers Res Ther, 9

112

(1):21.

Myllykangas, L., Polvikoski, T., Sulkava, R., Verkkoniemi, A., Crook, R., Tienari, P. J.et al. 1999. Genetic association of alpha2-macroglobulin with Alzheimer’s diseasein a Finnish elderly population. Ann Neurol, 46(3):382–90.

Nag, S., Yu, L., Capuano, A. W., Wilson, R. S., Leurgans, S. E., Bennett, D. A. et al.2015. Hippocampal sclerosis and TDP-43 pathology in aging and Alzheimer dis-ease. Ann Neurol, 77(6):942–52.

Nascimento, C., Di Lorenzo Alho, A. T., Bazan Conceicao Amaral, C., Leite, R. E. P.,Nitrini, R., Jacob-Filho, W. et al. 2018. Prevalence of transactive response DNA-binding protein 43 (TDP-43) proteinopathy in cognitively normal older adults: sys-tematic review and meta-analysis. Neuropathol Appl Neurobiol, 44(3):286–297.

Nascimento, K. K. F. d., Silva, K. P., Malloy-Diniz, L. F., Butters, M. A. and Diniz, B. S.2015. Plasma and cerebrospinal fluid amyloid-β levels in late-life depression: Asystematic review and meta-analysis. J Psychiatr Res, 69:35–41.

Nasrallah, I. M., Pajewski, N. M., Auchus, A. P., Chelune, G., Cheung, A. K., Cleve-land, M. L. et al. 2019. Association of intensive vs standard blood pressure controlwith cerebral white matter lesions. JAMA, 322(6):524–534.

National Academies of Sciences, Engineering, and Medicine and Health. 2017. Pre-venting Cognitive Decline and Dementia: A Way Forward. National Academies Press(US).

Nelson, P. T., Smith, C. D., Abner, E. L., Wilfred, B. J., Wang, W.-X., Neltner, J. H.et al. 2013. Hippocampal sclerosis of aging, a prevalent and high-morbidity braindisease. Acta Neuropathol, 126(2):161–77.

Nelson, P. T., Trojanowski, J. Q., Abner, E. L., Al-Janabi, O. M., Jicha, G. A., Schmitt,F. A. et al. 2016. “new old pathologies”: AD, PART, and cerebral age-related TDP-43 with sclerosis (CARTS). J Neuropathol Exp Neurol, 75(6):482–98.

Ngandu, T., Lehtisalo, J., Levalahti, E., Laatikainen, T., Lindstrom, J., Peltonen,M. et al. 2014. Recruitment and baseline characteristics of participants in theFinnish geriatric intervention study to prevent cognitive impairment and disability(FINGER)—a randomized controlled lifestyle trial. Int J Environ Res Public Health,11(9):9345–60.

Ngandu, T., Lehtisalo, J., Solomon, A., Levalahti, E., Ahtiluoto, S., Antikainen, R.et al. 2015. A 2 year multidomain intervention of diet, exercise, cognitive training,and vascular risk monitoring versus control to prevent cognitive decline in at-riskelderly people (FINGER): a randomised controlled trial. Lancet, 385(9984):2255–63.

NIA-RIA. 1997. Consensus recommendations for the postmortem diagnosisof Alzheimer’s disease. The National Institute on Aging, and Reagan Instituteworking group on diagnostic criteria for the neuropathological assessment ofAlzheimer’s disease. Neurobiol Aging, 18(S4):S1–2.

Noble, W. S. 2006. What is a support vector machine? Nat Biotechnol, 24(12):1565–7.

Norberg, J., Graff, C., Almkvist, O., Ewers, M., Frisoni, G. B., Frolich, L. et al. 2011. Re-

113

gional differences in effects of APOE ε4 on cognitive impairment in non-dementedsubjects. Dement Geriatr Cogn Disord, 32(2):135–42.

Notarianni, E. 2017. Cortisol: Mediator of association between Alzheimer’s diseaseand diabetes mellitus? Psychoneuroendocrinology, 81:129–137.

Nyberg, L., Nilsson, L. G., Olofsson, U. and Backman, L. 1997. Effects of divisionof attention during encoding and retrieval on age differences in episodic memory.Exp Aging Res, 23(2):137–43.

O’Donnell, C. A., Browne, S., Pierce, M., McConnachie, A., Deckers, K., van Boxtel,M. P. J. et al. 2015. Reducing dementia risk by targeting modifiable risk factors inmid-life: study protocol for the innovative midlife intervention for dementia deter-rence (In-MINDD) randomised controlled feasibility trial. Pilot Feasibility Stud, 1:40.

Oh, H., Madison, C., Haight, T. J., Markley, C. and Jagust, W. J. 2012. Effects of ageand β-amyloid on cognitive changes in normal elderly people. Neurobiol Aging, 33(12):2746–55.

Oinas, M., Polvikoski, T., Sulkava, R., Myllykangas, L., Juva, K., Notkola, I.-L.et al. 2009. Neuropathologic findings of dementia with Lewy bodies (DLB) in apopulation-based Vantaa 85+ study. J Alzheimers Dis, 18(3):677–89.

Onyike, C. U., Pletnikova, O., Sloane, K. L., Sullivan, C., Troncoso, J. C. and Rabins,P. V. 2013. Hippocampal sclerosis dementia: An amnesic variant of frontotemporaldegeneration. Dement Neuropsychol, 7(1):83–87.

Ossenkoppele, R., Jansen, W. J., Rabinovici, G. D., Knol, D. L., van der Flier, W. M.,van Berckel, B. N. M. et al. 2015. Prevalence of amyloid PET positivity in dementiasyndromes: a meta-analysis. JAMA, 313(19):1939–49.

Palmqvist, S., Insel, P. S., Zetterberg, H., Blennow, K., Brix, B., Stomrud, E. et al.2019. Accurate risk estimation of β-amyloid positivity to identify prodromalAlzheimer’s disease: Cross-validation study of practical algorithms. AlzheimersDement, 15(2):194–204.

Peltonen, M., Korpi-Hyovalti, E., Oksa, H., Puolijoki, H., Saltevo, J., Vanhala, M. et al.2006. Prevalence of obesity, type 2 diabetes, and other disturbances in glucosemetabolism in Finland—the FIN-D2D survey. Finnish Medical Journal, 61(3):163–170.

Pentikainen, H., Savonen, K., Ngandu, T., Solomon, A., Komulainen, P., Paajanen, T.et al. 2019. Cardiorespiratory fitness and cognition: Longitudinal associations inthe FINGER study. J Alzheimers Dis, 68(3):961–968.

Perk, J., De Backer, G., Gohlke, H., Graham, I., Reiner, Z., Verschuren, M. et al. 2012.European guidelines on cardiovascular disease prevention in clinical practice (ver-sion 2012). Eur Heart J, 33(13):1635–701.

Petersen, R. C., Wiste, H. J., Weigand, S. D., Rocca, W. A., Roberts, R. O., Mielke, M. M.et al. 2016. Association of elevated amyloid levels with cognition and biomarkersin cognitively normal people from the community. JAMA Neurol, 73(1):85–92.

114

Petrovitch, H., White, L. R., Izmirilian, G., Ross, G. W., Havlik, R. J., Markesbery, W.et al. 2000. Midlife blood pressure and neuritic plaques, neurofibrillary tangles,and brain weight at death: the HAAS. Neurobiol Aging, 21(1):57–62.

Pfeiffer, E. 1975. A short portable mental status questionnaire for the assessment oforganic brain deficit in elderly patients. J Am Geriatr Soc, 23(10):433–41.

Polvikoski, T., Sulkava, R., Haltia, M., Kainulainen, K., Vuorio, A., Verkkoniemi, A.et al. 1995. Apolipoprotein E, dementia, and cortical deposition of beta-amyloidprotein. N Engl J Med, 333(19):1242–7.

Powell, G. E. 1980. The subjective memory questionnaire (SMQ): An investigationinto the self-reporting of “real-life” memory skills. British Journal of Social & ClinicalPsychology, 19(2):177–188.

Power, M. C., Weuve, J., Gagne, J. J., McQueen, M. B., Viswanathan, A. and Blacker,D. 2011. The association between blood pressure and incident Alzheimer disease:a systematic review and meta-analysis. Epidemiology, 22(5):646–59.

Profenno, L. A., Porsteinsson, A. P. and Faraone, S. V. 2010. Meta-analysis ofAlzheimer’s disease risk with obesity, diabetes, and related disorders. Biol Psy-chiatry, 67(6):505–12.

Puska, P., Tuomilehto, J., Salonen, J., Neittaanmaki, L., Maki, J., Virtamo, J. et al.1979. Changes in coronary risk factors during comprehensive five-year communityprogramme to control cardiovascular diseases (North Karelia project). Br Med J, 2(6199):1173–8.

Puska, P., Salonen, J. T., Nissinen, A., Tuomilehto, J., Vartiainen, E., Korhonen, H.et al. 1983. Change in risk factors for coronary heart disease during 10 years ofa community intervention programme (North Karelia project). Br Med J (Clin ResEd), 287(6408):1840–4.

Qiu, C., Kivipelto, M., Aguero-Torres, H., Winblad, B. and Fratiglioni, L. 2004. Riskand protective effects of the APOE gene towards Alzheimer’s disease in the Kung-sholmen project: variation by age and sex. J Neurol Neurosurg Psychiatry, 75(6):828–33.

Rastas, S., Pirttila, T., Mattila, K., Verkkoniemi, A., Juva, K., Niinisto, L. et al. 2010.Vascular risk factors and dementia in the general population aged >85 years:prospective population-based study. Neurobiol Aging, 31(1):1–7.

Rhodius-Meester, H. F. M., Koikkalainen, J., Mattila, J., Teunissen, C. E., Barkhof, F.,Lemstra, A. W. et al. 2016. Integrating biomarkers for underlying Alzheimer’sdisease in mild cognitive impairment in daily practice: Comparison of a clinicaldecision support system with individual biomarkers. J Alzheimers Dis, 50(1):261–70.

Ritchie, C. W. and Muniz-Terrera, G. 2019. Models for dementia risk prediction:so much activity brings a need for coordination and clarity. J Neurol NeurosurgPsychiatry, 90(4):372.

Ritchie, C. W., Molinuevo, J. L., Truyen, L., Satlin, A., Van der Geyten, S., Love-

115

stone, S. et al. 2016. Development of interventions for the secondary prevention ofAlzheimer’s dementia: the European Prevention of Alzheimer’s Dementia (EPAD)project. Lancet Psychiatry, 3(2):179–86.

Roberts, R. and Knopman, D. S. 2013. Classification and epidemiology of MCI. ClinGeriatr Med, 29(4):753–72.

Roberts, R. O., Knopman, D. S., Cha, R. H., Mielke, M. M., Pankratz, V. S., Boeve, B. F.et al. 2014. Diabetes and elevated hemoglobin A1c levels are associated with brainhypometabolism but not amyloid accumulation. J Nucl Med, 55(5):759–64.

Roman, G. C., Tatemichi, T. K., Erkinjuntti, T., Cummings, J. L., Masdeu, J. C., Garcia,J. H. et al. 1993. Vascular dementia: diagnostic criteria for research studies. Reportof the NINDS-AIREN international workshop. Neurology, 43(2):250–60.

Rouch, L., Cestac, P., Hanon, O., Cool, C., Helmer, C., Bouhanick, B. et al. 2015. An-tihypertensive drugs, prevention of cognitive decline and dementia: a systematicreview of observational studies, randomized controlled trials and meta-analyses,with discussion of potential mechanisms. CNS Drugs, 29(2):113–30.

Rovio, S., Spulber, G., Nieminen, L. J., Niskanen, E., Winblad, B., Tuomilehto, J. et al.2010. The effect of midlife physical activity on structural brain changes in the el-derly. Neurobiol Aging, 31(11):1927–36.

Rowe, C. C., Ellis, K. A., Rimajova, M., Bourgeat, P., Pike, K. E., Jones, G. et al. 2010.Amyloid imaging results from the Australian imaging, biomarkers and lifestyle(AIBL) study of aging. Neurobiol Aging, 31(8):1275–83.

Rusanen, M., Kivipelto, M., Quesenberry, C. P., Jr, Zhou, J. and Whitmer, R. A. 2011.Heavy smoking in midlife and long-term risk of Alzheimer disease and vasculardementia. Arch Intern Med, 171(4):333–9.

Sabri, O., Sabbagh, M. N., Seibyl, J., Barthel, H., Akatsu, H., Ouchi, Y. et al. 2015.Florbetaben PET imaging to detect amyloid beta plaques in Alzheimer’s disease:phase 3 study. Alzheimers Dement, 11(8):964–74.

Sachdev, P., Kalaria, R., O’Brien, J., Skoog, I., Alladi, S., Black, S. E. et al. 2014. Di-agnostic criteria for vascular cognitive disorders: A VASCOG statement. AlzheimerDis Assoc Disord.

Salmon, D. P. and Bondi, M. W. 2009. Neuropsychological assessment of dementia.Annu Rev Psychol, 60:257–82.

Saunders, A. M., Strittmatter, W. J., Schmechel, D., George-Hyslop, P. H., Pericak-Vance, M. A., Joo, S. H. et al. 1993. Association of apolipoprotein E allele epsilon4 with late-onset familial and sporadic Alzheimer’s disease. Neurology, 43(8):1467–72.

Scarmeas, N., Anastasiou, C. A. and Yannakoulia, M. 2018. Nutrition and preventionof cognitive impairment. Lancet Neurol, 17(11):1006–1015.

Scheltens, P., Leys, D., Barkhof, F., Huglo, D., Weinstein, H. C., Vermersch, P. et al.1992. Atrophy of medial temporal lobes on mri in “probable” Alzheimer’s diseaseand normal ageing: diagnostic value and neuropsychological correlates. J Neurol

116

Neurosurg Psychiatry, 55(10):967–72.

Schilling, S., DeStefano, A. L., Sachdev, P. S., Choi, S. H., Mather, K. A., DeCarli,C. D. et al. 2013. APOE genotype and MRI markers of cerebrovascular disease:systematic review and meta-analysis. Neurology, 81(3):292–300.

Schultz, S. A., Boots, E. A., Almeida, R. P., Oh, J. M., Einerson, J., Korcarz, C. E. et al.2015. Cardiorespiratory fitness attenuates the influence of amyloid on cognition. JInt Neuropsychol Soc, 21(10):841–50.

Seblova, D., Quiroga, M. L., Fors, S., Johnell, K., Lovden, M., de Leon, A. P. et al.2018. Thirty-year trends in dementia: a nationwide population study of Swedishinpatient records. Clin Epidemiol, 10:1679–1693.

Seppala, T., Herukka, S.-K. and Remes, A. M. 2013. Alzheimerin taudin varhaisdiag-nostiikka. Duodecim, 129:2003–2010.

Sharp, S. I., Aarsland, D., Day, S., Sønnesyn, H., Alzheimer’s Society Vascular De-mentia Systematic Review Group and Ballard, C. 2011. Hypertension is a potentialrisk factor for vascular dementia: systematic review. Int J Geriatr Psychiatry, 26(7):661–9.

Shinohara, M., Murray, M. E., Frank, R. D., Shinohara, M., DeTure, M., Yamazaki,Y. et al. 2016. Impact of sex and APOE4 on cerebral amyloid angiopathy inAlzheimer’s disease. Acta Neuropathol, 132(2):225–34.

Singer, M., Deutschman, C. S., Seymour, C. W., Shankar-Hari, M., Annane, D., Bauer,M. et al. 2016. The third international consensus definitions for sepsis and septicshock (sepsis-3). JAMA, 315(8):801–10.

Skoog, I., Hesse, C., Aevarsson, O., Landahl, S., Wahlstrom, J., Fredman, P. et al.1998. A population study of apoE genotype at the age of 85: relation to dementia,cerebrovascular disease, and mortality. J Neurol Neurosurg Psychiatry, 64(1):37–43.

Smith, E. E. and Greenberg, S. M. 2009. Beta-amyloid, blood vessels, and brainfunction. Stroke, 40(7):2601–6.

Sofi, F., Valecchi, D., Bacci, D., Abbate, R., Gensini, G. F., Casini, A. et al. 2011. Physicalactivity and risk of cognitive decline: a meta-analysis of prospective studies. JIntern Med, 269(1):107–17.

Solomon, A., Mangialasche, F., Richard, E., Andrieu, S., Bennett, D. A., Breteler, M.et al. 2014a. Advances in the prevention of Alzheimer’s disease and dementia. JIntern Med, 275(3):229–50.

Solomon, A., Kivipelto, M., Wolozin, B., Zhou, J. and Whitmer, R. A. 2009. Midlifeserum cholesterol and increased risk of Alzheimer’s and vascular dementia threedecades later. Dement Geriatr Cogn Disord, 28(1):75–80.

Solomon, A., Ngandu, T., Soininen, H., Hallikainen, M. M., Kivipelto, M. andLaatikainen, T. 2014b. Validity of dementia and Alzheimer’s disease diagnosesin Finnish national registers. Alzheimers Dement, 10(3):303–9.

Solomon, A., Ngandu, T. and Kivipelto, M. From prediction to dementia preven-

117

tion. In Irving, K., Hogervorst, E., Oliveira, D. and Kivipelto, M., editors, NewDevelopments in Dementia Prevention Research: State of the Art and Future Possibilities.Routledge, 1st edition, 2019.

Sperling, R. A., Aisen, P. S., Beckett, L. A., Bennett, D. A., Craft, S., Fagan, A. M. et al.2011. Toward defining the preclinical stages of Alzheimer’s disease: Recommen-dations from the National Institute on Aging-Alzheimer’s Association workgroupson diagnostic guidelines for Alzheimer’s disease. Alzheimers Dement, 7(3):280–92.

Stephen, R., Hongisto, K., Solomon, A. and Lonnroos, E. 2017a. Physical activityand alzheimer’s disease: A systematic review. J Gerontol A Biol Sci Med Sci, 72(6):733–739.

Stephen, R., Liu, Y., Ngandu, T., Rinne, J. O., Kemppainen, N., Parkkola, R. et al.2017b. Associations of CAIDE dementia risk score with MRI, PIB-PET measures,and cognition. J Alzheimers Dis, 59(2):695–705.

Stern, Y., Arenaza-Urquijo, E. M., Bartres-Faz, D., Belleville, S., Cantilon, M., Chetelat,G. et al. 2018. Whitepaper: Defining and investigating cognitive reserve, brainreserve, and brain maintenance. Alzheimers Dement, Epub ahead of print.

Stiell, I. G., Greenberg, G. H., McKnight, R. D., Nair, R. C., McDowell, I. and Wor-thington, J. R. 1992. A study to develop clinical decision rules for the use of radio-graphy in acute ankle injuries. Ann Emerg Med, 21(4):384–90.

Stroop, J. 1935. Studies of inference in serial verbal reaction. Journal of ExperimentalPsychology, 18.

Suthaharan, S. 2015. Machine Learning Models and Algorithms for Big Data Classification:Thinking with Examples for Effective Learning. Springer.

Syvanen, A. C., Sajantila, A. and Lukka, M. 1993. Identification of individuals byanalysis of biallelic DNA markers, using PCR and solid-phase minisequencing.Am J Hum Genet, 52(1):46–59.

Tang, E. Y. H., Harrison, S. L., Errington, L., Gordon, M. F., Visser, P. J., Novak, G. et al.2015. Current developments in dementia risk prediction modelling: An updatedsystematic review. PLoS One, 10(9):e0136181.

Tanskanen, M., Makela, M., Myllykangas, L., Rastas, S., Sulkava, R. and Paetau, A.2012. Intracerebral hemorrhage in the oldest old: a population-based study (Vantaa85+). Front Neurol, 3:103.

ten Kate, M., Redolfi, A., Peira, E., Bos, I., Vos, S. J., Vandenberghe, R. et al. 2018. MRIpredictors of amyloid pathology: results from the EMIF-AD multimodal biomarkerdiscovery study. Alzheimers Res Ther, 10(1):100.

Tiffin, J. Purdue Pegboard Examiner’s Manual. London House, Rosemont, IL, 1968.

Toledo, J. B., Toledo, E., Weiner, M. W., Jack, C. R., Jr, Jagust, W., Lee, V. M. Y.et al. 2012. Cardiovascular risk factors, cortisol, and amyloid-β deposition inAlzheimer’s Disease Neuroimaging Initiative. Alzheimers Dement, 8(6):483–9.

Tolonen, A., Rhodius-Meester, H. F. M., Bruun, M., Koikkalainen, J., Barkhof, F., Lem-

118

stra, A. W. et al. 2018. Data-driven differential diagnosis of dementia using multi-class Disease State Index classifier. Front Aging Neurosci, 10:111.

Tosun, D., Joshi, S., Weiner, M. W. and Alzheimer’s Disease Neuroimaging Initiative.2013. Neuroimaging predictors of brain amyloidosis in mild cognitive impairment.Ann Neurol, 74(2):188–98.

Tosun, D., Joshi, S., Weiner, M. W. and the Alzheimer’s Disease Neuroimaging Initia-tive. 2014. Multimodal MRI-based imputation of the Aβ+ in early mild cognitiveimpairment. Ann Clin Transl Neurol, 1(3):160–170.

Tsukamoto, K., Watanabe, T., Matsushima, T., Kinoshita, M., Kato, H., Hashimoto, Y.et al. 1993. Determination by PCR-RFLP of apo E genotype in a Japanese popula-tion. J Lab Clin Med, 121(4):598–602.

Vartiainen, E., Puska, P., Jousilahti, P., Korhonen, H. J., Tuomilehto, J. and Nissinen,A. 1994. Twenty-year trends in coronary risk factors in north Karelia and in otherareas of Finland. Int J Epidemiol, 23(3):495–504.

Walker, K. A., Sharrett, A. R., Wu, A., Schneider, A. L. C., Albert, M., Lutsey, P. L.et al. 2019. Association of midlife to late-life blood pressure patterns with incidentdementia. JAMA, 322(6):535–545.

Walker, Z., Possin, K. L., Boeve, B. F. and Aarsland, D. 2015. Lewy body dementias.Lancet, 386(10004):1683–97.

Walters, K., Hardoon, S., Petersen, I., Iliffe, S., Omar, R. Z., Nazareth, I. et al. 2016.Predicting dementia risk in primary care: development and validation of the de-mentia risk score using routinely collected data. BMC Med, 14:6.

Wechsler, D. Wechsler Adult Intelligence Scale Manual. Psychological Corporation, NewYork, 1944.

Wermer, M. J. H. and Greenberg, S. M. 2018. The growing clinical spectrum of cerebralamyloid angiopathy. Curr Opin Neurol, 31(1):28–35.

Westwood, S., Liu, B., Baird, A. L., Anand, S., Nevado-Holgado, A. J., Newby, D.et al. 2017. The influence of insulin resistance on cerebrospinal fluid and plasmabiomarkers of Alzheimer’s pathology. Alzheimers Res Ther, 9(1):31.

Westwood, S., Baird, A. L., Hye, A., Ashton, N. J., Nevado-Holgado, A. J., Anand,S. N. et al. 2018. Plasma protein biomarkers for the prediction of CSF amyloid andtau and [18F]-flutemetamol PET scan result. Front Aging Neurosci, 10:409.

Willette, A. A., Johnson, S. C., Birdsill, A. C., Sager, M. A., Christian, B., Baker, L. D.et al. 2015. Insulin resistance predicts brain amyloid deposition in late middle-agedadults. Alzheimers Dement, 11(5):504–510.e1.

Williamson, J. D., Pajewski, N. M., Auchus, A. P., Bryan, R. N., Chelune, G., Cheung,A. K. et al. 2019. Effect of intensive vs standard blood pressure control on probabledementia: A randomized clinical trial. JAMA, 321(6):553–561.

Wirth, M., Madison, C. M., Rabinovici, G. D., Oh, H., Landau, S. M. and Jagust, W. J.2013. Alzheimer’s disease neurodegenerative biomarkers are associated with de-

119

creased cognitive function but not β-amyloid in cognitively normal older individ-uals. J Neurosci, 33(13):5553–63.

Wolters, F. J., Segufa, R. A., Darweesh, S. K. L., Bos, D., Ikram, M. A., Sabayan, B. et al.2018. Coronary heart disease, heart failure, and the risk of dementia: A systematicreview and meta-analysis. Alzheimers Dement, 14(11):1493–1504.

World Health Organization. 1993. The ICD-10 Classification of Mental and BehaviouralDisorders: Diagnostic criteria for research. World Health Organization, Geneva.

World Health Organization. 2019. Risk reduction of cognitive decline and dementia: WHOguidelines. World Health Organization.

Xu, W., Yu, J.-T., Tan, M.-S. and Tan, L. 2015. Cognitive reserve and Alzheimer’sdisease. Mol Neurobiol, 51(1):187–208.

Yasuno, F., Kazui, H., Morita, N., Kajimoto, K., Ihara, M., Taguchi, A. et al. 2015. Lowamyloid-β deposition correlates with high education in cognitively normal olderadults: a pilot study. Int J Geriatr Psychiatry, 30(9):919–26.

Zung, W. W., Richards, C. B. and Short, M. J. 1965. Self-rating depression scale in anoutpatient clinic. Further validation of the SDS. Arch Gen Psychiatry, 13(6):508–15.

120

ORIGINAL PUBLICATIONS (I–IV)

I

Development of a late-life dementia prediction index with supervised machinelearning in the population-based CAIDE study

Pekkala T, Hall A, Lotjonen J, Mattila J, Soininen H, Ngandu T, Laatikainen T,Kivipelto M and Solomon A

Journal of Alzheimer’s Disease 55: 1055–1067, 2017

Reprinted with the kind permission of IOS Press

Journal of Alzheimer’s Disease 55 (2017) 1055–1067DOI 10.3233/JAD-160560IOS Press

1055

Development of a Late-Life DementiaPrediction Index with Supervised MachineLearning in the Population-Based CAIDEStudy

Timo Pekkalaa, Anette Halla, Jyrki Lotjonenb,c, Jussi Mattilac, Hilkka Soininena,d, Tiia Ngandue,Tiina Laatikainene,f,g, Miia Kivipeltoa,e,h and Alina Solomona,h,∗aInstitute of Clinical Medicine, Neurology, University of Eastern Finland, Kuopio, FinlandbVTT Technical Research Centre of Finland, Tampere, FinlandcCombinostics, Tampere, FinlanddDepartment of Neurology, Kuopio University Hospital, Kuopio, FinlandeChronic Disease Prevention Unit, National Institute for Health and Welfare, Helsinki, Finlandf Institute of Public Health and Clinical Nutrition, University of Eastern Finland, Kuopio, FinlandgHospital District of North Karelia, Joensuu, FinlandhDivision of Clinical Geriatrics, Center for Alzheimer Research, Karolinska Institutet, Stockholm, Sweden

Accepted 12 September 2016

Abstract.Background and objective: This study aimed to develop a late-life dementia prediction model using a novel validatedsupervised machine learning method, the Disease State Index (DSI), in the Finnish population-based CAIDE study.Methods: The CAIDE study was based on previous population-based midlife surveys. CAIDE participants were re-examinedtwice in late-life, and the first late-life re-examination was used as baseline for the present study. The main study populationincluded 709 cognitively normal subjects at first re-examination who returned to the second re-examination up to 10 yearslater (incident dementia n = 39). An extended population (n = 1009, incident dementia 151) included non-participants/non-survivors (national registers data). DSI was used to develop a dementia index based on first re-examination assessments.Performance in predicting dementia was assessed as area under the ROC curve (AUC).Results: AUCs for DSI were 0.79 and 0.75 for main and extended populations. Included predictors were cognition, vascularfactors, age, subjective memory complaints, and APOE genotype.Conclusion: The supervised machine learning method performed well in identifying comprehensive profiles for predictingdementia development up to 10 years later. DSI could thus be useful for identifying individuals who are most at risk and maybenefit from dementia prevention interventions.

Keywords: Computer-assisted decision making, dementia, prediction, prevention, supervised machine learning

INTRODUCTION

Dementia prevention is a high public health prior-ity. With many reported modifiable risk factors [1],

∗Correspondence to: Alina Solomon, MD, PhD, Institute ofClinical Medicine, Neurology, University of Eastern Finland, POB1627, 70211 Kuopio, Finland. Tel.: +358403552015; E-mail:[email protected].

and several ongoing large multimodal preventiontrials [2, 3], the interest in dementia prediction mod-els has grown during the past years. Similarly torisk scores for cardiovascular disease [4], dementiarisk scores could be used to identify at-risk indi-viduals who would benefit most from preventiveinterventions. Dementia risk profiling could addition-ally facilitate the tailoring of preventive interventions

ISSN 1387-2877/17/$35.00 © 2017 – IOS Press and the authors. All rights reservedThis article is published online with Open Access and distributed under the terms of the Creative Commons Attribution Non-Commercial License (CC BY-NC 4.0).

1056 T. Pekkala et al. / Late-Life Dementia Prediction Index

to target the most relevant risk factors for a specificindividual or group.

Several dementia prediction models have beenreported [5, 6]. Model development has been basedmainly on a data analytical approach (logistic or Coxproportional hazards regression analyses), and in onecase on an Evidence-Based Medicine approach [5, 7].The increasing number and complexity of factors andbiomarkers related to dementia risk, and limitationsin visualizing and interpreting individual risk pro-files represent major challenges for such methods ofdeveloping dementia prediction models.

One of the few validated dementia risk scores[8, 9] has already been used to select at-risk elderlyfrom the general population participating in a suc-cessful prevention trial [2], and is available foruse with both pen-and-paper and computer-basedtechnology (mobile app, online tool) [10]. The use-fulness of computerized dementia prediction tools forprevention-related decision-making is only startingto be explored. As comprehensive online preventionresearch resources and e-Health solutions are startingto be developed for both health care profession-als and general public (e.g., Brain Health Registry,multinational data discovery and sharing platforms,internet-based prevention trials [11], clinical decisionsupport systems integratable with electronic healthrecords [12]), it is increasingly important to find suit-able methods for developing, updating, and easilyvisualizing and interpreting complex dementia riskprofiles.

The Disease State Index (DSI) is a supervisedmachine learning method designed for practicalimplementation as a clinical decision support sys-tem [12]. DSI has been extensively tested and shownto perform well in the context of improving earlydiagnosis of Alzheimer’s disease and differentialdiagnosis of neurodegenerative diseases [12–20].However, the use of DSI in a public health/dementiaprevention context has so far not been investigated,i.e., predicting dementia in a general population with-out cognitive impairment. Compared to previouslyused methods for developing dementia risk scores[5], the main strengths of DSI are its ability to dealwith larger amounts of heterogeneous data, to han-dle missing data well, and to use unprocessed data(i.e., without any pre-specified cut-offs for clinical orbiomarker variables). In addition, DSI is accompa-nied by the Disease State Fingerprint (DSF), a methodfor presenting DSI data in an easily and quicklyinterpretable visual form. The present study aims todevelop a late-life dementia prediction model using

DSI in the longitudinal population-based CAIDEstudy.

MATERIALS AND METHODS

The CAIDE study

The CAIDE study has been previously describedin detail [21–23]. In brief, participants were firstevaluated at midlife (1972, 1977, 1982, or 1987) incardiovascular surveys. A random sample of 2,000individuals aged 65–79 at the end of 1997, andliving in or close to Kuopio and Joensuu regionsin Eastern Finland were invited for a first late-lifere-examination in 1998 (Fig. 1). Altogether 1,449(72.5%) individuals participated. A second late-lifere-examination was conducted in 2005–2008. Of theinitial 2,000 persons, 1,426 were still alive and liv-ing in the region in the beginning of 2005, and909 (63.7%) participated. Mean age (SD) was 50.6(6.0) years at midlife, 71.3 (4.0) years at the firstre-examination, and 78.6 (3.7) years at the secondre-examination. The CAIDE study was approved bythe local ethics committee of Kuopio University Hos-pital and written informed consent was obtained fromall participants.

In both late-life re-examinations, cognition wasassessed using a three-step protocol (screening, clin-ical, and differential diagnostic phases). In 1998,participants with ≤24 points on the Mini-MentalState Examination (MMSE) [24] at screening werereferred for further evaluations. In 2005–2008, sub-jects with ≤24 points or decline ≥3 points on MMSE,<70% delayed recall in the CERAD word list [25], orwith informant concerns about the participant’s cog-nition were referred for further evaluations. In bothre-examinations, the clinical phase included detailedmedical and neuropsychological assessments, and thedifferential diagnostic phase included brain imaging(MRI/CT), blood tests, and if needed cerebrospinalfluid analysis.

A review board including the study physician,neuropsychologist, a senior neuropsychologist, anda senior neurologist ascertained the primary diag-nosis based on all available information. Dementiaand mild cognitive impairment (MCI) diagnoses weremade according to established criteria [26–28].

Design of the present study

The present study focused on CAIDE partic-ipants without dementia or MCI in 1998 (first

T. Pekkala et al. / Late-Life Dementia Prediction Index 1057

Fig. 1. Formation of the study populations.

late-life re-examination, used here as baseline).The main study population included 709 indi-viduals who also participated in the 2005–2008re-examination (39 diagnosed with dementia). Meanfollow-up (SD) was 8.3 (1.0) years. To accountfor non-participants/non-survivors in 2005–2008, anextended study population (n = 1,009) was definedusing additional data on dementia diagnoses until the

end of 2008 from the Hospital Discharge Register,Drug Reimbursement Register and Causes of DeathRegister [22]. Dementia cases in the extended pop-ulation (n = 151) were defined according to CAIDEor register diagnoses (CAIDE diagnoses had priority,except when registers indicated dementia diagnosesafter the second re-examination and before the endof 2008). Mean follow-up (SD) was 9.0 (1.4) years,


and mean time (SD) to dementia diagnosis was 7.1(1.9) years. Non-participants in 2005–2008 who haddied without a recorded dementia diagnosis (n = 244)could not be classified as cases or controls and wereexcluded. Additionally 13 subjects without cognitiveimpairment in 1998 who had a dementia diagnosisin any register before the end of 2000 were excluded(they were considered too close to dementia onset).

Factors included in prediction models

Survey methods were carefully standardized andcomplied with international recommendations [29].Cognitive performance in 1998 was included inprediction models. Five cognitive domains wereassessed as previously described [30]: global cog-nition (MMSE), episodic memory (mean numberof recalled words from three 10-word lists), verbalexpression (one-minute animal naming test), psy-chomotor speed (mean of normalized scores fromLetter Digit Substitution and bimanual Purdue Peg-board tests), executive functioning (time differencebetween the color word interference and naming tasksin the Stroop test), and prospective memory (remind-ing the investigator to make a phone call at the end ofthe testing session; score 1–4 from not rememberingto remembering without reminders).

Vascular factors (to (blood pressure (BP), bodymass index (BMI), waist-hip ratio, total choles-terol, high-density lipoprotein (HDL) cholesterol andtriglycerides) were assessed at each examination.Assessments from 1998 were included in the basicmodel. Changes in BP, BMI, and total cholesterolfrom midlife to the first re-examination in 1998were included in an additional model. Diagnoses ofstroke, transient ischemic attack, myocardial infarc-tion, coronary heart disease, atrial fibrillation, heartfailure, or diabetes (Hospital Discharge Register)were combined into a dichotomous comorbidity vari-able.

Other assessments from 1998 used in thepresent study included data from a self-administeredquestionnaire on sociodemographic characteristics,medical history and health-related behavior, e.g.,leisure-time physical activity, alcohol use, smoking,self-rated health, and fitness, feelings of hopelessness[31], Beck Depression Inventory [32], and SubjectiveMemory Questionnaire (SMQ) [33].

Apolipoprotein E (APOE) genotypes wereassessed from blood leucocytes using polymerasechain reaction and HhaI digestion [34]. APOEwas modeled as a dichotomous variable (�4 allele

carrier/non-carrier), and also as an ordered variable(genotype 23 < 24 and 33 < 34 < 44) [35, 36].

Disease state index and disease state fingerprint

DSI has been previously described in detail [12,13]. In brief, DSI is a validated supervised machinelearning method that provides numeric index valuesranging from 0 to 1. The DSI value is computedby comparing an individual to a previously knownpopulation (training data). The DSI value can beinterpreted as the share of data corresponding to asubsequent dementia profile. DSI value 0 correspondsto an ideal control, and 1 to an ideal subsequentdementia case. Higher DSI values thus denote greaterprofile similarity to individuals known to subse-quently develop dementia in the training population.

DSI values are computed in three steps. First,each measurement is compared with the trainingdata using a monotonically increasing fitness func-tion that provides a likelihood of the measured factorbelonging to an individual who will develop demen-tia. The fitness as a function of measurement valuex, is defined as f (x) = FN(x)

FN(x)+FP(x) , where FN(x)is the false negative error rate and FP(x) the falsepositive error rate in the training data, when usingx as the classification threshold. Second, the rele-vance of each measurement is calculated, indicatinghow well the measurement can discriminate betweenindividuals who will develop dementia and thosewho will not. Relevance is computed as relevance =sensitivity + specificity − 1, where sensitivity andspecificity are obtained by classifying the diagnosedpopulation. Third, fitness and relevance values arecombined into a composite factor group DSI valueusing a weighted average, where the fitness valuesare weighted according to their relevance: DSI =∑

relevance × fitness∑relevance

. The process of evaluating fitness

and relevance and combining measurements into acomposite group DSI are repeated recursively until anoverall DSI value from all available data is obtainedfor the individual.

DSI can process heterogeneous data, and the mea-sured factors/biomarkers are structured into groups,e.g., different cognitive tests into a Cognition groupor vascular factors into a Vascular group. A com-posite DSI value is calculated for each group basedon the included individual factors. Grouping is thususeful for assessing the combined effect of concep-tually related measurements, and it has other effectssuch as filtering out noise at group level, and ensuring


that strongly correlated factors are not added into themodel multiple times. Missing data does not affectmodel building as long as there is enough data foreach factor to give a reliable distribution.

The DSF visualization gives a comprehensiveoverview of an individual’s predictive profile [13],showing which factors are most relevant and to whatextent they correspond to a subject who will developdementia. An example with explanations is shown inthe Supplementary Material.

Data analysis

Differences between control and dementia groupswere determined with Mann-Whitney U test forcontinuous or ordinal variables, and χ2 test forother categorical variables. Significance level wasset at p < 0.05. Only factors significantly differentbetween control and dementia groups were pre-selected into the DSI model. Additional p-valuesignificance thresholds for selecting factors into themodel were also tested to assess effects on predictiveperformance.

Performance of DSI in predicting dementia wasevaluated using a stratified cross-validation proce-dure. Analysis was performed using 50 × 5-folds.The performance of DSI was measured as the areaunder the receiver operating characteristic curve(AUC), by averaging AUCs from individual folds.DSI classification results were validated by compar-ison with a commonly used machine learning model,support vector machine (SVM), using the same data.Analyses were conducted using Matlab R2014a.

RESULTS

Population characteristics

Population characteristics in 1998 by dementia sta-tus until the end of 2008 are shown in Table 1. Inthe main study population, individuals with subse-quent dementia were older, had significantly poorerperformance on four of the six cognitive tests,had lower systolic blood pressure (SBP) and dias-tolic blood pressure (DBP), higher frequency ofcardio/cerebrovascular comorbidity and the APOE�4 allele, and more pronounced subjective memorycomplaints (total SMQ score and four items aboutforgetting phone numbers, clothing size, names ofactors, and forgetting what to say in mid-sentence).SBP, DBP, and BMI decreased more between midlife

and 1998 in subjects with subsequent dementia com-pared with controls.

In the extended study population, individuals withdementia were older, had significantly poorer per-formance on all six cognitive tests, higher frequencyof cardio/cerebrovascular comorbidity and the APOE�4 allele, and more pronounced subjective memorycomplaints (total SMQ score and one item about for-getting phone numbers). No differences were foundin SBP or DBP. Changes in DBP, total cholesterol, andBMI (but not SBP) between midlife and 1998 weredifferent between controls and subsequent dementiacases.

Performance of DSI in predicting dementia

Table 2 shows AUCs (95% CI) for the compos-ite DSI including factor groups Cognition, Vascularfactors, Demographics, Subjective memory ques-tionnaire, and APOE genotype (basic model). Thecomposite DSI achieved an AUC of 0.79 (0.79–0.80)in the main study population, and 0.75 (0.74–0.75) inthe extended study population. Training the DSI onthe entire main or extended population and using itto classify the same cases yielded AUCs of 0.84 and0.76, respectively.

There was an overall pattern of similar to somewhatlower AUCs for individual factors and factor groupsin the extended population compared with the mainstudy population. ROC curves for the composite DSIin both populations are shown in Fig. 2. Accuracy,sensitivity and specificity for different composite DSIcut-off values are shown in Table 3.

AUC (95% CI) for the composite DSI includingthe basic model plus changes in vascular factors frommidlife to late-life are shown in Table 2. There was aslight increase in AUCs for composite DSI comparedwith the basic model. AUCs for changes in vascu-lar factors considered together were slightly higherthan AUCs for the group of late-life vascular factors,and this difference was most pronounced in theextended study population. Change in BMI had thehighest AUC (0.68) for both main and extended studypopulations.

Sensitivity analyses

Table 4 shows the effects of p-value threshold filter-ing on the number of factors included in the predictionmodel, and on AUCs (95% CI) for the composite DSI.Analyses focused on p-values from Mann-WhitneyU-tests comparing controls and subsequent dementia


Tabl

e1

Cha

ract

eris

tics

ofth

est

udy

popu

latio

ns

Mai

nst

udy

popu

latio

nE

xten

ded

stud

ypo

pula

tion

(par

ticip

ants

/sur

vivo

rs)

nC

ontr

olD

emen

tiap-

valu

en

Con

trol

Dem

entia

p-va

lue

Soci

o-de

mog

raph

icch

arac

teri

stic

sA

ge67

0/39

70.0

(3.4

)72

.4(4

.1)

<0.0

0185

8/15

170

.2(3

.6)

72.4

(3.9

)<0

.001

Edu

catio

n(y

ears

)65

9/39

9.3

(3.5

)9.

2(4

.1)

0.47

845/

148

9.0

(3.5

)9.

0(3

.8)

0.38

Wom

en67

0/39

64.2

%74

.4%

0.20

858/

151

65.4

%66

.9%

0.72

Cog

niti

onM

MSE

(0-3

0p)

669/

3926

.6(1

.6)

26.2

(1.8

)0.

3385

3/14

926

.5(1

.7)

26.1

(1.9

)0.

044

Ver

bale

xpre

ssio

n67

0/39

21.3

(6.0

)19

.5(5

.0)

0.15

854/

148

20.9

(6.0

)19

.5(5

.6)

0.02

8Pr

ospe

ctiv

em

emor

y(1

-4p)

585/

342.

9(0

.8)

2.4

(0.9

)0.

001

756/

129

2.8

(0.8

)2.

4(0

.8)

<0.0

01E

piso

dic

mem

ory

(0-1

0p)

662/

395.

4(1

.1)

4.8

(1.0

)0.

002

846/

150

5.3

(1.1

)4.

7(1

.3)

<0.0

01Ps

ycho

mot

orsp

eed

635/

380.

02(0

.8)

–0.4

(1.0

)0.

011

809/

136

0.1

(0.8

)–0

.4(0

.9)

<0.0

01E

xecu

tive

func

tioni

ng64

9/38

36.2

(17.

5)48

.3(2

1.5)

<0.0

0182

4/14

337

.5(1

8.7)

46.6

(25.

9)<0

.001

Vasc

ular

&li

fest

yle

fact

ors

SBP

(mm

Hg)

669/

3915

1.1

(22.

3)14

0.4

(22.

7)0.

004

857/

151

150.

9(2

2.6)

152.

6(2

5.8)

0.48

DB

P(m

mH

g)66

9/39

81.3

(10.

8)75

.4(1

0.3)

0.00

285

7/15

180

.9(1

1.0)

79.9

(11.

1)0.

40B

MI

(kg/

m2)

670/

3927

.8(4

.0)

27.0

(4.0

)0.

2985

8/15

127

.8(4

.0)

27.2

(4.1

)0.

09W

aist

-hip

ratio

669/

390.

9(0

.1)

0.9

(0.1

)0.

1285

2/15

00.

9(0

.1)

0.9

(0.1

)0.

35To

talc

hole

ster

ol(m

mol

/l)66

8/39

5.9

(1.0

)5.

6(1

.1)

0.08

855/

150

5.9

(1.0

)5.

9(1

.0)

0.56

HD

L(m

mol

/l)66

8/39

1.4

(0.4

)1.

5(0

.4)

0.48

855/

150

1.4

(0.4

)1.

5(0

.4)

0.09

Tri

glyc

erid

es(m

mol

/l)66

8/39

1.5

(0.7

)1.

4(0

.7)

0.38

855/

150

1.5

(0.7

)1.

5(0

.8)

0.35

Phys

ical

activ

ity(1

-6p)

666/

392.

0(1

.1)

2.5

(1.8

)0.

4784

9/14

92.

1(1

.2)

2.3

(1.5

)0.

32A

lcoh

olus

e(1

-3p)

663/

391.

8(0

.8)

1.9

(0.7

)0.

2384

8/14

81.

8(0

.8)

1.9

(0.8

)0.

22Sm

oker

665/

3833

.1%

29.0

%0.

6084

8/14

832

.9%

35.8

%0.

49Pr

esen

ceof

com

orbi

dity

670/

3920

.9%

41.0

%0.

003

858/

151

21.1

%31

.1%

0.00

7A

PO

Ege

noty

pe�23

/24

or33

/34/

44(N

)66

9/39

8/63

/27/

25/

41/4

4/10

0.00

185

0/15

08/

62/2

8/2

3/48

/38/

11<0

.001

�4

carr

ier

670/

3930

.4%

53.8

%0.

002

858/

151

31.1

%50

.3%

<0.0

01Se

lf-r

ated

heal

thm

easu

res

Self

-rat

edhe

alth

(1-5

p)66

7/39

2.6

(0.7

)2.

7(0

.7)

0.71

852/

150

2.6

(0.7

)2.

8(0

.8)

0.06

Self

-rat

edfit

ness

(1-5

p)66

6/37

2.7

(0.7

)2.

7(0

.8)

0.84

852/

148

2.7

(0.7

)2.

8(0

.8)

0.10

Hop

eles

snes

s(0

-8p)

624/

385.

2(1

.8)

4.7

(2.1

)0.

1678

2/12

75.

1(1

.8)

4.9

(1.8

)0.

44B

DI

(0-6

3p)

572/

299.

2(6

.2)

10.3

(7.3

)0.

4971

9/11

19.

2(6

.3)

10.0

(7.0

)0.

38


Subj

ecti

veM

emor

yQ

uest

ionn

aire

(1-4

p/qu

estio

n)To

tals

core

510/

2846

.1(8

.1)

50.1

(9.0

)0.

021

648/

9845

.7(8

.2)

48.1

(10.

0)0.

020

Forg

ettin

gph

one

num

bers

659/

382.

5(0

.8)

2.9

(0.7

)0.

003

842/

143

2.5

(0.8

)2.

75(0

.77)

0.00

1Fo

rget

ting

clot

hing

size

632/

322.

3(0

.9)

2.8

(1.1

)0.

023

808/

132

2.3

(1.0

)2.

5(1

.0)

0.19

Forg

ettin

gna

me

ofac

tors

638/

372.

8(0

.8)

3.2

(0.8

)0.

005

807/

142

2.8

(0.8

)2.

9(0

.8)

0.10

Forg

ettin

gw

hatt

osa

yin

mid

-sen

tenc

e66

2/38

1.8

(0.6

)2.

1(0

.7)

0.00

984

4/14

11.

8(0

.6)

1.9

(0.7

)0.

09C

hang

esin

vasc

ular

fact

ors

(lat

e-li

fe–

mid

life

)SB

P(m

mH

g)66

9/39

10.3

(23.

2)–1

.2(2

5.2)

0.00

285

7/15

19.

4(2

3.3)

7.6

(26.

5)0.

29D

BP

(mm

Hg)

669/

39–7

.0(1

2.1)

–13.

3(1

4.0)

0.01

785

7/15

1–7

.4(1

2.1)

–9.8

(12.

7)0.

031

Tota

lcho

lest

erol

(mm

ol/l)

668/

39–0

.8(1

.2)

–1.0

(1.4

)0.

2485

5/15

0–0

.8(1

.2)

–0.9

(1.2

)0.

035

BM

I(k

g/m

2)

670/

391.

6(2

.7)

–0.1

(2.3

)<0

.001

858/

151

1.6

(2.7

)0.

2(2

.9)

<0.0

01

All

show

nch

arac

teri

stic

sw

ere

asse

ssed

atth

efir

stla

te-l

ife

re-e

xam

inat

ion

(199

8),e

xcep

tfo

rC

hang

esin

vasc

ular

fact

ors,

whi

chsh

owdi

ffer

ence

sbe

twee

n19

98an

dm

idlif

e(2

1ye

ars

earl

ier)

.V

alue

sar

em

eans

(sta

ndar

dde

viat

ions

)un

less

othe

rwis

esp

ecifi

ed.F

orC

ogni

tion,

high

erre

sults

indi

cate

bette

rpe

rfor

man

ce,e

xcep

tfor

exec

utiv

efu

nctio

ning

whe

relo

wer

resu

ltsin

dica

tebe

tter

perf

orm

ance

.Phy

sica

lac

tivity

was

asse

ssed

as6

orde

red

cate

gori

es:

1=

daily

;2

=2-

3tim

esa

wee

k;3

=on

cea

wee

k;4

=2-

3tim

esa

mon

th;

5=

afe

wtim

esa

year

;an

d6

=no

tat

all.

Alc

ohol

use

was

asse

ssed

as3

orde

red

cate

gori

es:

1=

mon

thly

;2

=le

ssth

anm

onth

ly;

and

3=

not

atal

l.Se

lf-r

ated

heal

than

dfit

ness

wer

eas

sess

edas

5or

dere

dca

tego

ries

:1

=ve

rygo

od;

2=

good

;3

=sa

tisfa

ctor

y;4

=re

lativ

ely

poor

;an

d5

=ve

rypo

or.F

orho

pele

ssne

ss,h

ighe

rsc

ore

indi

cate

dle

ssho

pele

ssne

ss.F

orB

DI

(Bec

kD

epre

ssio

nIn

vent

ory)

,hig

her

scor

ein

dica

tem

ore

pron

ounc

edde

pres

sive

sym

ptom

s.In

the

Subj

ectiv

eM

emor

yQ

uest

ionn

aire

,eac

hqu

estio

nha

d4

orde

red

answ

erca

tego

ries

:1=

neve

r;2

=so

met

imes

;3=

ofte

n;an

d4

=al

mos

tall

the

time

(i.e

.,hi

gher

scor

ein

dica

tes

mor

epr

onou

nced

mem

ory

com

plai

nts)

.Onl

yqu

estio

nsw

ithsi

gnifi

cant

diff

eren

ces

betw

een

grou

psar

esh

own

here

.


Table 2Performance of DSI, included individual factors and factor groups in predicting dementia

Main study population (participants/survivors) Extended study populationAUC (95% CI) AUC (95% CI)

Basic model

Total DSI 0.79 (0.79–0.80) 0.75 (0.74–0.75)Cognition 0.73 (0.73–0.74) 0.69 (0.69–0.70)Executive functioning 0.68 (0.67–0.69) 0.62 (0.62–0.63)Episodic memory 0.64 (0.62–0.65) 0.61 (0.61–0.62)Prospective memory 0.62 (0.61–0.63) 0.63 (0.62–0.63)Psychomotor speed 0.62 (0.61–0.63) 0.67 (0.66–0.68)MMSE 0.54 (0.54–0.55)Verbal Expression 0.55 (0.55–0.56)Socio-demographic characteristics 0.67 (0.65–0.68) 0.66 (0.66–0.67)Age 0.67 (0.65–0.68) 0.66 (0.66–0.67)Vascular factors 0.65 (0.64–0.66) 0.53 (0.52–0.53)DBP 0.64 (0.63–0.65)SBP 0.63 (0.62–0.64)Presence of comorbidity 0.56 (0.55–0.57) 0.53 (0.52–0.53)Subjective Memory Questionnaire 0.64 (0.63–0.66) 0.58 (0.57–0.58)Total score 0.62 (0.61–0.64) 0.57 (0.56–0.58)Forgetting phone numbers 0.61 (0.60–0.62) 0.57 (0.56–0.57)Forgetting name of actors 0.60 (0.59–0.61)Forgetting clothing size 0.59 (0.57–0.60)Forgetting what to say in mid-sentence 0.58 (0.57–0.59)APOE genotype 0.59 (0.58–0.60) 0.60 (0.59–0.61)Genotype risk order 0.60 (0.59–0.61) 0.60 (0.60–0.61)�4 carrier 0.57 (0.55–0.58) 0.57 (0.57–0.58)

Basic model + changes in vascular factors from midlife

Total DSI 0.80 (0.79–0.81) 0.78 (0.77–0.79)Vascular changes 0.68 (0.66–0.69) 0.65 (0.64–0.66)BMI change 0.68 (0.67–0.69) 0.68 (0.67–0.69)SBP change 0.65 (0.63–0.66)DBP change 0.61 (0.59–0.62) 0.61 (0.59–0.62)Total cholesterol change 0.55 (0.54–0.57)

Values are AUC (95% CI) for the composite DSI, factor groups (Cognition, Vascular factors, Demographics, Subjective memory questionnaire,and APOE genotype), and individual factors within each group. In the basic model + changes in vascular factors from midlife, the total DSIvalue includes all factors and factor groups from the basic model plus the Vascular changes group. Only factors with significant differencesbetween control and dementia groups (as per Table 1 p-values) are shown here.

cases, and on factors showing significant differencesat various p-value thresholds. Results suggest that themodel is not improved after adding variables withp > 0.01.

Additional analyses were conducted to accountfor previously described J- or U-shaped associa-tions between BMI, BP, cholesterol, and dementia[1] (the current DSI version includes a monotonicallyincreasing fitness function). Dichotomous variableswere created for values higher or lower than chosencut-offs for BMI, BP, and total cholesterol, and thevariables were added to the models to investigate thesignificance of the distribution tails. Several cut-offswere tested, but the combined predictive performanceof these variables was low and did not affect theoverall performance of the model (results not shown).

Results were validated by comparison with a SVMclassification, trained with a linear kernel using the

same set of factors and cross-validation procedure.We used the MATLAB fitcsvm function with parame-ter values that empirically gave the best results (kernelscale 103 and box constraint 10−3 for both models).Population mean values were used for missing values,and factors were entered into the model as individ-ual standardized values. The SVM achieved an AUCof 0.77 (0.76–0.78) for the main study population,and 0.74 (0.73–0.74) for the extended population, aslightly lower performance compared with DSI.

DISCUSSION

The late-life DSI dementia index developed usinga supervised machine learning method performedwell in predicting dementia up to 10 years later inan older general population without MCI or demen-tia at baseline. Performance was in the upper range


Table 3Late-life DSI dementia index cut-offs (basic model) with accuracy, sensitivity, specificity, and the percentage of individuals classified as

developing dementia in the future

Main study population Extended study populationCut-off Accuracy Sensitivity Specificity Classified Accuracy Sensitivity Specificity Classified

dementia + (%) dementia + (%)

0.10 0.06 1.00 0.01 99 0.15 1.00 0.00 1000.20 0.15 1.00 0.10 91 0.19 1.00 0.05 960.30 0.33 0.93 0.30 72 0.34 0.95 0.23 800.40 0.56 0.85 0.54 48 0.51 0.83 0.45 590.45 0.66 0.82 0.65 38 0.60 0.78 0.57 480.50 0.74 0.73 0.74 29 0.67 0.69 0.67 380.55 0.81 0.61 0.83 20 0.74 0.59 0.76 290.60 0.87 0.45 0.89 13 0.80 0.49 0.85 200.70 0.93 0.24 0.97 5 0.85 0.26 0.95 80.80 0.94 0.09 0.99 1 0.85 0.07 0.99 20.90 0.95 0.03 1.00 0 0.85 0.00 1.00 0

Table 4Effects of p-value threshold filtering on the number of factors included in the model, and on the predictive performance (AUC) of the DSI

dementia index

Main study population Extended study populationp-value thresholds No. of factors included AUC (95% CI) No. of factors included AUC (95% CI)

in model in model

p < 0.000001 0 5 0.76 (0.75–0.76)p < 0.001 4 0.76 (0.75–0.78) 9 0.77 (0.76–0.77)p < 0.01 14 0.82 (0.81–0.83) 10 0.77 (0.76–0.77)p < 0.05 18 0.80 (0.79–0.81) 15 0.75 (0.75–0.76)p < 0.1 21 0.79 (0.79–0.80) 23 0.75 (0.74–0.75)p < 0.2 30 0.79 (0.78–0.80) 27 0.75 (0.74–0.75)no threshold 49 0.74 (0.73–0.76) 49 0.73 (0.72–0.73)

p-values calculated from Mann-Whitney U-tests comparing controls and subsequent dementia cases were used for the thresholds shown.Only factors showing significant differences between groups below a specific threshold are included in the model and factors not showingsignificant differences are filtered out of the model.

of reported performance for previous dementia riskscores [5], and close to the performance level of estab-lished risk scores for cardiovascular conditions [4,37, 38]. The late-life DSI dementia index and midlifeCAIDE Dementia Risk Score, both developed withinthe CAIDE study but with very different methods,had similar predictive power [8, 9].

As emphasized by a recent multidomain vascu-lar care trial to prevent dementia [39], preventiveinterventions may not be effective in unselectedolder populations. A risk-based selection could facil-itate targeting preventive interventions to individualswho are most likely to benefit. The midlife CAIDEDementia Risk Score has been used for this purposein another population-based multidomain lifestyletrial that showed significant beneficial interventioneffects on cognitive performance [2]. However, theselection required data pre-processing according topre-set cut-offs, and additional cognitive testing ref-erenced to population norms (separate from thedementia risk score). The late-life DSI dementia

index could facilitate faster and more detailed riskassessment, with easier to interpret individual riskprofiles, thus enabling risk-based selection of targetpopulations, and also potential tailoring of preven-tive interventions based on the most relevant riskfactors. Such advantages derive from the ability ofDSI to quickly handle large amounts of hetero-geneous data in raw form (i.e., as collected fromsubjects), and the provision of DSI data to humanreaders in an easily interpretable visual form. Whilemany available classifiers process data as a ‘blackbox’ requiring machine learning expertise to scruti-nize, DSF clearly discloses the factors contributingto the results, and supports clinical judgment byhighlighting what is most relevant. Such character-istics are particularly important for dementia riskassessment tools in the context of recent databasedevelopments such as large population-based onlineBrain Health Registries, multinational data discoveryand sharing platforms, or internet-based preventiontrials [11].


Fig. 2. ROC curves for the late-life DSI dementia index in the mainand extended study populations.

Factors included in the DSI index

A large number of heterogeneous factors weretested in the present study, and DSI performed wellin identifying the main types of late-life risk factorsrelated to subsequent dementia: objective and subjec-tive measures of cognition, age, vascular factors, andAPOE genotype, in overall agreement with previousstudies using other statistical methods [5]. Detailed,factor-specific comparisons with available dementiarisk scores are difficult because these have often pre-processed raw data according to different cut-offs,and/or combined variables in different ways, leadingto variability in individual factors and their weights.However, some general patterns can be observed.

Long-term (i.e., decades) dementia predictionmodels tend to differ from shorter-term (i.e.,<10years) prediction models, and they also tend to per-form poorly when applied outside the age groupsthey were designed for [5, 6]. The relatively longpre-clinical stage of dementia-related diseases (e.g.,Alzheimer’s disease or cerebrovascular disease) isa major challenge for dementia risk scores, partic-ularly at older ages [5, 6]. The links between riskfactors and dementia development can be bidirec-tional, i.e., a factor may increase dementia risk, butit may also be influenced by ongoing disease pro-cesses once the dementia-related disease starts [1].While the mechanisms are not yet fully clear, a pat-tern of more pronounced decline in, for example, BP,BMI, and total cholesterol from midlife to late-life

has been consistently described in people who sub-sequently develop dementia [1]. Whereas traditionalvascular risk factors (e.g., high BP, BMI, and/or totalcholesterol) are important for midlife dementia riskscores, their predictive value decreases in late-liferisk scores (some of which may even include low BPand/or low BMI as predictors) [1, 5]. AUCs for thevascular factors group in the DSI dementia index arein agreement with this pattern. Interestingly, groupAUCs for changes in vascular factors prior to base-line were slightly higher that group AUCs for vascularfactors at baseline in the DSI model. Declining BMIfrom midlife to late-life was the most important pre-dictor in the vascular changes group, while BMI inlate-life was not predictive of subsequent dementia.The predictive value of one-time late-life measure-ments versus midlife-to-latelife changes has so far notbeen investigated in late-life dementia risk scores.

However, overall performance of the DSI dementiaindex was not greatly affected by leaving out changesin vascular factors. The most important predictor wascognitive performance, which is perhaps not surpris-ing for late-life dementia risk scores [5]. Cognitiveperformance was also more predictive of subsequentdementia than age. As our study focused on individ-uals aged 65–79 years, it remains to be determinedwhether this finding applies to other age groups orpopulations. APOE genotype had the lowest AUCscompared to the other groups of factors included inthe DSI models. While in some previous dementiaprediction models APOE genotype appeared to besomewhat informative, other models have excludedit as not informative enough [5].

Strengths, limitations, and future directions

The main strengths of the present study are thepopulation-based design, long follow-up time, anddetailed late-life cognitive assessments at two timepoints, thus increasing diagnostic accuracy. Mortal-ity and non-participation were at least partly takeninto account by including both the main popula-tion (survivors/participants) and extended population(additional register dementia diagnoses for non-survivors/non-participants) in analyses. Results forboth populations were relatively similar, although inthe extended population AUCs tended to be some-what lower, and some factors were excluded fromthe models. Individuals who do not participate instudies or die during follow-up usually have poorerhealth, and are more likely to either develop demen-tia or die at younger ages, before dementia onset.


Although dementia diagnoses in Finnish national reg-isters were accurate (positive predictive values above90%), their combined sensitivity was around 70%[22], thus underestimating the actual number of cases.Also, individuals who died without recorded demen-tia diagnoses had to be excluded from analyses.

The comorbidity variable used in DSI modelswas based on Hospital Discharge Register diagnoses,thus including only cardio/cerebrovascular condi-tions severe enough to require hospitalization (dataon pharmacological treatment and conditions diag-nosed in outpatient clinics were not available). Also,brain MRI measurements were not included in thepresent study due to insufficient sample size. Aprevious late-life risk index including MRI measure-ments had somewhat better predictive performance(AUC 0.81) [40], but the shorter version withoutMRI had similar predictive performance to DSI (AUC0.77) [41].

The present study tested many heterogeneous fac-tors, and results from p-value thresholds filteringanalyses indicated that the DSI dementia index ben-efited from selection of factors. DSI was originallybuilt with the assumption that all included factorsare already established as likely classifiers, and theireffectiveness is ranked by relevance. If several fac-tors with unclear predictive value for dementia areincluded, the need for factor selection arises. A largeamount of poor classifiers with little relevance canoverpower the factors with higher relevance and skewthe final results. Also, if the training groups are toosmall, a non-significant difference between controlsand cases can lead to a higher relevance by chance.

The late-life DSI dementia prediction model wasdesigned for shorter-term dementia prediction (upto 10 years). External validation is needed to ver-ify its predictive performance. Long-term predictiveperformance will also need to be tested. In addition,analyses of changes in overall risk level over time areessential for determining whether the DSI dementiaindex can be used for longitudinal risk monitoringand assessing response to preventive interventions.

Conclusion

DSI performed well in identifying comprehensiveprofiles for predicting dementia development up to10 years later. The DSI dementia index could thusbe useful for identifying individuals who are most atrisk and may benefit from dementia prevention inter-ventions. The detailed and visually easy to interpretindividual risk profiles may also facilitate tailoring of

preventive interventions based on the most relevantrisk factors.

ACKNOWLEDGMENTS

This study was funded by the European Union7th Framework Program for research, technologi-cal development and demonstration VPH-DARE@IT(GrantAgreementNo:601055);MIND-ADAcademyof Finland 291803 and Swedish Research Council529-2014-7503 (EU Joint Programme - Neurodegen-erative Disease Research, JPND); strategic fundingfor UEF-BRAIN from University of Eastern Fin-land; Academy of Finland grants 287490 and 294061;Center for Innovative Medicine (CIMED), Sweden;Alzheimerfonden Sweden; AXA Research Fund.

The funding sources had no involvement in studydesign; in the collection, analysis and interpretationof data; in the writing of the report; and in the decisionto submit the article for publication.

Authors’ disclosures available online (http://j-alz.com/manuscript-disclosures/16-0560r1).

SUPPLEMENTARY MATERIAL

The supplementary material is available in theelectronic version of this article: http://dx.doi.org/10.3233/JAD-160560.

REFERENCES

[1] Solomon A, Mangialasche F, Richard E, Andrieu S, BennettDA, Breteler M, Fratiglioni L, Hooshmand B, KhachaturianAS, Schneider LS, Skoog I, Kivipelto M (2014) Advancesin the prevention of Alzheimer’s disease and dementia.J Intern Med 275, 229-250.

[2] Ngandu T, Lehtisalo J, Solomon A, Levalahti E, Ahtiluoto S,Antikainen R, Backman L, Hanninen T, Jula A, LaatikainenT, Lindstrom J, Mangialasche F, Paajanen T, Pajala S,Peltonen M, Rauramaa R, Stigsdotter-Neely A, StrandbergT, Tuomilehto J, Soininen H, Kivipelto M (2015) A 2 yearmultidomain intervention of diet, exercise, cognitive train-ing, and vascular risk monitoring versus control to preventcognitive decline in at-risk elderly people (FINGER): Arandomized controlled trial. Lancet 385, 2255-2263.

[3] European Dementia Prevention Initiative, http://www.edpi.org, Accessed on September 19, 2016

[4] Lloyd-Jones DM (2010) Cardiovascular risk prediction:Basic concepts, current status, and future directions.Circulation 121, 1768-1777.

[5] Tang EY, Harrison SL, Errington L, Gordon MF, Visser PJ,Novak G, Dufouil C, Brayne C, Robinson L, Launer LJ,Stephan BC (2015) Current developments in dementia riskprediction modelling: An updated systematic review. PLoSOne 10, e0136181.

[6] Solomon A, Soininen H (2015) Dementia: Risk predictionmodels in dementia prevention. Nat Rev Neurol 11, 375-377.


[7] Anstey KJ, Cherbuin N, Herath PM (2013) Developmentof a new method for assessing global risk of Alzheimer’sdisease for use in population health approaches to preven-tion. Prev Sci 14, 411-421.

[8] Kivipelto M, Ngandu T, Laatikainen T, Winblad B, SoininenH, Tuomilehto J (2006) Risk score for the prediction ofdementia risk in 20 years among middle aged people: Alongitudinal, population-based study. Lancet Neurol 5, 735-741.

[9] Exalto LG, Quesenberry CP, Barnes D, Kivipelto M,Biessels GJ, Whitmer RA (2013) Midlife risk score forthe prediction of dementia four decades later. AlzheimersDement 10, 562-570.

[10] Sindi S, Calov E, Fokkens J, Ngandu T, Soininen H,Tuomilehto J, Kivipelto M (2015) The CAIDE DementiaRisk Score App: The development of an evidence-based mobile application to predict the risk of dementia.Alzheimers Dement (Amst) 1, 328-333.

[11] Healthy Ageing Through Internet Counselling in the Elderlyclinical trial, http://www.hatice.eu, Accessed on September19, 2016.

[12] Mattila J, Koikkalainen J, Virkki A, van Gils M, LotjonenL, and Alzheimer’s Disease Neuroimaging Initiative (2012)Design and application of a generic clinical decision supportsystem for multiscale data. IEEE Trans Biomed Eng 59,234-240.

[13] Mattila J, Koikkalainen J, Virkki A, Simonsen A,van Gils M, Waldemar G, Soininen H, Lotjonen J;Alzheimer’s Disease Neuroimaging Initiative (2011) A Dis-ease State Fingerprint for evaluation of Alzheimer’s disease.J Alzheimers Dis 27, 163-176.

[14] Mattila J, Soininen H, Koikkalainen J, Rueckert D, Wolz R,Waldemar G, Lotjonen J (2012) Optimizing the diagnosisof early Alzheimer’s disease in mild cognitive impairmentsubjects. J Alzheimers Dis 32, 969-979.

[15] Simonsen AH, Mattila J, Hejl AM, Frederiksen KS,Herukka SK, Hallikainen M, van Gils M, Lotjonen J, Soini-nen H, Waldemar G (2012) Application of the PredictADsoftware tool to predict progression in patients with mildcognitive impairment. Dement Geriatr Cogn Disord 34,344-350.

[16] Liu Y, Mattila J, Ruiz MA, Paajanen T, Koikkalainen J,van Gils M, Herukka SK, Waldemar G, Lotjonen J, Soini-nen H, Alzheimer’s Disease Neuroimaging Initiative (2013)Predicting AD conversion: Comparison between prodromalAD guidelines and computer assisted PredictAD tool. PLoSOne 8, e55246.

[17] Hall A, Mattila J, Koikkalainen J, Lotjonen J, Wolz R,Scheltens P, Frisoni G, Tsolaki M, Nobili F, Freund-LeviY, Minthon L, Frolich L, Hampel H, Visser PJ, Soininen H(2015) Predicting progression from cognitive impairmentto Alzheimer’s disease with the Disease State Index. CurrAlzheimer Res 12, 69-79.

[18] Hall A, Munoz-Ruiz M, Mattila J, Koikkalainen J, Tso-laki M, Mecocci P, Kloszewska I, Vellas B, LovestoneS, Visser PJ, Lotjonen J, Soininen H, Alzheimer Dis-ease Neuroimaging Initiative, AddNeuroMed consortium,DESCRIPA, Kuopio L-MCI (2015) Generalizability of thedisease state index prediction model for identifying patientsprogressing from mild cognitive impairment to Alzheimer’sdisease. J Alzheimers Dis 44, 79-92.

[19] Munoz-Ruiz MA, Hartikainen P, Hall A, Mattila J,Koikkalainen J, Herukka SK, Julkunen V, Vanninen R,Liu Y, Lotjonen J, Soininen H (2013) Disease State Fin-gerprint in frontotemporal degeneration with reference

to Alzheimer’s disease and mild cognitive impairment.J Alzheimers Dis 35, 727-739.

[20] Simonsen AH, Mattila J, Hejl AM, Garde E, van Gils M,Thomsen C, Lotjonen J, Soininen H, Waldemar G (2013)Application of the PredictAD decision support tool to a Dan-ish cohort of patients with Alzheimer’s disease and otherdementias. Dement Geriatr Cogn Disord 37, 207-213.

[21] Kivipelto M, Helkala EL, Laakso MP, Hanninen T,Hallikainen M, Alhainen K, Soininen H, TuomilehtoJ, Nissinen A (2001) Midlife vascular risk factors andAlzheimer’s disease in later life: Longitudinal, populationbased study. BMJ 322, 1447-1451.

[22] Solomon A, Ngandu T, Soininen H, Hallikainen M,Kivipelto M, Laatikainen T (2014) Validity of dementia andAlzheimer’s disease diagnoses in Finnish national registers.Alzheimers Dement 10, 303-309.

[23] CAIDE-Cardiovascular Risk Factors, Aging and Dementia,http://www.uef.fi/caide, Accessed on September 19, 2016.

[24] Folstein MF, Folstein SE, McHugh PR (1975) Mini-mentalstate. A practical method for grading the cognitive state ofpatients for the clinician. J Psychiatr Res 12, 189-198.

[25] Morris JC, Heyman A, Mohs RC, Hughes JP, van BelleG, Fillenbaum G, Mellits ED, Clark C (1989) The Con-sortium to Establish a Registry for Alzheimer’s disease(CERAD). Part I. Clinical and neuropsychological assess-ment of Alzheimer’s disease. Neurology 39, 1159-1165.

[26] American Psychiatric Association (1994) Diagnostic andStatistical Manual of Mental Disorders, 4th edn. AmericanPsychiatric Association, Washington, DC.

[27] McKhann G, Drachman D, Folstein M, Katzman R, Price D,Stadlan EM (1984) Clinical diagnosis of Alzheimer’s dis-ease: Report of the NINCDS-ADRDA Work Group underthe auspices of Department of Health and Human Ser-vices Task Force on Alzheimer’s Disease. Neurology 34,939-944.

[28] Petersen RC, Smith GE, Ivnik RJ, Tangalos EG, Schaid DJ,Thibodeau SN, Kokmen E, Waring SC, Kurland LT (1995)Apolipoprotein E status as a predictor of the developmentof Alzheimer’s disease in memory-impaired individuals.JAMA 273, 1274-1278.

[29] Kuulasmaa K, Tunstall-Pedoe H, Dobson A, Fortmann S,Sans S, Tolonen H, Evans A, Ferrario M, Tuomilehto J(2000) Estimation of contribution of changes in classic riskfactors to trends in coronary-event rates across the WHOMONICA Project populations. Lancet 355, 675-687.

[30] Ngandu T, Helkala EL, Soininen H, Winblad B, Tuomile-hto J, Nissinen A, Kivipelto M (2007) Alcohol drinkingand cognitive functions: Findings from the CardiovascularRisk Factors Aging and Dementia (CAIDE) Study. DementGeriatr Cogn Disord 23, 140-149.

[31] Everson SA, Goldberg DE, Kaplan GA, Cohen RD, PukkalaE, Tuomilehto J, Salonen JT (1996) Hopelessness and risk ofmortality and incidence of myocardial infarction and cancer.Psychosom Med 58, 113-121.

[32] Beck AT, Ward CH, Mendelson M, Mock J, Erbaugh J(1961) An inventory for measuring depression. Arch GenPsychiatry 4, 561-571.

[33] Bennett-Levy J, Powell GE (1980) The SubjectiveMemory Questionnaire (SMQ). An investigation into theself-reporting of ‘real-life’ memory skills. Br J Soc ClinPsychol 19, 177-188.

[34] Tsukamoto K, Watanabe T, Matsushima T, Kinoshita M,Kato H, Hashimoto Y, Kurokawa K, Teramoto T (1993)Determination by PCR-RFLP of ApoE genotype in aJapanese population. J Lab Clin Med 121, 598-602.


[35] Corder EH, Saunders AM, Strittmatter WJ, Schmechel DE,Gaskell PC, Small GW, Roses AD, Haines JL, Pericak-Vance MA (1993) Gene dose of apolipoprotein E type 4allele and the risk of Alzheimer’s disease in late onset fam-ilies. Science 261, 921-923.

[36] Corder EH, Saunders AM, Risch NJ, Strittmatter WJ,Schmechel DE, Gaskell PC Jr, Rimmler JB, Locke PA, Con-neally PM, Schmader KE, Small GW, Roses AD, Haines JL,Pericak-Vance MA (1994) Protective effect of apolipopro-tein E type 2 allele for late onset Alzheimer disease. NatGenet 7, 180-184.

[37] Wilson PW, D’Agostino RB, Levy D, Belanger AM,Silbershatz H, Kannel WB (1998) Prediction of coronaryheart disease using risk factor categories. Circulation 97,1837-1847.

[38] Conroy RM, Pyorala K, Fitzgerald AP, Sans S, Menotti A,De Backer G, De Bacquer D, Ducimetiere P, Jousilahti P,Keil U, Njølstad I, Oganov RG, Thomsen T, Tunstall-PedoeH, Tverdal A, Wedel H, Whincup P, Wilhelmsen L, Graham

IM; SCORE project group (2003) Estimation of ten-yearrisk of fatal cardiovascular disease in Europe: The SCOREproject. Eur Heart J 24, 987-1003.

[39] Moll van Charante EP, Richard E, Eurelings LS, van DalenJW, Ligthart SA, van Bussel EF, Hoevenaar-Blom MP,Vermeulen M, van Gool WA (2016) Effectiveness of a6-year multidomain vascular care intervention to preventdementia (preDIVA): A cluster-randomised controlled trial.Lancet 388, 797-805.

[40] Barnes DE, Covinsky KE, Whitmer RA, Kuller LH, LopezOL, Yaffe K (2009) Predicting risk of dementia in olderadults: The late-life dementia risk index. Neurology 73, 173-179.

[41] Barnes DE, Covinsky KE, Whitmer RA, Kuller LH, LopezOL, Yaffe K (2010) Commentary on Developing a nationalstrategy to prevent dementia: Leon Thal Symposium 2009.Dementia risk indices: A framework for identifying indi-viduals with a high dementia risk. Alzheimers Dement 6,138-141.

II

Prediction models for dementia and neuropathology in the oldest old:the Vantaa 85+ cohort study

Hall A, Pekkala T, Polvikoski T, van Gils M, Kivipelto M, Lotjonen J, Mattila J,Kero M, Myllykangas L, Makela M, Oinas M, Paetau A, Soininen H, Tanskanen M

and Solomon A

Alzheimer’s Research & Therapy 11, 2019

Reprinted under the Creative Commons Attribution License 4.0

RESEARCH Open Access

Prediction models for dementia andneuropathology in the oldest old: theVantaa 85+ cohort studyAnette Hall1, Timo Pekkala1, Tuomo Polvikoski2, Mark van Gils3, Miia Kivipelto1,4,5,6, Jyrki Lötjönen7, Jussi Mattila7,Mia Kero8, Liisa Myllykangas8, Mira Mäkelä8, Minna Oinas8,9, Anders Paetau8, Hilkka Soininen1,Maarit Tanskanen8 and Alina Solomon1,4*

Abstract

Background: We developed multifactorial models for predicting incident dementia and brain pathology in theoldest old using the Vantaa 85+ cohort.

Methods: We included participants without dementia at baseline and at least 2 years of follow-up (N = 245) fordementia prediction or with autopsy data (N = 163) for pathology. A supervised machine learning method wasused for model development, considering sociodemographic, cognitive, clinical, vascular, and lifestyle factors, aswell as APOE genotype. Neuropathological assessments included β-amyloid, neurofibrillary tangles and neuriticplaques, cerebral amyloid angiopathy (CAA), macro- and microscopic infarcts, α-synuclein pathology, hippocampalsclerosis, and TDP-43.

Results: Prediction model performance was evaluated using AUC for 10 × 10-fold cross-validation. Overall AUCswere 0.73 for dementia, 0.64–0.68 for Alzheimer’s disease (AD)- or amyloid-related pathologies, 0.72 for macroinfarcts,and 0.61 for microinfarcts. Predictors for dementia were different from those in previous reports of youngerpopulations; for example, age, sex, and vascular and lifestyle factors were not predictive. Predictors for dementia versuspathology were also different, because cognition and education predicted dementia but not AD- or amyloid-relatedpathologies. APOE genotype was most consistently present across all models. APOE alleles had a different impact: ε4did not predict dementia, but it did predict all AD- or amyloid-related pathologies; ε2 predicted dementia, but it wasprotective against amyloid and neuropathological AD; and ε3ε3 was protective against dementia, neurofibrillarytangles, and CAA. Very few other factors were predictive of pathology.

Conclusions: Differences between predictors for dementia in younger old versus oldest old populations, as well as fordementia versus pathology, should be considered more carefully in future studies.

Keywords: Dementia, Neuropathology, Oldest old, Prediction, Supervised machine learning

BackgroundThe oldest old constitute the largest and fastest growingpopulation with dementia [1], but they are less often thefocus of dementia prevention studies. Cohort studieswith participants aged 85+ years [2–7] have investigatedindividual risk factors in association with dementia, but

the predictive value of more complex multifactorial riskprofiles in the oldest old is still unclear. Several dementiarisk scores have been developed in younger populations,but they tend to perform poorly for predicting dementiain the oldest old age groups [8, 9]. The association ofvascular and lifestyle-related factors with dementia risk,for example, has been shown to vary with age [10], andrisk profiles predictive of subsequent dementia can differbetween midlife and older age [9].While most multifactorial prediction models or risk

scores have focused on dementia, less is known about

* Correspondence: [email protected] of Clinical Medicine, Neurology, University of Eastern Finland, P.O.Box 1627, 70211 Kuopio, Finland4Division of Clinical Geriatrics, NVS, Karolinska Institutet, Stockholm, SwedenFull list of author information is available at the end of the article

© The Author(s). 2019 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, andreproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link tothe Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Hall et al. Alzheimer's Research & Therapy (2019) 11:11 https://doi.org/10.1186/s13195-018-0450-3

longitudinal prediction of neuropathology in peoplewithout dementia. In the oldest old, multiple dementia-related pathologies are common [11], but the associationwith a dementia diagnosis may be less straightforwardthan in younger age groups [10]. In this context, it be-comes particularly important to investigate potential dif-ferences between predictors for dementia and forspecific types of neuropathologies.The main aims of the present study based on the Van-

taa 85+ cohort are to develop multifactorial models for(1) predicting incident dementia in the oldest old, con-sidering sociodemographic, cognitive, clinical, lifestyle,and apolipoprotein E (APOE) genotype data; and (2) pre-dicting dementia-related neuropathologies at death inthe oldest old, including Alzheimer’s disease (AD)-re-lated pathology (amyloid plaques and neurofibrillary tan-gles), cerebral amyloid angiopathy (CAA), cerebralmacro- and microinfarcts, and Lewy body pathology(α-synuclein).

MethodsStudy populationThe Vantaa 85+ study has been described in detail previ-ously [4, 12]. In brief, the study focused on residents inthe City of Vantaa in southern Finland who were at least85 years old in 1991. Of the 601 people invited to par-ticipate, 11 refused, 1 could not be reached, and 1 died,leaving 588 (98%) participants who gave informed con-sent to participate in the study. Additionally, 35 peopledied before the baseline clinical examination, which wasdone for 553 participants. At baseline, 214 participantswere diagnosed with dementia, and 339 did not have de-mentia. Clinical reexaminations were conducted in 1994,

1996, 1999, and 2001. At the time of death, additionally101 participants had been diagnosed with dementia.Postmortem examination was conducted for 288 partici-pants who attended the baseline clinical examinationand 16 who had died before baseline. The Vantaa 85+study was approved by the ethics committee of theHealth Centre of the City of Vantaa. Written consent forthe autopsies was given by the nearest relatives of thedeceased.To reduce the effects of mortality, of the 339 without

dementia at baseline, 94 participants who died withinthe first 2 years of follow-up were excluded from the de-mentia prediction model. This eliminated significanttime-to-death differences between individuals who diedwith and without dementia. Of the 245 remaining partic-ipants without baseline dementia and who were includedin the model development (Fig. 1), 97 subsequently de-veloped dementia.The study population used for neuropathological pre-

diction model development included 163 participantswho attended the baseline examination, did not have de-mentia at baseline, and had available autopsy data. Par-ticipants with baseline dementia were excluded to enablecomparison with the dementia prediction model.

Assessment of factors included in prediction modelsFactors included in prediction models were assessed atthe baseline clinical evaluation, when participants wereexamined by a physician and interviewed on their health,health-related behavior, and medication by a trainednurse. Medical history was additionally verified usingprimary health care records. Sociodemographic factorsincluded age, sex, years of formal education, and social

Fig. 1 Study design flowchart

Hall et al. Alzheimer's Research & Therapy (2019) 11:11 Page 2 of 12

class [13]. Cognition was assessed with the Mini MentalState Examination (MMSE) [14] and the Short PortableMental Status Questionnaire (SPMSQ; recorded numberof errors) [15]. Participants self-reported subjectivememory decline on a scale of no/a little/yes. Functionalabilities were evaluated using activities of daily living(ADL) and the Instrumental Activities of Daily LivingScale (IADL) [16, 17]. Competence in daily activities wasalso assessed in the interview with a single question withsix answer options ranging from “independent” (1 point)to “needs help in all activities” (6 points). The ZungSelf-Rating Depression Scale was administered to assessdepressive symptoms [18]. Comorbidities included inprediction models were diabetes, cardiovascular condi-tions (angina pectoris, heart infarction, atrial fibrillation,heart failure, arteriosclerosis obliterans, or hypertension)and cerebrovascular conditions (stroke or transient is-chemic attack). Other vascular and lifestyle-related fac-tors were systolic and diastolic blood pressure, bodymass index (BMI), alcohol use, and smoking.Total cholesterol as well as high-density lipoprotein

(HDL) and low-density lipoprotein (LDL) cholesterolwere quantified from baseline blood samples using en-zymatic methods [4]. APOE genotyping was done with acombination of DNA minisequencing [19] and DNAamplification through PCR followed by restriction en-zyme digestion with HhaI [20].

Dementia diagnosisDementia was diagnosed according to the revised cri-teria of the Diagnostic and Statistical Manual of MentalDisorders, Third Edition [21]. AD and vascular dementiawere diagnosed using the National Institute of Neuro-logical and Communicative Disorders–Alzheimer’s Dis-ease and Related Disorders Association [22] andNational Institute of Neurological Disorders and Stroke–Association Internationale pour la Recherche en l’En-seignement en Neurosciences [23] criteria. Diagnosiswas based on a broad range of information, including in-terviews, health examinations, cognitive and functionalassessments, and health and social work records (e.g., in-formation on home services or other social care servicesprovided to participants based on diminished functionalor cognitive capacity). Diagnoses were made by consen-sus of two neurologists.Incident dementia cases were identified from medical

and social work records, as well as from the informa-tion collected at the study follow-up visits using exami-nations and interviews with participants and theirrelatives or caregivers [24]. Although clinicians werenot blinded to cognitive/functional assessments duringthese visits (e.g., MMSE, SPMSQ, ADL, IADL), diagno-ses relied primarily on overall clinical judgment basedon all available information.

Neuropathological assessmentNeuropathological assessments have been described indetail previously [12, 24–28]. In brief, brains obtained atautopsy were fixed in phosphate-buffered 4% formalde-hyde for at least 2 weeks and examined independently ofclinical data. For AD-related pathology, the Consortiumto Establish a Registry for Alzheimer’s disease (CERAD)protocol was followed [29]. Methenamine silver stainingwas used for β-amyloid [30], and the modifiedBielschowsky method was used for neurofibrillary tan-gles and neuritic plaques [31]. As described previously,β-amyloid load was determined as the average fractionof cortical area covered by methenamine silver-stainedplaques in four neocortical samples [12]. The averagenumber of neurofibrillary tangles was also determined inthe four samples [25]. The CERAD scores and Braakstages were defined as originally described [29, 32]. CAAwas analyzed in six brain regions (frontal, parietal, tem-poral and occipital lobes, hippocampus, and cerebellum)based on Congo red staining and confirmed using IHCagainst β-amyloid peptide [26]. Macroscopic infarcts(cavitary lesions or solid cerebral infarcts visible to thenaked eye) were identified from sliced cerebral hemi-spheres, brainstem, and cerebellum. All lesions were his-tologically ascertained to be infarcts. Corticalmicroinfarcts were analyzed in the H&E-stained tissuesections in the same six brain regions as CAA [26]. Theywere focal lesions smaller than 2 mm invisible to thenaked eye with neuronal loss, glial cell and macrophagereaction, and/or cystic tissue necrosis. Sections of sub-stantia nigra stained with the H&E method and sectionsof substantia nigra and hippocampus stained with anti-bodies against α-synuclein were used to screen for Lewybody-related pathology [27].Hippocampal sclerosis (HS) and TDP-43 (transactive

response binding protein 43) immunopositivity in thegranular cell layer of the hippocampus were assessed aspreviously described [28]. In summary, HS and hemi-spheric symmetry/asymmetry were determined on H&Estaining by estimating the severity of neuronal loss. Thedensity of neurons was assessed semiquantitatively bythree observers. For TDP-43 immunostaining, right-sided hippocampus tissue blocks were cut into4-μm-thick sections and stained with the Lab Visionimmunostainer (Thermo Fisher Scientific, Waltham,MA, USA), and polymer-based kits were used fordetection.

Disease State IndexThe Disease State Index (DSI) is a supervised machinelearning method designed for predicting disease out-comes and differential diagnostics as a clinicaldecision-making tool. A detailed description has beenpublished previously [33]. Compared with traditional


methods for developing dementia risk scores, DSI is ableto deal with larger amounts of heterogeneous data, tohandle missing data, and to use unprocessed data with-out prespecified cutoffs for predictors. Conceptuallyrelated factors are structured into groups, such as com-bining all cognitive tests. This is useful for filtering noiseand preventing strongly correlated factors from beingmultiplied. DSI thus provides detailed information aboutpredictive performance on multiple levels simultan-eously: the independent performance of each factor, thecombined performance of a group of similar factors, andthe overall performance of the entire model. DSI hasaccuracy comparable to that of methods such as logisticregression, support vector machines, and Bayes inference[33] and has previously been used for predicting demen-tia [34], progression of mild cognitive impairment [33,35, 36], and differential diagnosis of neurodegenerativediseases [37].DSI builds a model from the distributions of data

using a population with known outcomes. For testedindividuals, DSI gives index values ranging from 0 to1, describing similarity of the data to the distributionsin the model. A value close to 0 indicates that the dataare similar to controls (no subsequent dementia orpathology), while a value close to 1 shows similarity tocases (subsequent dementia or pathology). The dataused for the predictions can be dichotomous, continu-ous, or categorical.First, a fitness function is calculated for each factor.

Fitness function f(x) is the share of false-negative er-rors divided by the sum of false-negative andfalse-positive errors, using measurement value x as athreshold for classification. It goes through the distri-bution using each point as a classification threshold toevaluate the shares of false-negatives and false-posi-tives, assigning 0 to values unique to controls and 1 tovalues unique to cases.To complement fitness, a relevance value is calculated

for each measure. The relevance value ranges from 0 to1 and indicates the ability, based on the data, to differen-tiate between cases and controls in general. Relevance isdefined as the sum of sensitivity and specificity minus 1,also known as the Youden index. Two data distributionsthat are completely overlapping will receive a relevanceof 0, while two distributions with no overlap will get arelevance of 1.Conceptually related factors are structured into groups

to combine the effect of possibly correlating factors to asingle predictor. Individual factors are combined into agroup DSI value through a weighted average based ontheir relevance values. This process is then repeated re-cursively for all groups to obtain a total DSI value. Anymissing values are ignored as part of the model, and thetotal score is calculated only from the available values.

Data analysisWe built DSI models for predicting dementia and thedifferent neuropathologies. AUCs with 95% CIs for a10 × 10-fold cross-validation were calculated to evaluatemodel performance. The dataset was divided into tenrandom subgroups, where nine were combined to formthe training group and one acted as the test group. Thisprocess was completed for each subgroup, and thecross-validation itself was repeated ten times. Thus, weshow mean AUCs and 95% CIs resulting from the 10 ×10 cross-validation process. Factor selection was con-ducted before the model building; that is, only factorsthat were significantly different (p < 0.05) between thegroups with and without the outcome of interest wereincluded in the final models. The initial list of factorgroups and individual factors included sociodemo-graphics (age, sex, education, social class), cognition(MMSE total score, MMSE orientation, MMSE word list- sum of registration and recall tasks, MMSE calculation,MMSE other tasks, and SPMSQ), functioning (sum ofADL and IADL, competence in daily activities question,and subjective memory decline), APOE genotype (binaryvariables: ε2 carrier versus noncarrier, ε4 carrier versusnoncarrier, genotype ε3ε3 versus others; and a categor-ical variable: all genotypes [ε2ε2, ε2ε3, ε2ε4, ε3ε3, ε3ε4,or ε4ε4]), comorbidities (cardiovascular, cerebrovascular,and diabetes), cholesterol (total, LDL, and HDL), bloodpressure (systolic and diastolic), lifestyle (BMI, alcoholuse, and smoking), and depressive symptoms (ZungSelf-Rating Depression Scale).The following neuropathological outcomes were di-

chotomized as present versus absent: β-amyloid load(average fraction of cortical area covered by methena-mine silver-stained plaques > 0), tangle count (averagenumber of neurofibrillary tangles > 0), CAA (averagepercentage of blood vessels with CAA > 0), cerebralmacroinfarcts (total number > 0), microinfarcts (num-ber > 0), α-synuclein (brainstem, limbic, or diffuse neo-cortical pathology present versus absent), HS (severemarked/total loss versus no/minor loss of pyramidalneurons in the CA1 and subiculum), and TDP-43(immunopositivity in the granular cell layer present ver-sus absent). Neuropathological AD was defined on thebasis of National Institute on Aging–Alzheimer’s Associ-ation criteria [38] using the combination of Braak andCERAD scores, and it was dichotomized as present(intermediate or high likelihood of AD) versus absent(low likelihood of AD).To investigate relationships between neuropathological

variables and how they predict dementia, we con-ducted principal component analysis (PCA) in 159 par-ticipants without dementia at baseline and who hadcomplete neuropathological data. The pca function inMATLAB R2015b (MathWorks, Natick, MA, USA) was


used. The pathology variables were centered but notweighted by variance. The principal component (PC)scores of each participant were used as predictors for de-mentia, and predictive performance for each PC wasassessed using AUC values.

ResultsPredicting dementiaCharacteristics for the dementia prediction populationare shown in Table 1. The population consisted of 245participants without dementia at baseline who werealive for at least 2 more years. Mean follow-up was 5.6years, and 97 (40%) of the participants developed de-mentia before death. Education, performance in MMSE,SPMSQ, and competence in daily activities were signifi-cantly lower in people with subsequent dementia. Dif-ferences were also detected for APOE genotype. Otherbaseline characteristics were not significantly differentbetween groups (Table 1) and were excluded from theprediction model.Cross-validation results for the DSI model for pre-

dicting dementia development are shown in Table 2.AUC was 0.73 for the entire model. According to AUCvalues for the groups of predictors, cognition includingSPMSQ and both MMSE total score and its four sub-categories (orientation, calculation, word list, and othertasks), were the most important predictors of dementia,followed by functioning (competence in daily activities),sociodemographics (education), and APOE status.APOEε2 carrier status predicted dementia developmentbefore death, while the ε3ε3 genotype was protectiveagainst dementia development, although AUCs wererelatively low. The impact of APOEε2 and other majorpredictors was similar in further analyses consideringclinical diagnosis of AD and vascular dementia separ-ately (results not shown).

Predicting pathologyFor predicting pathology, we included the 163 partici-pants with no dementia at baseline and available autopsydata. This population had a mean age of 88.7 years, afollow-up time of 4.1 years, and 4.3 years of education.Thirty-one (19%) of these participants were male, 33(21%) of them were APOEε4 carriers, and 26 (17%) wereε2 carriers. Fifty-nine (36%) had dementia at death.Cross-validation results (AUCs) of the DSI pathology

prediction models are shown in Table 3. Sensitivities andspecificities are shown in Additional file 1: Table S1. Thetotal AUCs for AD- or amyloid-related pathologies were0.66 for amyloid load, 0.64 for tangle count, 0.68 forneuropathological AD, and 0.66 for CAA. APOE geno-type had the highest AUCs for all these pathologies, butthere were differences in the impact of different alleles.APOEε4 carrier status was predictive for all four

pathology outcomes, while APOEε2 carrier status wasprotective against β-amyloid load and neuropathologicalAD. The ε3ε3 genotype was protective against tanglecount and CAA, but it was not related to amyloid loador neuropathological AD.Very few other factors had predictive value (Table 3):

poorer competence in daily activities for β-amyloid load,higher total and LDL cholesterol, and subjective memorydecline for tangle count; lower social class and subjectivememory decline for neuropathological AD; and absenceof cardiovascular comorbidity and male sex for CAA.The model for cerebral macroinfarcts had the best

predictive performance with total AUC of 0.72 (Table 3).The predictors in descending order of AUC values werehistory of cerebrovascular conditions, poorer MMSEscore (total and word list learning and recall tasks),higher BMI, and poorer competence in daily activities.We also modeled the two most common subtypes ofcerebral macroinfarcts: cortical and white matter. Corticalmacroinfarcts (AUC of 0.71) were predicted by cerebro-vascular comorbidity and APOE genotype. APOEε4 car-riers were more likely to develop cortical macroinfarcts,while genotype ε3ε3 was protective. White matter macro-infarcts (AUC of 0.76) were predicted by cholesterol, cere-brovascular comorbidity, and APOEε2 carrier status. TheAUC for the cerebral microinfarcts model was 0.61, witheducation as the only predictor.HS (AUC of 0.78) was predicted by cognition (MMSE

total score, word list learning and recall, and othertasks). Current smokers were also more likely to haveHS (Table 3). TDP-43 was only predicted by less pro-nounced depressive symptoms (AUC of 0.69). Therewere no significant predictors found for α-synuclein.Overall, APOE genotype was the predictor that emerged

most consistently across all models. The impact of APOEon dementia versus pathology is summarized in Fig. 2.Associations between pathology and dementia at death

in participants without dementia at baseline are shown inAdditional file 1: Table S2, and results of the PCA analysisare provided in Additional file 1: Table S3. The first threecomponents of PCA explained 56% of the variance in aut-opsy findings and reflected three mostly independentpathological processes: AD-/amyloid-related pathology,including β-amyloid, neurofibrillary tangles, and CAA(PC1); vascular-type pathology, including primarilymacroscopic infarcts (PC2); and Lewy body-type path-ology, including α-synuclein (PC3). PC1 was most predict-ive of dementia (AUC of 0.71), followed by PC2(AUC of 0.60). The other PCs did not predict dementia.

DiscussionPredicting dementia in the oldest oldThe DSI model performance for predicting dementia on-set before death, on average 6 years later, in people aged


Table 1 Baseline characteristics of the dementia prediction population

Characteristics No dementia at death (n = 148) Dementia at death (n = 97) p Value

Follow-up time, years 5.4 (2.7) 5.8 (2.6) 0.3

Sociodemographics

Age at baseline, years 88.4 (2.6) 88.3 (2.6) 0.7

Men, n (%) 33 (22%) 19 (20%) 0.6

Education, years 4.6 (3.3) 3.7 (2.0) 0.01

Social class 6.0 (1.5) 6.2 (1.3) 0.2

Cognition

MMSE Total 25.3 (3.3) 22.2 (4.5) < 0.001

MMSE Calculation 3.4 (1.6) 2.9 (1.6) 0.02

MMSE Orientation 9.5 (0.8) 8.7 (1.6) < 0.001

MMSE Other tasks 7.4 (1.2) 6.7 (1.4) < 0.001

MMSE Wordlist 5.0 (1.1) 4.2 (1.3) < 0.001

SPMSQ 0.8 (1.4) 1.8 (1.9) < 0.001

Functioning

Competence in daily activities 2.6 (1.3) 3.2 (1.4) 0.001

ADL sum (ADL + IADL) 29.7 (10.2) 31.6 (10.1) 0.2

Subjective memory decline 1.7 (0.6) 1.9 (0.7) 0.05

APOE genotype

ε2ε3 17 (12%) 18 (19%) 0.02

ε2ε4 1 (1%) 5 (6%)

ε3ε3 101 (69%) 53 (55%)

ε3ε4 27 (18%) 18 (19%)

ε4ε4 0 (0%) 1 (1%)

Comorbidity, n (%)

Cardiovascular 114 (77%) 66 (68%) 0.1

Cerebrovascular 22(15%) 19 (20%) 0.3

Diabetes 29 (20%) 28 (29%) 0.09

Cholesterol, mmol/L

Total cholesterol 5.9 (1.3) 5.7 (1.1) 0.2

LDL cholesterol 4.0 (1.2) 3.8 (1.0) 0.2

HDL cholesterol 1.0 (0.3) 1.1 (0.3) 0.2

Blood pressure, mmHg

Systolic 161 (25) 157 (27) 0.2

Diastolic 85 (11) 84 (12) 0.6

Lifestyle factors

BMI 25.4 (4.4) 24.9 (3.6) 0.3

No alcohol use, n (%) 99 (67%) 67 (69%) 0.8

Nonsmokers, n (%) 144 (97%) 95 (98%) 0.7

Depressive symptoms

Zung Self-Rating Depression Scale 26.8 (5.8) 26.7 (5.5) 1

Abbreviations: ADL Activities of daily living, APOE Apolipoprotein E, BMI body mass index, HDL/LDL High-/low-density lipoprotein, IADL Instrumental activities ofdaily living, MMSE Mini Mental State Examination, SPMSQ Short Portable Mental Status QuestionnaireValues are shown as mean (SD) or number (percent). p Values were calculated with the Mann-Whitney U test or χ2 test for categorical variables. Social class iscategorized on a scale from 1 (lowest) to 10 (highest) [13]


85+ years was close to the 10-year DSI dementia predic-tion model in a younger-old population [34] and was inthe upper range of reported performance for previousdementia risk scores in younger populations [9]. Simi-larly to other dementia risk scores [9, 34], cognition wasthe main predictor, followed by functioning and educa-tion levels. However, there were several important differ-ences compared with younger populations. Age, sex, andvascular and lifestyle factors were not predictive of de-mentia in the present study, although they are usuallyimportant predictors in midlife. The age range for 85+populations is inherently smaller than for younger co-horts, potentially limiting the predictive value of age. In-dividuals who survive to the age of 85 years withoutdementia are also a selected group. While mechanismsare not fully clear, associations of vascular and lifestylefactors with dementia have been reported to differ inmidlife versus late life [10].APOE genotype was related to incident dementia, but

the pattern was different from that in younger popula-tions, where the ε4 allele increases dementia risk, whilethe ε2 allele seems protective (www.alzgene.org). In thepresent study, APOEε4 carrier status was not importantfor dementia prediction, in line with previous findings inthe oldest old [39, 40]. The ε3ε3 genotype was protect-ive, while the ε2 allele was predictive of subsequent de-mentia. Compared with younger populations, a lower

proportion of ε4 carriers and a higher proportion of ε2carriers have been reported in the oldest old [10, 40], in-cluding the Vantaa 85+ cohort [41]. Three previouspopulation-based studies with shorter follow-up thanVantaa 85+ reported no protective effect of the ε2 alleleagainst incident dementia after the age of 85 years [40,42, 43]. Additionally, the ε2 allele increased the risk ofincident vascular dementia in one study [42]. Previousreports on lower risk of dementia among the oldest oldAPOEε2 carriers have come from cross-sectional studiesof dementia prevalence at death [44], and this may notnecessarily apply longitudinally to dementia incidenceafter the age of 85 years.

Predicting dementia versus predicting neuropathologyThe APOE genotype consistently predicted AD- oramyloid-related pathologies at death on average 6 yearslater, but with a different pattern than for incident de-mentia. The ε4 allele predicted all these pathologies.ε3ε3 genotype was protective against neurofibrillary tan-gles and CAA. The ε2 allele was protective againstβ-amyloid load and neuropathological AD. This patternis closer to findings derived from younger-old popula-tions, where the ε4 allele increases the risk and ε2 alleledecreases the risk of subsequent AD-related pathology[45]. A conflicting finding was reported in the 90+Study, where ε2 carriers had increased CERAD scores in

Table 2 Dementia prediction model

Predictors AUC [95% CI] Sensitivity [95% CI] Specificity [95% CI]

Entire modela 0.73 [0.68–0.78] 0.66 [0.63–0.69] 0.68 [0.66–0.71]

Cognitionb 0.72 [0.66–0.78] 0.55 [0.51–0.59] 0.55 [0.52–0.58]

MMSE Calculationc 0.60 [0.53–0.68] 0.53 [0.49–0.56] 0.68 [0.66–0.70]

MMSE Orientationc 0.64 [0.58–0.70] 0.68 [0.65–0.71] 0.54 [0.51–0.56]

MMSE Other tasksc 0.65 [0.58–0.72] 0.56 [0.53–0.58] 0.67 [0.65–0.69]

MMSE Wordlistc 0.68 [0.62–0.75] 0.77 [0.75–0.80] 0.57 [0.55–0.60]

SPMSQc 0.71 [0.65–0.77] 0.67 [0.65–0.70] 0.64 [0.62–0.67]

MMSE totalc 0.71 [0.64–0.77] 0.62 [0.59–0.65] 0.61 [0.58–0.63]

Functioningb 0.61 [0.55–0.67] 0.62 [0.59–0.65] 0.61 [0.58–0.63]

Competence in daily activitiesc 0.61 [0.55–0.67] 0.83 [0.81–0.86] 0.35 [0.32–0.38]

Sociodemographicsb 0.60 [0.54–0.65] 0.83 [0.81–0.86] 0.35 [0.32–0.38]

Education, yearsc 0.60 [0.54–0.65] 0.66 [0.63–0.69] 0.68 [0.66–0.71]

APOE genotypeb 0.58 [0.52–0.64] 0.45 [0.42–0.47] 0.69 [0.67–0.71]

APOEε2 carriersc 0.56 [0.51–0.61] 0.25 [0.23–0.27] 0.88 [0.86–0.89]

APOEε3ε3 genotypec 0.57 [0.51–0.63] 0.45 [0.42–0.47] 0.69 [0.67–0.71]

All genotypesc,d (23/24/33/34/44) 0.58 [0.51–0.64] 0.67 [0.64–0.70] 0.66 [0.63–0.68]

Abbreviations: APOE Apoliprotein E, MMSE Mini Mental State Examination, SPMSQ Short Portable Mental Status QuestionnaireAUC, sensitivity and specificity [95% CI] values using the cutoff point Disease State Index (DSI) = 0.5 are shown for 10 × 10-fold cross-validation of the DSI model.Numbers of participants with missing data were 3 for education and 3 for APOE genotypeaOverall model performancebOverall performance of each group of related predictorscPerformance of each individual predictordCategorical variable including genotype ε2ε3, ε2ε4, ε3ε3, ε3ε4, or ε4ε4


Table 3 Neuropathology prediction models

Neuropathologicaloutcomes

Predictors AUC [95%CI] Description of predictors by neuropathological outcomecategories

Absent Present

β-Amyloid load Overall modela 0.66 [0.56 - 0.77] N=37 (23%) N=126 (77%)

APOE genotypeb 0.63 [0.56 - 0.71]

APOEε2 carriersc 0.57 [0.50 - 0.64] 10 (27%) 16 (13%)

APOEε4 carriersc 0.60 [0.55 - 0.65] 2 (5%) 31 (25%)

All genotypes (23/24/33/34/44)c,d 0.62 [0.54 - 0.70] 9/1/26/1/0(24/3/70/3/0%)

13/3/78/28/0(11/2/64/23/0%)

Functioningb 0.61 [0.51 - 0.70]

Competence in Daily Activitiesc 0.61 [0.51 - 0.70] 2.7 (1.4) 3.2 (1.4)

Tangle count Overall modela 0.64 [0.55 - 0.73] N=64 (39%) N=99 (61%)

APOE genotypeb 0.60 [0.54 - 0.67]

APOEε4 carriersc 0.61 [0.55 - 0.67] 5 (8%) 28 (29%)

APOE ε3ε3 genotypec 0.59 [0.52 - 0.66] 48 (75%) 56 (58%)

All genotypes (23/24/33/34/44)c,d 0.55 [0.48 - 0.62] 10/1/48/4/0(16/2/76/6/0%)

12/3/56/25/0(13/3/58/26/0%)

Cholesterolb 0.60 [0.51 - 0.70]

Totalc 0.61 [0.51 - 0.70] 5.6 (1.4) 5.9 (1.2)

LDLc 0.60 [0.50 - 0.69] 3.6 (1.1) 3.9 (1.0)

Functioningb 0.59 [0.50 - 0.67]

Subjective memory declinec 0.59 [0.50 - 0.67] 1.7 (0.6) 1.9 (0.7)

Neuropathological AD Overall modela 0.68 [0.61 - 0.76] N=86 (53%) N=77 (47%)

APOE genotypeb 0.65 [0.59 - 0.71]

APOEε2 carriersc 0.57 [0.51 - 0.62] 19 (23%) 7 (9%)

APOEε4 carriersc 0.62 [0.57 - 0.67] 8 (10%) 25 (33%)

All genotypes (23/24/33/34/44)c,d 0.64 [0.58 - 0.70] 17/2/59/6/0(20/2/70/7/0%)

5/2/45/23/0(7/3/60/31/0%)

Sociodemographicsb 0.62 [0.53 - 0.70]

Social classc 0.62 [0.53 - 0.70] 6.4 (1.5) 5.9 (1.2)

Functioningb 0.58 [0.50 - 0.66]

Subjective memory declinec 0.58 [0.50 - 0.66] 1.7 (0.6) 1.9 (0.7)

Cerebral amyloid angiopathy Overall modela 0.66 [0.58 - 0.74] N=56 (35%) N=103 (65%)

APOE genotypeb 0.62 [0.55 - 0.69]

APOEε4 carriersc 0.63 [0.58 - 0.68] 2 (4%) 30 (30%)


All genotypes (23/24/33/34/44)c,d 0.61 [0.55 - 0.68] 10/1/43/1/0(18/2/78/2/0%)

11/3/59/27/0(11/3/59/27/0%)

Comorbidityb 0.59 [0.52 - 0.65]

Cardiovascularc 0.59 [0.52 - 0.65] 48 (86%) 70 (68%)

Sociodemographics 0.58 [0.53 - 0.64]

Gender, menc 0.58 [0.53 - 0.64] 5 (9%) 26 (25%)

Cerebral macroinfarcts Overall modela 0.72 [0.64 - 0.79] N=83 (51%) N=80 (49%)

Comorbidityb 0.64 [0.58 - 0.70]


Table 3 Neuropathology prediction models (Continued)

Neuropathologicaloutcomes

Predictors AUC [95%CI] Description of predictors by neuropathological outcomecategories

Absent Present

Cerebrovascularc 0.64 [0.58 - 0.70] 4 (5%) 26 (33%)

Cognitionb 0.63 [0.54 - 0.72]

MMSE Wordlistc 0.63 [0.55 - 0.71] 4.6 (1.3) 4.0 (1.3)

MMSE Totalc 0.62 [0.52 - 0.71] 23.8 (4.6) 22.0 (4.5)

Lifestyleb 0.62 [0.53 - 0.70]

BMIc 0.62 [0.53 - 0.70] 23.9 (4.1) 25.5 (4.2)

Functioningb 0.59 [0.51 - 0.66]

Competence in Daily Activitiesc 0.59 [0.51 - 0.66] 2.9 (1.4) 3.3 (1.3)

Cortical macroinfarcts Overall modela 0.71 [0.63 - 0.79] N=116 (71%) N=47 (29%)

Comorbidityb 0.64 [0.57 - 0.71]


APOE genotypeb 0.60 [0.51 - 0.69]

APOEε4 carriersc 0.58 [0.51 - 0.65] 18 (16%) 15 (33%)


White matter macroinfarcts Overall modela 0.76 [0.65 - 0.87] N=140 (86%) N=23 (14%)

Cholesterolb 0.72 [0.60 - 0.83]

HDLc 0.68 [0.58 - 0.79] 1.0 (0.3) 0.9 (0.3)

LDLc 0.70 [0.58 - 0.83] 3.9 (1.0) 3.2 (1.0)

Comorbidityb 0.62 [0.51 - 0.73]


APOE genotypeb 0.61 [0.50 - 0.71]

APOEε2 carriersc 0.61 [0.50 - 0.71] 18 (13%) 8 (35%)

Cerebral microinfarcts Overall modela 0.61 [0.51 - 0.71] N=130 (83%) N=26 (17%)

Education yearsc 0.61 [0.51 - 0.71] 4.5 (2.9) 3.3 (1.8)

Hippocampal Sclerosis Overall modela 0.78 [0.64 - 0.91] N=151 (93%) N=11 (7%)

Cognitionb 0.75 [0.59 - 0.92]

MMSE Wordlistc 0.68 [0.54 - 0.83] 4.4 (1.3) 3.6 (0.9)

MMSE Other tasksc 0.74 [0.57 - 0.90] 6.9 (1.5) 5.8 (1.4)

MMSE Totalc 0.72 [0.55 - 0.90] 23.1 (4.6) 22.3 (4.1)

Lifestyleb 0.57 [0.42 - 0.71]

Current smokingc 0.57 [0.42 - 0.71] 4 (3%) 2 (18%)

TDP-43 Overall modela 0.69 [0.56 - 0.81] N=139 (86%) N=21 (13%)

Zung depression scalec 0.69 [0.56 - 0.81] 27.5 (5.8) 23.7 (2.8)

Abbreviations: AD Alzheimer’s Disease, APOE Apolipoprotein E, BMI body mass index, HDL/LDL High/low density lipoprotein, MMSE Mini-Mental State Examination,SPMSQ Short portable mental status questionnaire, TDP-43 TAR DNA-binding protein 43.AUC [95% CI] values are shown for 10*10-fold cross-validation of the DSI model. In the “Description of predictors…” columns, values are shown as mean (standarddeviation) or number (percentage). Number of participants with missing data was 4 for cerebral amyloid angiopathy, 4 for cerebral microinfarcts, , 4 for APOEgenotype, 5 for MMSE, 6 for subjective memory, 3 for education, 1 for social class, 1 for smoking, 3 for Zung scale, 10 for cholesterol, and 39 for BMI.aOverall model performance for each neuropathological outcomebOverall performance of each group of related predictors.cPerformance of each individual predictordCategorical variable including genotype ε2ε3, ε2ε4, ε3ε3, ε3ε4 or ε4ε4


cross-sectional analyses at death [44]. However, furtheranalyses showed lower cortical β-amyloid percentageareas in ε2 carriers [46].While APOE was not related to cerebral macroinfarcts

in general in the present study, the ε4 allele predictedcortical macroinfarcts, and the ε2 allele predicted whitematter macroinfarcts. A meta-analysis of studies inyounger populations has linked both the ε4 and ε2 al-leles to increasing burden in magnetic resonance im-aging markers of cerebrovascular disease, includingwhite matter hyperintensities [47]. However, longitudinalassociations of APOE genotype with subsequent cerebro-vascular lesions in the oldest old are still unclear. In theVantaa 85+ study, while white matter infarcts alone werenot significantly related to dementia diagnosis at death,they may suggest a potential explanation for the predict-ive effect of APOEε2 on incident dementia.Very few other factors besides APOE predicted neuro-

pathology. Vascular and lifestyle factors did not predictβ-amyloid load or neuropathological AD. It is still de-bated whether vascular and lifestyle risk factors for de-mentia are actually related to amyloid pathology andwhether such relationships may be age-dependent. Ourfinding that higher LDL and total cholesterol predictedtangle count needs to be verified in other 85+ cohorts.Cognitive performance was not predictive of AD- or

amyloid-related pathologies, although it predicted de-mentia, cerebral macroinfarcts, and HS. Of the includedsociodemographic factors, only lower social class pre-dicted neuropathological AD.Predictors for HS and TDP-43 pathology are still un-

clear. While current smoking was related to HS and lesspronounced depressive symptoms were related to TDP-43, the number of participants with these pathologies

was very small in this study, and these findings requireverification in other cohorts.Overall, predictive performance of the models (AUC,

sensitivity, specificity) was not very high. Whilestudy-specific limitations may have contributed to this,it is also possible that neither incident dementia nor spe-cific neuropathologies can be predicted with very highaccuracy in the oldest old using predictors commonlyemphasized in younger-old populations. This is also sug-gested by the failure of previous dementia risk scoreswhen extrapolated from younger-old to oldest-old popu-lations. Different approaches may be needed that betteraccount for the heterogeneity and multipathology oftenexisting within the 85+ age group.

Strengths and limitations of the present studyThe main strength of the present study is the prospect-ive population-based design with a high autopsy rateover 10 years, the inclusion of participants aged > 85years, and the multicomponent longitudinal predictionmodels for both dementia and specific neuropathologies.However, the developed prediction models are applicableonly to a highly selected group of individuals who sur-vive to the age of 85 years without developing dementia.External validation in other oldest-old cohorts will alsobe needed. The Vantaa 85+ population may differ frompopulations that are currently 85+ years old (e.g., forrelatively low education). Health-related measures priorto the age of 85 years were not available. Sample sizemay have limited statistical power, especially for patho-logical outcomes with smaller numbers of participants.Participants with autopsy were more likely to have in-

cident dementia and lower MMSE at baseline than thosewithout autopsy, which may have affected the pathology

OutcomesAPOE genotype

Dementia

AD/amyloid-related pathology

-Amyloid load

Tangle count

Neuropathological AD

Cerebral amyloid angiopathy

Macroinfarcts

All

Cortical

White matter

MicroinfarctsHippocampal sclerosisTDP-43

4 2 3 3

Fig. 2 Impact of apolipoprotein E (APOE) genotype on dementia versus neuropathology. Red indicates alleles that predicted dementia/pathology.Blue indicates alleles that were protective against dementia/pathology. White indicates alleles with no significant impact


models. Participants with dementia at baseline were ex-cluded from pathology models, but owing to their oldage, some pathology may have been present at baseline.Quantitative, systematic methods were used for neuro-pathological assessments, but findings based on trad-itional silver-staining methods may be somewhatdifferent from IHC methods for AD pathology. Otherpathologies, such as aging-related tau astrogliopathy,could not be included owing to lack of data.

ConclusionsThis is the first study combining longer-term dementiaand neuropathology multicomponent prediction modelsamong the oldest old. The dementia risk profile in thisage group was very different from risk profiles previouslydescribed at younger ages. Predictors of dementia didnot necessarily predict pathology. APOE genotype wasthe most consistent predictor across all models, but withdifferent impact for different alleles.The predictive models in the present study were devel-

oped for early identification of individuals with elevatedrisk of subsequent dementia. Longitudinal predictionmodels in the oldest old are more complex than inyounger-old populations, and multifactorial risk profilesincluding both genetic and nongenetic factors need tobe further investigated.

Additional file

Additional file 1: Table S1. Sensitivity and specificity of neuropathologyprediction models. Table S2. Neuropathology characteristics by dementiastatus at death for participants without dementia at baseline. Table S3. Thefirst three components of PCA for autopsy findings and prediction ofdementia at death for participants with complete autopsy data andno dementia at baseline. (PDF 40 kb)

AbbreviationsAD: Alzheimer’s disease; ADL: Activities of daily living; APOE: ApolipoproteinE; BMI: Body mass index; CAA: Cerebral amyloid angiopathy; CERAD: Consortiumto Establish a Registry for Alzheimer’s Disease; DSI: Disease State Index;HDL: High-density lipoprotein; HS: Hippocampal sclerosis; IADL: InstrumentalActivities of Daily Living Scale; LDL: Low-density lipoprotein; MMSE: Mini MentalState Examination; PCA: Principal component analysis; SPMSQ: Short PortableMental Status Questionnaire; TDP-43: Transactive response binding protein 43

AcknowledgementsNot applicable.

FundingThis study was funded by the European Union 7th Framework Program forresearch, technological development, and demonstration VPH-DARE@IT(601055); MIND-AD Academy of Finland 291803 and Swedish ResearchCouncil 529-2014-7503 (EU Joint Programme - Neurodegenerative DiseaseResearch, JPND); strategic funding for UEF-BRAIN from the University ofEastern Finland; VTR funding from Kuopio University Hospital; the Academy ofFinland (287490, 294061, 278457, 319318); Center for Innovative Medicine(CIMED) at Karolinska Institutet Sweden; Stiftelsen Stockholms sjukhem Sweden;the Knut and Alice Wallenberg Foundation (Sweden); Konung Gustaf V:s ochDrottning Victorias Frimurarstiftelse Sweden; Alzheimerfonden Sweden; SwedishResearch Council 2017-06105; and the Stockholm County Council (ALF20150589, 20170304). The study was supported by UEF Bioinformatics

computing infrastructure and HUS ERVA fund. The funding sources hadno involvement in study design; in the collection, analysis, and interpretation ofdata; in the writing of the report; or in the decision to submit the articlefor publication.

Availability of data and materialsThe datasets generated and/or analyzed during the current study are notpublicly available, owing to ethics rules and legislation in Finland. For moreinformation, please contact LM ([email protected]).

Authors’ contributionsAll authors took part in drafting or revising the manuscript for content. Thestudy concept and design and the analysis and interpretation of data weredone by AH, TPe, and AS. JL and JM provided the DSI software used foranalysis. TPo, LM, MKe, MM, MO, AP, and MT took part in the acquisition ofdata. AH, TPe, MvG, and AS contributed to the statistical analysis. AH, JL, andAS took part in study supervision and coordination. MKi, HS, and AS obtainedfunding for the study. All authors read and approved the final manuscript.

Ethics approval and consent to participateThe Vantaa 85+ study was approved by the ethics committee of theHealth Centre of the City of Vantaa, and all patients provided informedconsent. Written consent for the autopsies was given by the nearestrelatives of the deceased.

Consent for publicationNot applicable.

Competing interestsJM and JL are shareholders and cofounders of Combinostics Ltd. CombinosticsLtd. owns the following intellectual property rights related to the paper: Amethod for inferring the state of a system (US7,840,510 B2, PCT/FI2007/050277)(JL) and State inference in a heterogeneous system (PCT/FI2010/050545.FI20125177) (JL and JM).

Publisher’s NoteSpringer Nature remains neutral with regard to jurisdictional claims inpublished maps and institutional affiliations.

Author details1Institute of Clinical Medicine, Neurology, University of Eastern Finland, P.O.Box 1627, 70211 Kuopio, Finland. 2Institute for Neuroscience, NewcastleUniversity, Newcastle upon Tyne, UK. 3VTT Technical Research Centre ofFinland Ltd., Tampere, Finland. 4Division of Clinical Geriatrics, NVS, KarolinskaInstitutet, Stockholm, Sweden. 5Chronic Disease Prevention Unit, NationalInstitute for Health and Welfare, Helsinki, Finland. 6Institute of Public Healthand Clinical Nutrition, University of Eastern Finland, Kuopio, Finland.7Combinostics, Tampere, Finland. 8Department of Pathology, University ofHelsinki and HUSLAB, Helsinki, Finland. 9Department of Neurosurgery,University of Helsinki and Helsinki University Hospital, Helsinki, Finland.

Received: 6 April 2018 Accepted: 21 November 2018

References1. Prince M, Bryce R, Albanese E, Wimo A, Ribeiro W, Ferri CP. The global

prevalence of dementia: a systematic review and metaanalysis. AlzheimersDement. 2013;9(1):63–75.

2. Corrada MM, Brookmeyer R, Berlau D, Paganini-Hill A, Kawas CH. Prevalenceof dementia after age 90: results from the 90+ study. Neurology. 2008;71(5):337–43.

3. Heeren TJ, Lagaay AM, Hijmans W, Rooymans HG. Prevalence ofdementia in the ‘oldest old’ of a Dutch community. J Am Geriatr Soc.1991;39(8):755–9.

4. Rastas S, Pirttilä T, Mattila K, et al. Vascular risk factors and dementia in thegeneral population aged >85 years: prospective population-based study.Neurobiol Aging. 2010;31(1):1–7.

5. von Strauss E, Viitanen M, De Ronchi D, Winblad B, Fratiglioni L. Aging andthe occurrence of dementia: findings from a population-based cohort witha large sample of nonagenarians. Arch Neurol. 1999;56(5):587–92.


6. Tschanz JT, Treiber K, Norton MC, et al. A population study of Alzheimer’sdisease: findings from the Cache County Study on Memory, Health, andAging. Care Manag J. 2005;6(2):107–14.

7. Ebly EM, Parhad IM, Hogan DB, Fung TS. Prevalence and types of dementiain the very old: results from the Canadian Study of Health and Aging.Neurology. 1994;44(9):1593–600.

8. SJB V, MPJ VB, OJG S, et al. Modifiable risk factors for prevention ofdementia in midlife, late life and the oldest-old: validation of the LIBRAindex. J Alzheimers Dis. 2017;58(2):537–47.

9. EYH T, Harrison SL, Errington L, et al. Current developments indementia risk prediction modelling: an updated systematic review.PLoS One. 2015;10(9):e0136181.

10. Gardner RC, Valcour V, Yaffe K. Dementia in the oldest old: a multi-factorialand growing public health issue. Alzheimers Res Ther. 2013;5(4):27.

11. Kawas CH, Kim RC, Sonnen JA, Bullain SS, Trieu T, Corrada MM. Multiplepathologies are common and related to dementia in the oldest-old: the 90+ Study. Neurology. 2015;85(6):535–42.

12. Polvikoski T, Sulkava R, Haltia M, et al. Apolipoprotein E, dementia,and cortical deposition of beta-amyloid protein. N Engl J Med. 1995;333(19):1242–7.

13. Rauhala U. The social stratification of Finnish society [in Finnish]. Doctoralthesis. Porvoo: Werner Söderström Osakeyhtiö; 1966.

14. Folstein MF, Folstein SE, McHugh PR. “Mini Mental State”: a practical methodfor grading the cognitive state of patients for clinician. J Psychiatr Res. 1975;12:189–98.

15. Pfeiffer E. A short portable mental status questionnaire for the assessmentof organic brain deficit in elderly patients. J Am Geriatr Soc. 1975;23:433–41.

16. Katz S, Ford AB, Moscowitch RW, et al. Studies of illness in the aged. JAMA.1963;185:914–9.

17. Lawton MP, Brody EM. Assessment of older people: self maintaining andinstrumental activities of daily living. Gerontologist. 1969;9:179–86.

18. Zung WW, Richards CB, Short MJ. Self-rating depression scale in anoutpatient clinic. Arch Gen Psychiatry. 1965;13:508–15.

19. Syvänen AC, Sajantila A, Lukka M. Identification of individuals by analysis ofbiallelic DNA markers, using PCR and solid-phase minisequencing. Am JHum Genet. 1993;52:46–59.

20. Hixson JE, Vernier DTJ. Restriction isotyping of human apolipoprotein E bygene amplification and cleavage with HhaI. J Lipid Res. 1990;31:545–8.

21. American Psychiatric Association. Diagnostic and statistical manual of mentaldisorders. 3rd ed. Washington, DC: American Psychiatric Association; 1987.

22. McKhann G, Drachman D, Folstein M, Katzman R, Price D, Stadlan EM.Clinical diagnosis of Alzheimer’s disease: report of the NINCDS-ADRDA WorkGroup under the auspices of Department of Health and Human ServicesTask Force on Alzheimer’s Disease. Neurology. 1984;34(7):939–44.

23. Román GC, Tatemichi TK, Erkinjuntti T, et al. Vascular dementia: diagnosticcriteria for research studies: report of the NINDS-AIREN InternationalWorkshop. Neurology. 1993;43(2):250–60.

24. Ahtiluoto S, Polvikoski T, Peltonen M, et al. Diabetes, Alzheimer disease, andvascular dementia: a population-based neuropathologic study. Neurology.2010;75(13):1195–202.

25. Myllykangas L, Polvikoski T, Sulkava R, Verkkoniemi A, Crook R, Tienari PJ,Pusa AK, Niinistö L, O’Brien P, Kontula K, Hardy J, Haltia M, Pérez-Tur J.Genetic association of α2-macroglobulin with Alzheimer’s disease in aFinnish elderly population. Ann Neurol. 1999;46(3):382–90.

26. Tanskanen M, Mäkelä M, Myllykangas L, Rastas S, Sulkava R, Paetau A.Intracerebral hemorrhage in the oldest old: a population-based study(Vantaa 85+). Front Neurol. 2012;3:103.

27. Oinas M, Polvikoski T, Sulkava R, et al. Neuropathologic findings ofdementia with Lewy bodies (DLB) in a population-based Vantaa 85+study. J Alzheimers Dis. 2009;18(3):677–89.

28. Kero M, Raunio A, Polvikoski T, Tienari PJ, Paetau A, Myllykangas L.Hippocampal sclerosis in the oldest old: a Finnish population-based study. JAlzheimers Dis. 2018;63(1):263–72.

29. Mirra SS, Heyman A, McKeel D, et al. The Consortium to Establish a Registryfor Alzheimer’s Disease (CERAD): II. Standardization of the neuropathologicassessment of Alzheimer’s disease. Neurology. 1991;41:479–86.

30. Yamaguchi H, Haga C, Hirai S, Nakazato Y, Kosaka K. Distinctive, rapid, andeasy labeling of diffuse plaques in the Alzheimer brains by a newmethenamine silver stain. Acta Neuropathol. 1990;79:569–72.

31. Yamamoto T, Hirano A. A comparative study of modified Bielschowsky,Bodian and thioflavin S stains on Alzheimer’s neurofibrillary tangles.Neuropathol Appl Neurobiol. 1986;12:3–9.

32. Braak H, Braak E. Neuropathological stageing of Alzheimer-related changes.Acta Neuropathol (Berl). 1991;82:239–59.

33. Mattila J, Koikkalainen J, Virkki A, et al. A disease state fingerprint forevaluation of Alzheimer’s disease. J Alzheimers Dis. 2011;27(1):163–76.

34. Pekkala T, Hall A, Lötjönen J, et al. Development of a late-life dementiaprediction index with supervised machine learning in the population-basedCAIDE study. J Alzheimers Dis. 2017;55(3):1055–67.

35. Hall A, Mattila J, Koikkalainen J, et al. Predicting progression from cognitiveimpairment to Alzheimer’s disease with the Disease State Index. CurrAlzheimer Res. 2015;12(1):69–79.

36. Hall A, Muñoz-Ruiz M, Mattila J, et al. Generalizability of the Disease State Indexprediction model for identifying patients progressing from mild cognitiveimpairment to Alzheimer’s disease. J Alzheimers Dis. 2015;44(1):79–92.

37. Koikkalainen J, Rhodius-Meester H, Tolonen A, et al. Differential diagnosis ofneurodegenerative diseases using structural MRI data. Neuroimage Clin.2016;11:435–49.

38. Hyman BT, Phelps CH, Beach TG, et al. National Institute on Aging-Alzheimer’s Association guidelines for the neuropathologic assessment ofAlzheimer’s disease. Alzheimers Dement. 2012;8(1):1–13.

39. Juva K, Verkkoniemi A, Viramo P, et al. APOEε4 does not predict mortality,cognitive decline, or dementia in the oldest old. Neurology. 2000;54(2):412–5.

40. Corrada MM, Paganini-Hill A, Berlau DJ, Kawas CH. Apolipoprotein Egenotype, dementia, and mortality in the oldest old: the 90+ Study.Alzheimers Dement. 2013;9(1):12–8.

41. Sulkava R, Kainulainen K, Verkkoniemi A, et al. APOE alleles in Alzheimer’sdisease and vascular dementia in a population aged 85+. Neurobiol Aging.1996;17(3):373–6.

42. Skoog I, Hesse C, Aevarsson O, et al. A population study of apoE genotypeat the age of 85: relation to dementia, cerebrovascular disease, andmortality. J Neurol Neurosurg Psychiatry. 1998;64(1):37–43.

43. Qiu C, Kivipelto M, Agüero-Torres H, Winblad B, Fratiglioni L. Risk andprotective effects of the APOE gene towards Alzheimer’s disease in theKungsholmen project: variation by age and sex. J Neurol NeurosurgPsychiatry. 2004;75(6):828–33.

44. Berlau DJ, Corrada MM, Head E, Kawas CH. APOEε2 is associated with intactcognition but increased Alzheimer pathology in the oldest old. Neurology.2009;72(9):829–34.

45. Nicoll JA, Savva GM, Stewart J, Matthews FE, Brayne C, Ince P. MedicalResearch Council Cognitive Function and Ageing Study. Associationbetween APOE genotype, neuropathology and dementia in the olderpopulation of England and Wales. Neuropathol Appl Neurobiol. 2011;37(3):285–94.

46. Berlau DJ, Corrada MM, Robinson JL, et al. Neocortical β-amyloid area isassociated with dementia and APOE in the oldest-old. Alzheimers Dement.2013;9(6):699–705.

47. Schilling S, DeStefano AL, Sachdev PS, et al. APOE genotype and MRImarkers of cerebrovascular disease: systematic review and meta-analysis.Neurology. 2013;81(3):292–300.


III

Detecting amyloid positivity in elderly with increased risk of cognitive decline

Pekkala T, Hall A, Ngandu T, van Gils M, Helisalmi S, Hanninen T, Kemppainen N,Liu Y, Lotjonen J, Paajanen T, Rinne J O, Soininen H, Kivipelto M, Solomon A

Submitted to journal for publication

1

Detecting amyloid positivity in elderly with increased risk of cognitive 1 decline 2

Timo Pekkala1, Anette Hall1*, Tiia Ngandu2,3, Mark van Gils4, Seppo Helisalmi1, Tuomo 3 Hänninen5, Nina Kemppainen6,7, Yawu Liu1,8, Jyrki Lötjönen9, Teemu Paajanen10, Juha O. Rinne6,7, 4 Hilkka Soininen1,5, Miia Kivipelto1,3,l1,12, Alina Solomon1,3 5

6 1Institute of Clinical Medicine/Neurology, University of Eastern Finland, Kuopio, Finland 7 2Public Health Promotion Unit, Finnish Institute for Health and Welfare, Helsinki, Finland 8 3Division of Clinical Geriatrics, Center for Alzheimer Research, NVS, Karolinska Institutet, 9 Stockholm, Sweden 10 4VTT Technical Research Centre of Finland Ltd., Tampere, Finland; 11 5Neurocenter/ Neurology, Kuopio University Hospital, Kuopio, Finland. 12 6Turku PET Centre, University of Turku, Turku, Finland 13 7Division of clinical neurosciences, Turku University Hospital, Turku, Finland 14 8Department of Clinical Radiology, Kuopio University Hospital, Kuopio, Finland 15 9Combinostics Plc, Tampere, Finland 16 10Finnish Institute of Occupational Health, Helsinki, Finland. 17 11Institute of Public Health and Clinical Nutrition, University of Eastern Finland, Kuopio, Finland 18 l2Ageing Epidemiology Research Unit, School of Public Health, Imperial College London, London, 19 United Kingdom. 20 21 Running title: Detecting amyloid positivity in elderly 22 23 *Correspondence: Anette Hall 24 Neurology, Institute of Clinical Medicine, University of Eastern Finland 25 P.O. Box 1627, 70211 Kuopio, Finland 26 Email: [email protected] 27 Phone: +358505392167 28 29 Keywords: amyloid beta, positron emission tomography (PET), cognition, magnetic resonance 30 imaging (MRI), apolipoprotein E (APOE), machine learning, Alzheimer’s disease 31

2

Abstract 32

The importance of early interventions in Alzheimer’s disease (AD) emphasizes the need to 33 accurately and efficiently identify at-risk individuals. Although many dementia prediction models 34 have been developed, there are fewer studies focusing on detection of brain pathology. 35

We developed a model for identification of amyloid-PET positivity using data on demographics, 36 vascular factors, cognition, APOE genotype, and structural MRI, including regional brain volumes, 37 cortical thickness and a visual medial temporal lobe atrophy (MTA) rating. We also analyzed the 38 relative importance of different factors when added to the overall model. The model used baseline 39 data from the Finnish Geriatric Intervention Study to Prevent Cognitive Impairment and Disability 40 (FINGER) exploratory PET sub-study. 41

Participants were at risk for dementia, but without dementia or cognitive impairment. Their mean 42 age was 71 years. Participants underwent a brain 3T MRI and PiB-PET imaging. PiB images were 43 visually determined as positive or negative. Cognition was measured using a modified version of 44 the Neuropsychological Test Battery. Body mass index (BMI) and hypertension were used as 45 cardiovascular risk factors in the model. Demographic factors included age, gender and years of 46 education. The model was built using the Disease State Index (DSI) machine learning algorithm. 47

Of the 48 participants, 20 (42%) were rated as Aβ positive. Compared with the Aβ negative group, 48 the Aβ positive group had a higher proportion of APOE e4 carriers (53% vs. 14%), lower executive 49 functioning, lower brain volumes, and higher visual MTA rating. 50

AUC [95% CI] for the complete model was 0.78 [0.65–0.91]. MRI was the most effective factor, 51 especially brain volumes and visual MTA rating but not cortical thickness. APOE was nearly as 52 effective as MRI in improving detection of amyloid positivity. The model with the best 53 performance (AUC 0.82 [0.71–0.93]) was achieved by combining APOE and MRI. 54

Our findings suggest that combining demographic data, vascular risk factors, cognitive 55 performance, APOE genotype, and brain MRI measures can help identify Aβ positivity. Detecting 56 amyloid positivity could reduce invasive and costly assessments during the screening process in 57 clinical trials. 58

59

60

Abstract words: 334/350 61

3

1 Introduction 62

The importance of dementia prevention and early interventions in Alzheimer’s disease (AD) 63 (Winblad et al., 2016) has emphasized the increasing need for accurate identification of at-risk 64 individuals who may benefit most from such interventions. Although many dementia prediction 65 models have been developed (Hou et al., 2019), there are considerably fewer studies focusing on 66 detection of brain pathology. Given the central role attributed to beta-amyloid (Aβ) pathology in 67 AD (Dubois et al., 2007), early identification of individuals with Aβ pathology has become 68 particularly important. 69

The prevalence of Aβ pathology from ages 50 to 80 years has been estimated to range from 10 to 70 33% in cognitively normal individuals, and from 27 to 60% in individuals with mild cognitive 71 impairment (MCI) (Jansen et al., 2015). This complicates the screening process in e.g. randomized 72 controlled trials testing interventions that target Aβ, since assessment of Aβ pathology in 73 cerebrospinal fluid (CSF) or on positron emission tomography (PET) scans can easily become 74 inefficient due to invasiveness, costs, and/or PET availability. Developing models for detecting Aβ 75 pathology based on less invasive, less costly, and more easily available factors could help identify a 76 target population with high prevalence of Aβ pathology. More selective use of CSF or PET 77 assessments to confirm the presence of Aβ pathology could thus reduce costly screening failures 78 and improve screening efficiency. 79

Previous models for Aβ pathology, were most commonly developed in mixed populations including 80 individuals with AD dementia and/or MCI (e.g. Ansart et al., 2020; Apostolova et al., 2015; Bahar-81 Fuchs et al., 2013; Burnham et al., 2014; Haghighi et al., 2015; Lee et al., 2018; Palmqvist et al., 82 2019; Tosun et al., 2013, 2014; Westwood et al., 2018), with area under the receiver operating 83 characteristic curve (AUC) values up to 0.87–0.88. Very few studies have focused specifically on 84 cognitively normal populations, despite the key importance of this group who could potentially 85 benefit from interventions that are started early, before the onset of cognitive impairment. Lower 86 performance has been reported for models developed in cognitively normal individuals, with AUC 87 values up to 0.74–0.77 (Ansart et al., 2020; Insel et al., 2016; Mielke et al., 2012; ten Kate et al., 88 2018). Models for detecting Aβ pathology have most often been developed based on demographic 89 data, cognitive performance, and apolipoprotein E (APOE) genotype. Of the studies in cognitively 90 normal individuals, two have also included structural magnetic resonance imaging (MRI) (Ansart et 91 al., 2020; ten Kate et al., 2018). 92

In this study, we first aim to develop a model for detecting Aβ pathology in individuals with risk 93 factors for dementia, but without dementia or substantial cognitive impairment. We assess a broad 94 range of factors, including demographic data, cardiovascular factors, cognitive performance, APOE 95 genotype, and brain MRI measures. Both visual rating of medial temporal lobe atrophy (MTA) and 96 quantitative measures of regional brain volumes and cortical thickness are considered. The second 97 aim is to conduct a pragmatic analysis on the added value of the different factors, taking into 98 account how easily obtainable they are in clinical settings, i.e. from less complex to more 99 specialized. The model uses baseline data from the Finnish Geriatric Intervention Study to Prevent 100 Cognitive Impairment and Disability (FINGER) exploratory PET sub-study. 101

4

2 Materials and methods 102

2.1 Participants 103

The FINGER main study design and population characteristics have been previously described 104 (Kivipelto et al., 2013; Ngandu et al., 2014). In brief, FINGER (ClinicalTrials.gov identifier 105 NCT01041989) was a 2-year randomized controlled trial testing a multidomain lifestyle 106 intervention versus regular health advice in 1260 older individuals at risk for dementia from the 107 general population. Results showing intervention benefits on the primary and secondary cognitive 108 outcomes, as well as on several other outcomes, have been published (Ngandu et al., 2015). 109

The exploratory PET sub-study included 48 individuals from the FINGER main study and was 110 conducted at the Turku PET Centre. Participants had to be eligible for MRI and PET scans, in 111 addition to meeting all inclusion criteria for the FINGER main study: age 60 to 77 years; 112 Cardiovascular Risk Factors, Aging, and Dementia (CAIDE) score at or above 6 points indicating 113 elevated risk for dementia; and cognitive performance at the mean level or slightly lower than 114 expected for age according to Finnish population norms for the Consortium to Establish a Registry 115 for Alzheimer’s Disease (CERAD) test as previously described in detail (Kivipelto et al., 2013). 116 Individuals with diagnosed dementia or suspected dementia or substantial cognitive impairment 117 based on screening assessments were excluded from the study. 118

The FINGER PET population was not significantly different from the rest of the Turku cohort or 119 the rest of the FINGER participants regarding education, vascular risk factors, or APOE e4 carrier 120 status (Kemppainen et al., 2018). FINGER PET participants were slightly older than the rest of the 121 FINGER population at the baseline visit (mean 70.8 vs. 69.3 years), most likely due to a later 122 initiation of recruitment at the Turku study site (Kemppainen et al., 2018). 123

The FINGER study was approved by the Coordinating Ethics Committee of the Hospital District of 124 Helsinki and Uusimaa. All participants gave written informed consent at the screening and baseline 125 visits, and also for the exploratory neuroimaging sub-study. 126

2.2 Clinical assessments and APOE genotyping 127

The present study used data from the FINGER baseline visit, before the start of the intervention. 128 Cognition was measured using a modified version of the Neuropsychological Test Battery (mNTB) 129 (Harrison et al., 2007). A standardized composite mNTB score was determined based on 14 130 individual tests measuring three different cognitive domains, i.e. memory, executive function, and 131 processing speed. Domain-specific standardized mNTB scores were also calculated as previously 132 described (Ngandu et al., 2015), with higher scores indicating better performance. Height, weight, 133 and blood pressure were measured (Ngandu et al., 2014), and body mass index (BMI) and 134 hypertension (systolic blood pressure ≥140 mmHg and/or diastolic blood pressure ≥90 mmHg) 135 were used as vascular risk factors in the model. Genomic DNA was extracted from venous blood 136 samples with Chemagic MSM1 (PerkinElmer) using magnetic beads. APOE genotype was 137 determined by polymerase chain reaction using TaqMan genotyping assays (Applied Biosystems) 138 for 2 single-nucleotide polymorphisms (rs429358 and rs7412) and an allelic discrimination method 139 on the Applied Biosystems 7500 platform (De La Vega et al., 2005). 140

2.3 MRI and PET imaging 141

Participants in the FINGER PET sub-study underwent a brain 3T MRI (Philips Ingenuity TF 142 PET/MR, Amsterdam, the Netherlands) and 11C-Pittsburgh compound B (PiB)-PET imaging. The 143 MRI and PiB-PET protocols have been previously published (Kemppainen et al., 2018). PiB images 144 were visually determined as positive or negative by two-party consensus agreement. PiB negative 145 individuals had only non-specific 11C-PiB-PET retention in white matter, whereas PiB positive 146

5

individuals had 11C-PiB-PET retention in at least one AD-specific cortical region. 147

Regional brain volumes and cortical thickness were measured on MRI scans using FreeSurfer 148 (version 5.3, http://surfer.nmr.mgh.harvard.edu). Brain volumes were normalized to total 149 intracranial volume to take into account variations in head size. An AD-signature cortical thickness 150 measure was calculated as the average of cortical thickness in the entorhinal, inferior and middle 151 temporal, and fusiform regions (Jack et al., 2015). Additionally, visual assessment of MTA was 152 conducted by a single rater blinded to clinical data based on a T1-weighted coronal slice. MTA was 153 graded on the Scheltens scale from 0 to 4 (Scheltens et al., 1992). 154

2.4 Statistical analysis 155

The population was characterized by calculating group means and standard deviations. Statistical 156 significance of group differences was examined using the Wilcoxon rank sum test for continuous 157 and categorical data. 158

We used a machine-learning algorithm (Disease State Index, DSI) to detect PiB-PET positivity with 159 clinical, APOE and MRI data as factors. DSI is a supervised machine learning method developed at 160 the VTT Technical Research Centre of Finland (Mattila et al., 2011). Its accuracy is comparable to 161 other methods such as logistic regression, support vector machines, and Bayes inference (Mattila et 162 al., 2011), and it has been successful in modelling MCI progression (Hall et al., 2015) and 163 discriminating between dementia types (Koikkalainen et al., 2016; Tolonen et al., 2018). DSI 164 classifies individuals into Aβ positive and negative based on a population with known Aβ status 165 (training population). An individual’s data are compared with value distributions in the training 166 population. The analysis puts more weight on factors that show more pronounced dissimilarities 167 between the positive and negative groups in the training population. The resulting DSI value for an 168 individual represents a point on the scale 0–1, where higher values denote higher similarity to Aβ 169 positive individuals in the training population. A separate training population was not used in this 170 study, but the data were cross-validated by randomly selecting 80% of the population for training 171 and 20% for testing, and then repeating the procedure 100 times for statistically reliable results 172 (100×5-fold cross validation). The classification results are shown as area under the receiver-173 operator curve (AUC) for the model, with 95% confidence intervals (CI) averaged from the folds. 174

Advantages of the DSI include the ability to incorporate a large number of factors simultaneously, 175 and permissive requirements for types and distribution characteristics of the data. Missing values 176 are ignored as part of the model, and the total DSI value is calculated from the available data. Over 177 learning is a challenge especially when the sample size is small. DSI defines a classifier for each 178 predictor separately and combines these classifiers making it less sensitive to over learning than 179 methods based on complex decision boundaries. Conceptually related and potentially correlated 180 factors can be structured into groups to assess their combined effect. Individual factors are 181 combined into a group DSI value through a weighted average, and the process is then repeated 182 recursively for all groups to obtain a total DSI value. DSI thus provides detailed information about 183 performance on multiple levels simultaneously: the independent performance of each factor, the 184 combined performance of a group of similar factors, and the overall performance of the entire 185 model. 186

In this study, factors were organized into groups according to conceptual likeness: Demographic 187 (age, sex, education), Cardiovascular (BMI, hypertension), APOE genotype (ε4 carrier versus non-188 carrier), Cognition (mNTB total, memory, processing speed, and executive function), and MRI. 189 Subgroups were defined for MRI measures (Volumes, Visual MTA score, and AD-signature 190 cortical thickness) for a more detailed assessment of performance. Additional analyses were 191 conducted to assess the added value of different factor groups (modalities), taking into account how 192 easily obtainable they are in clinical settings. This was done by assessing performance of the model 193

6

after step-by-step inclusion or exclusion of different factor groups. 194

All analyses were performed using MATLAB R2015b. DSI values were calculated using 195 Fingerprint Toolbox version 0.9 on MATLAB. 196

7

3 Results 197

Population characteristics according to Aβ status on PiB-PET scans are shown in Table 1. Of the 48 198 participants, 20 (42%) were rated as Aβ positive. Compared with the Aβ negative group, the Aβ 199 positive group had a higher proportion of APOE e4 carriers (53% vs. 14%), lower executive 200 functioning, lower brain volumes (total cortical and gray matter volumes, cerebellar cortex, 201 thalamus proper, putamen, hippocampus, amygdala, the accumbens area, and ventral diencephalon), 202 and higher visual MTA rating. No significant differences were found in other characteristics (Table 203 1). 204

The performance of the complete DSI model, factor groups and individual factors is shown in Table 205 2. The AUC of the complete model after cross-validation was 0.78 (95% CI 0.65–0.91). Model 206 AUC without cross-validation (training and testing with all individuals) was 0.88. Table 3 shows 207 sensitivity, specificity, and positive and negative predictive values for different DSI cutoff values. 208 For example, setting the DSI cutoff value for positive classification at 0.5 would identify a sub-209 population with a true Aβ positivity prevalence (positive predictive value, PPV) of 65%, with 69% 210 sensitivity, 69% specificity and 77% negative predictive value (NPV). If only individuals with a 211 DSI value ≥0.5 undergo PiB-PET scans, this would lead to an increase in the rate of Aβ positive 212 scans from 42% (observed in FINGER PET study participants) to 65%. Using a higher cutoff for 213 Aβ positivity prediction, such as 0.6, could increase the positive scan rate to 74%, but at a lower 214 sensitivity (39%). 215

Among the groups of factors included in the model, structural MRI measures together had the best 216 performance, with an overall AUC (95% CI) of 0.75 (0.61–0.89). Within the MRI group, the most 217 effective subgroups were volumetric measures (AUC 0.72, CI 0.57–0.88) and visual MTA rating 218 (AUC 0.71, CI 0.59–0.84), while AD-signature cortical thickness had lower performance (AUC 219 0.65, CI 0.48–0.82). APOE ε4 carrier status had an AUC (95% CI) of 0.69 (0.56–0.82). The 220 Cognition group of factors had an AUC (95% CI) of 0.65 (0.49–0.81), and within this group 221 executive functioning had an AUC of 0.69 (0.53–0.84). Other cognitive measures did not have 222 detection power. BMI was the strongest factor (AUC 0.65, CI 0.50–0.79) in the cardiovascular 223 group, although the group level AUC (95% CI) was low at 0.60 (0.46–0.75). Demographic factors 224 had the poorest performance with years of education having the highest within-group AUC of 0.59 225 (0.43–0.75), and age and sex showing no effect. 226

Table 4 shows the added value of different groups of factors in terms of model performance. In the 227 first scenario, a base model with only demographic and cardiovascular data (AUC 0.56, CI 0.41–228 0.72) was augmented by adding a single factor group (Cognition, APOE, Visual MTA rating, or all 229 MRI measurements). The AUC improved to 0.62–0.71, but only addition of APOE or MRI data led 230 to a CI above 0.50 indicating a significant model. All MRI measurements together had the highest 231 added performance (AUC 0.71, 95% CI 0.56-0.85), followed by APOE ε4 carrier status (AUC 0.69, 232 CI 0.56–0.83). 233

The second scenario used a base model including demographic and cardiovascular factors, and 234 cognition (AUC 0.62, 95% CI 0.46–0.77). Adding APOE or all MRI measures enhanced the model 235 performance to an AUC (95% CI) of 0.71 (0.56–0.85) and 0.72 (0.58–0.87), respectively. Visual 236 MTA rating improved the AUC (95% CI) only to 0.66 (0.51–0.82). 237

The third scenario tested the added value of the simple visual MTA rating instead of the more 238 comprehensive automated MRI measures. The base model using all factors except MRI led to AUC 239 (95% CI) 0.71 (0.56–0.85). Visual MTA rating increased AUC to 0.75 (0.62–0.89), with the 240 complete model performing at 0.78 (0.65–0.91). We also tested a fourth scenario focusing on the 241 combination of APOE and MRI, the two best performing factor groups. APOE together with either 242 visual MTA rating (AUC 0.81, 95% CI 0.69–0.92) or all MRI measurements (AUC 0.82, 95% CI 243 0.71–0.93) performed better as a combination than the complete model. 244

8

The relative importance of the different subgroups of MRI factors is shown in Table 4. MRI factor 245 subgroups were removed one-by-one from a base model. With a base model including all MRI 246 measures, AUC (95% CI) decreased from 0.76 (0.62–0.90) to 0.72 (0.57–0.86) by removing 247 volumetric measures, and to 0.74 (0.60–0.89) by removing visual MTA rating. Removing AD-248 signature cortical thickness did not affect the performance of the base model. With a base model 249 including all MRI measures and APOE, removing AD-signature cortical thickness slightly 250 improved the model performance, while removing volumetric measures decreased the AUC. 251

9

4 Discussion 252

Findings from the exploratory FINGER PET sub-study suggest that a model combining 253 demographic data, vascular risk factors, cognitive performance, APOE genotype, and brain MRI 254 measures can detect Aβ positivity in older at-risk individuals without dementia or substantial 255 cognitive impairment. Given the lower prevalence of Aβ pathology among cognitively normal 256 individuals (Jansen et al., 2015), such a model would facilitate the identification of populations with 257 a considerably higher prevalence, thus reducing the number of invasive, time-consuming and costly 258 assessments during e.g. the screening process in clinical trials. Performance of the complete DSI 259 model including MRI was AUC 0.78, and 0.71 without MRI. Both could be considered 260 “acceptable” as per Hosmer et al. (2013) criteria. Previously reported models in cognitively normal 261 individuals have had AUCs in the range of 0.60–0.74 (Ansart et al., 2020; Mielke et al., 2012; ten 262 Kate et al., 2018), with the highest performance for a support vector machine (SVM) model 263 combining demographics, cognitive performance, APOE genotype and detailed structural MRI 264 measures (ten Kate et al., 2018). 265

Similar to the abovementioned SVM model (ten Kate et al., 2018), MRI and APOE were the best 266 factors in this study. Brain volumes with the highest performance (AUC≥0.70) were total cortical 267 and gray matter volumes, and hippocampus, accumbens, thalamus and putamen volumes, which 268 have been previously reported to be lower in cognitively normal Aβ positive individuals (ten Kate 269 et al., 2018). Visual MTA rating was almost as effective as brain volumes, although it was not 270 selected in the previous SVM model (ten Kate et al., 2018). The AD-signature cortical thickness 271 had lower performance than brain volumes or visual MTA rating in the FINGER PET population. 272

Very few studies in cognitively normal participants have investigated the added value of structural 273 MRI in the detection of Aβ pathology. One study reported that best results were obtained without 274 MRI, and that change in cognition over time was a superior substitute to MRI in a multimodal 275 prediction model (Ansart et al., 2020). In another study, MRI measures did have an added value 276 above other factors (ten Kate et al., 2018). Similar findings emphasizing the added value of MRI 277 were observed in the FINGER PET sub-study. In addition, the leave-one-out analysis of the MRI 278 factor group indicated brain volumes as the best factors, while AD-signature cortical thickness did 279 not have any added value. Given that visual MTA rating, which is easier to obtain in clinical 280 settings, performed almost as well as brain volumes, it may represent a useful alternative to the 281 more complex volumetric measures. 282

APOE ε4 carrier status was very effective in improving the results and adding APOE to basic 283 clinical data was almost as effective as performing an MRI scan. APOE and MRI together, in the 284 absence of any other factors, led to better performance compared with the complete model (AUC 285 0.81–0.82 versus 0.78). This is because the model showed that several factors in the demographic, 286 cardiovascular and cognitive groups were not useful in detecting amyloid positivity. 287

Regarding cognition, executive functioning was most effective, with an individual AUC of about 288 the same magnitude as APOE genotype. Cognition as a group was, however, not as valuable in 289 different combinations of factor modalities as APOE or MRI measures. In contrast, the previous 290 SVM model (ten Kate et al., 2018) emphasized memory among the tested cognitive domains. This 291 may be due to population differences, i.e. FINGER participants underwent cognitive screening to 292 select individuals with performance at the mean level or slightly lower than expected for age, thus 293 limiting the distribution of cognitive test scores in the present study. 294

Among vascular factors, BMI had some ability to detect amyloid positivity, but not hypertension. 295 Low BMI at younger-old ages has previously been associated with Aβ load (Ewers et al., 2012; 296 Toledo et al., 2012), although these studies also included individuals with MCI and dementia at 297 baseline. However, the performance of vascular factors may have been influenced by the use of the 298 CAIDE dementia risk score (including age, sex, education, BMI, systolic blood pressure, total 299

10

serum cholesterol and physical activity) (Kivipelto et al., 2006) to select the at-risk FINGER study 300 participants. FINGER eligibility criteria may also explain why age did not associate with Aβ 301 positivity in this population, despite being reported as a clear determinant of Aβ pathology in 302 individuals with normal cognition or MCI, with Aβ pathology prevalence growing rapidly after 303 about the age of 70 years (Jansen et al., 2015). 304

The main limitations of the present study are the small sample size leading to potential model 305 overfitting effects, and the lack of external validation. The same dataset was used for both training 306 and testing the DSI model, although we reported results following nested 100x5 cross-validation. 307 Findings need to be interpreted keeping in mind that FINGER participants had already undergone a 308 screening process based on cognitive testing and the CAIDE dementia risk score, i.e. they represent 309 a population at risk for dementia, but without dementia or substantial cognitive impairment. Studies 310 in independent populations will be needed to further validate the results. 311

Compared with previous studies in cognitively normal populations, the present study assessed a 312 broader range of factors, and performance at multiple levels simultaneously, i.e. from the overall 313 model to groups of conceptually related factors and also individual factors. We also investigated 314 different screening strategies, i.e. the benefit of adding more complex factor modalities, by testing 315 the performance of increasingly comprehensive models, from easily obtainable demographic, 316 clinical and cognitive data, to APOE genotyping and structural brain MRI. 317

11

Acknowledgements 318

We thank all participants and members of the FINGER study group for their cooperation in data 319 collection and management. 320

321

Funding 322

The study was funded by European Research Council grant 804371, Academy of Finland; Finnish 323 Social Insurance Institution, Alzheimer’s Research and Prevention Foundation, Juho Vainio 324 Foundation, Swedish Research Council, Alzheimerfonden, Region Stockholm ALF and NSV, 325 Center for Innovative Medicine (CIMED) at Karolinska Institutet, Knut and Alice Wallenberg 326 Foundation, Stiftelsen Stockholms sjukhem, Konung Gustaf V:s och Drottning Victorias 327 Frimurarstiftelse (Sweden); Joint Program of Neurodegenerative Disorders – prevention (MIND-328 AD), VTR grants of Turku University Hospital. JR was funded by the Sigrid Juselius Foundation, 329 Finnish State Research Funding, Academy of Finland (grant 310962). 330

331

332

Author contributions 333

Designing the study: TPe, AH, TN, JL, JR, AS, HS, MK, AS 334

Acquisition of the data: TN, SH, TH, NK, YL, TPa, JR, HS, MK 335

Interpretation of the data: TPe, AH, TN, MvG, JL, NK, JR, AS 336

Analysis of the data: TPe 337

Drafting the manuscript: TPe, AH, AS 338

Revising the manuscript, read and approved the final version: All 339

340

341

Conflict of interest 342

JL is a shareholder and co-founder of Combinostics Ltd. Combinostics Ltd owns the following IPR 343 related to the paper: 1. A method for inferring the state of a system, US7, 840,510 B2, 344 PCT/FI2007/050277. 2. State Inference in a heterogeneous system, PCT/FI2010/050545. 345 FI20125177. TPe, AH, TN, MvG, SH, TH, NK, YL, TPa, JR, HS, MK and AS have nothing to 346 disclose. 347

348

Data Availability Statement 349

The datasets for this manuscript are not publicly available because of privacy issues of sensitive 350 personal data. Requests to access the datasets should be directed to Alina Solomon, 351 [email protected]. 352

353

354

355

356

12

References 357

Ansart M, Epelbaum S, Gagliardi G, Colliot O, Dormont D, Dubois B, Hampel H, Durrleman S. 358 Reduction of recruitment costs in preclinical AD trials: validation of automatic pre-screening 359 algorithm for brain amyloidosis. Stat Methods Med Res 2020;29:151-64. 360

Apostolova LG, Hwang KS, Avila D, Elashoff D, Kohannim O, Teng E, Sokolow S, Jack CR, 361 Jagust WJ, Shaw L, Trojanowski JQ, Weiner MW, Thompson PM. Brain amyloidosis 362 ascertainment from cognitive, imaging, and peripheral blood protein measures. Neurology 363 2015;84:729-37. 364

Bahar-Fuchs A, Villemagne V, Ong K, Chetélat G, Lamb F, Reininger CB, Woodward M, Rowe 365 CC. Prediction of amyloid- β pathology in amnestic mild cognitive impairment with 366 neuropsychological tests. J Alzheimers Dis 2013;33:451-62. 367

Burnham SC, Faux NG, Wilson W, Laws SM, Ames D, Bedo J, Bush AI, Doecke JD, Ellis KA, 368 Head R, Jones G, Kiiveri H, Martins RN, Rembach A, Rowe CC, Salvado O, Macaulay SL, 369 Masters CL, Villemagne VL. A blood-based predictor for neocortical Aβ burden in Alzheimer's 370 disease: results from the AIBL study. Mol Psychiatry 2014;19:519-26. 371

De La Vega FM, Lazaruk KD, Rhodes MD, Wenz MH. Assessment of two flexible and compatible 372 SNP genotyping platforms: TaqMan SNP Genotyping Assays and the SNPlex Genotyping System. 373 Mutat Res 2005;573:111-35. 374

Dubois B, Feldman HH, Jacova C, Dekosky ST, Barberger-Gateau P, Cummings J, Delacourte A, 375 Galasko D, Gauthier S, Jicha G, Meguro K, O'brien J, Pasquier F, Robert P, Rossor M, Salloway S, 376 Stern Y, Visser PJ, Scheltens P. Research criteria for the diagnosis of Alzheimer's disease: revising 377 the NINCDS-ADRDA criteria. Lancet Neurol 2007;6:734-46. 378

Ewers M, Schmitz S, Hansson O, Walsh C, Fitzpatrick A, Bennett D, Minthon L, Trojanowski JQ, 379 Shaw LM, Faluyi YO, Vellas B, Dubois B, Blennow K, Buerger K, Teipel SJ, Weiner M, Hampel 380 H. Body mass index is associated with biological CSF markers of core brain pathology of 381 Alzheimer's disease. Neurobiol Aging 2012;33:1599-608. 382

Haghighi M, Smith A, Morgan D, Small B, Huang S. Identifying cost-effective predictive rules of 383 amyloid- β level by integrating neuropsychological tests and plasma-based markers. J Alzheimers 384 Dis 2015;43:1261-70. 385

Hall A, Muñoz-Ruiz M, Mattila J, Koikkalainen J, Tsolaki M, Mecocci P, Kloszewska I, Vellas B, 386 Lovestone S, Visser PJ, Lötjonen J, Soininen H. Generalizability of the Disease State Index 387 prediction model for identifying patients progressing from mild cognitive impairment to 388 Alzheimer's disease. J Alzheimers Dis 2015;44:79-92. 389

Harrison J, Minassian SL, Jenkins L, Black RS, Koller M, Grundman M. A neuropsychological test 390 battery for use in Alzheimer disease clinical trials. Arch Neurol 2007;64:1323-9. 391

Hosmer DW, Lemeshow S, and Sturdivant RX. Applied Logistic Regression. 3rd ed. Hoboken, 392 New Jersey: John Wiley & Sons; 2013. 393

Hou X, Feng L, Zhang C, Cao X, Tan L, Yu J. Models for predicting risk of dementia: a systematic 394 review. J Neurol Neurosurg Psychiatry 2019;90:373-9. 395

Insel PS, Palmqvist S, Mackin RS, Nosheny RL, Hansson O, Weiner MW, Mattsson N. Assessing 396 risk for preclinical β-amyloid pathology with APOE, cognitive, and demographic information. 397 Alzheimers Dement (Amst) 2016;4:76-84. 398

Jack CR Jr, Wiste HJ, Weigand SD, Knopman DS, Mielke MM, Vemuri P, Lowe V, Senjem ML, 399 Gunter JL, Reyes D, Machulda MM, Roberts R, Petersen RC. Different definitions of 400

13

neurodegeneration produce similar amyloid/neurodegeneration biomarker group findings. Brain 401 2015;138:3747-59. 402

Jansen WJ, Ossenkoppele R, Knol DL, Tijms BM, Scheltens P, Verhey FRJ, Visser PJ. Prevalence 403 of cerebral amyloid pathology in persons without dementia: a meta-analysis. JAMA 404 2015;313:1924-38. 405

Kemppainen N, Johansson J, Teuho J, Parkkola R, Joutsa J, Ngandu T, Solomon A, Stephen R, Liu 406 Y, Hänninen T, Paajanen T, Laatikainen T, Soininen H, Jula A, Rokka J, Rissanen E, Vahlberg T, 407 Peltoniemi J, Kivipelto M, Rinne JO. Brain amyloid load and its associations with cognition and 408 vascular risk factors in FINGER study. Neurology 2018;90:e206-e213. 409

Kivipelto M, Ngandu T, Laatikainen T, Winblad B, Soininen H, Tuomilehto J. Risk score for the 410 prediction of dementia risk in 20 years among middle aged people: a longitudinal, population-based 411 study. Lancet Neurol 2006;5:735-41. 412

Kivipelto M, Solomon A, Ahtiluoto S, Ngandu T, Lehtisalo J, Antikainen R, Bäckman L, Hänninen 413 T, Jula A, Laatikainen T, Lindström J, Mangialasche F, Nissinen A, Paajanen T, Pajala S, Peltonen 414 M, Rauramaa R, Stigsdotter-Neely A, Strandberg T, Tuomilehto J, Soininen H. The Finnish 415 Geriatric Intervention Study to Prevent Cognitive Impairment and Disability (FINGER): study 416 design and progress. Alzheimers Dement 2013;9:657-65. 417

Koikkalainen J, Rhodius-Meester H, Tolonen A, Barkhof F, Tijms B, Lemstra AW, Tong T, 418 Guerrero R, Schuh A, Ledig C, Rueckert D, Soininen H, Remes AM, Waldemar G, Hasselbalch S, 419 Mecocci P, van der Flier W, Lötjönen J. Differential diagnosis of neurodegenerative diseases using 420 structural MRI data. Neuroimage Clin 2016;11:435-49. 421

Lee JH, Byun MS, Yi D, Sohn BK, Jeon SY, Lee Y, Lee J, Kim YK, Lee Y, Lee DY. Prediction of 422 cerebral amyloid with common information obtained from memory clinic practice. Front Aging 423 Neurosci 2018;10:309. 424

Mattila J, Koikkalainen J, Virkki A, Simonsen A, van Gils M, Waldemar G, Soininen H, Lötjönen 425 J. A Disease State Fingerprint for evaluation of Alzheimer's disease. J Alzheimers Dis 2011;27:163-426 76. 427

Mielke MM, Wiste HJ, Weigand SD, Knopman DS, Lowe VJ, Roberts RO, Geda YE, Swenson-428 Dravis DM, Boeve BF, Senjem ML, Vemuri P, Petersen RC, Jack CR Jr. Indicators of amyloid 429 burden in a population-based study of cognitively normal elderly. Neurology 2012;79:1570-7. 430

Ngandu T, Lehtisalo J, Levälahti E, Laatikainen T, Lindström J, Peltonen M, Solomon A, Ahtiluoto 431 S, Antikainen R, Hänninen T, Jula A, Mangialasche F, Paajanen T, Pajala S, Rauramaa R, 432 Strandberg T, Tuomilehto J, Soininen H, Kivipelto M. Recruitment and baseline characteristics of 433 participants in the Finnish Geriatric Intervention Study to Prevent Cognitive Impairment and 434 Disability (FINGER)—a randomized controlled lifestyle trial. Int J Environ Res Public Health 435 2014;11:9345-60. 436

Ngandu T, Lehtisalo J, Solomon A, Levälahti E, Ahtiluoto S, Antikainen R, Bäckman L, Hänninen 437 T, Jula A, Laatikainen T, Lindström J, Mangialasche F, Paajanen T, Pajala S, Peltonen M, 438 Rauramaa R, Stigsdotter-Neely A, Strandberg T, Tuomilehto J, Soininen H, Kivipelto M. A 2 year 439 multidomain intervention of diet, exercise, cognitive training, and vascular risk monitoring versus 440 control to prevent cognitive decline in at-risk elderly people (FINGER): a randomised controlled 441 trial. Lancet 2015;385:2255-63. 442

Palmqvist S, Insel PS, Zetterberg H, Blennow K, Brix B, Stomrud E, Mattsson N, Hansson O. 443 Accurate risk estimation of β -amyloid positivity to identify prodromal Alzheimer's disease: Cross-444 validation study of practical algorithms. Alzheimers Dement 2019;15:194-204. 445

14

Scheltens P, Leys D, Barkhof F, Huglo D, Weinstein HC, Vermersch P, Kuiper M, Steinling M, 446 Wolters EC, Valk J. Atrophy of medial temporal lobes on MRI in "probable" Alzheimer's disease 447 and normal ageing: diagnostic value and neuropsychological correlates. J Neurol Neurosurg 448 Psychiatry 1992;55:967-72. 449

ten Kate M, Redolfi A, Peira E, Bos I, Vos SJ, Vandenberghe R, Gabel S, Schaeverbeke J, 450 Scheltens P, Blin O, Richardson JC, Bordet R, Wallin A, Eckerstrom C, Molinuevo JL, 451 Engelborghs S, Van Broeckhoven C, Martinez-Lage P, Popp J, Tsolaki M, Verhey FRJ, Baird AL, 452 Legido-Quigley C, Bertram L, Dobricic V, Zetterberg H, Lovestone S, Streffer J, Bianchetti S, 453 Novak GP, Revillard J, Gordon MF, Xie Z, Wottschel V, Frisoni G, Visser PJ, Barkhof F. MRI 454 predictors of amyloid pathology: results from the EMIF-AD Multimodal Biomarker Discovery 455 study. Alzheimers Res Ther 2018;10:100. 456

Toledo JB, Toledo E, Weiner MW, Jack CR Jr, Jagust W, Lee VMY, Shaw LM, Trojanowski JQ. 457 Cardiovascular risk factors, cortisol, and amyloid-��β deposition in Alzheimer's Disease 458 Neuroimaging Initiative. Alzheimers Dement 2012;8:483-9. 459

Tolonen A, Rhodius-Meester HFM, Bruun M, Koikkalainen J, Barkhof F, Lemstra AW, Koene T, 460 Scheltens P, Teunissen CE, Tong T, Guerrero R, Schuh A, Ledig C, Baroni M, Rueckert D, 461 Soininen H, Remes AM, Waldemar G, Hasselbalch SG, Mecocci P, van der Flier WM, Lötjönen J. 462 Data-Driven Differential Diagnosis of Dementia Using Multiclass Disease State Index Classifier. 463 Front Aging Neurosci 2018;10:111. 464

Tosun D, Joshi S, Weiner MW. Neuroimaging predictors of brain amyloidosis in mild cognitive 465 impairment. Ann Neurol 2013;74:188-98. 466

Tosun D, Joshi S, Weiner MW. Multimodal MRI-based imputation of the Aβ + in early mild 467 cognitive impairment. Ann Clin Transl Neurol 2014;1:160-70. 468

Westwood S, Baird AL, Hye A, Ashton NJ, Nevado-Holgado AJ, Anand SN, Liu B, Newby D, 469 Bazenet C, Kiddle SJ, Ward M, Newton B, Desai K, Tan Hehir C, Zanette M, Galimberti D, 470 Parnetti L, Lleó A, Baker S, Narayan VA, van der Flier WM, Scheltens P, Teunissen CE, Visser PJ, 471 Lovestone S. Plasma protein biomarkers for the prediction of CSF amyloid and tau and [18F]-472 Flutemetamol PET scan result. Front Aging Neurosci 2018;10:409. 473

Winblad B, Amouyel P, Andrieu S, Ballard C, Brayne C, Brodaty H, Cedazo-Minguez A, Dubois 474 B, Edvardsson D, Feldman H, Fratiglioni L, Frisoni GB, Gauthier S, Georges J, Graff C, Iqbal K, 475 Jessen F, Johansson G, Jönsson L, Kivipelto M, Knapp M, Mangialasche F, Melis R, Nordberg A, 476 Rikkert MO, Qiu C, Sakmar TP, Scheltens P, Schneider LS, Sperling R, Tjernberg LO, Waldemar 477 G, Wimo A, Zetterberg H. Defeating Alzheimer's disease and other dementias: a priority for 478 European science and society. Lancet Neurol 2016;15:455-532. 479

15

Table 1. Population characteristics according to amyloid status on PiB-PET scans. 480

Mean (SD) Amyloid- (n=28) Amyloid+ (n=20) p-value Demographic Sex/Female 14 (50 %) 8 (40 %) 0.505 Age (years) 70.2 (5.8) 71.6 (3.5) 0.310 Education (years) 9.7 (2.9) 8.9 (2.0) 0.320 Cardiovascular Body mass index 27.9 (3.6) 26.2 (2.6) 0.088 High blood pressure 10 (36 %) 9 (45 %) 0.529 APOE (e4 carrier)† 4 (14 %) 10 (53 %) 0.005 Cognition mNTB Total 0.04 (0.53) - 0.09 (0.50) 0.421 mNTB Memory - 0.11 (0.52) 0.04 (0.64) 0.385 mNTB Processing speed 0.16 (0.95) - 0.10 (0.77) 0.184 mNTB Executive function 0.16 (0.58) - 0.22 (0.44) 0.026 MRI Volumes Total cortex 0.29 (0.03) 0.27 (0.03) 0.007 Total gray matter 0.39 (0.05) 0.36 (0.04) 0.009 Cerebellum cortex 0.063 (0.009) 0.059 (0.008) 0.027 Thalamus proper 9.3E-3 (1.2E-3) 8.4E-3 (1.1E-3) 0.022 Caudate 4.9E-3 (8.6E-4) 4.5E-3 (7.7E-4) 0.070 Putamen 7.0E-3 (1.3E-3) 6.1E-3 (1.0E-3) 0.014 Pallidum 1.9E-3 (3.5E-4) 1.8E-3 (2.9E-4) 0.198 Brain Stem 0.014 (0.002) 0.014 (0.002) 0.229 Hippocampus 5.2E-3 (9.7E-4) 4.6E-3 (8.1E-4) 0.019 Amygdala 2.3E-3 (4.6E-4) 2.0E-3 (3.0E-4) 0.030 Accumbens area 6.6E-4 (1.3E-4) 5.6E-4 (1.3E-4) 0.004 Ventral diencephalon 5.0E-3 (5.7E-4) 4.7E-3 (4.1E-4) 0.037 Cerebrospinal fluid 8.8E-4 (1.4E-4) 8.1E-4 (1.4E-4) 0.171 Optic chiasm 1.4E-4 (3.6E-5) 1.2E-4 (4.6E-5) 0.164 Total corpus callosum 2.0E-3 (4.3E-4) 1.7E-3 (3.8E-4) 0.058 Visual MTA rating (Scheltens) 1.0 (0.7) 1.6 (0.7) 0.007 AD-signature cortical thickness 2.8 (0.1) 2.7 (0.1) 0.084 The Wilcoxon rank sum test was used to calculate p-values for differences between amyloid positive and negative groups. Key: APOE, Apolipoprotein E; mNTB, modified Neuropsychological Test Battery; MRI, magnetic resonance imaging; MTA, medial temporal lobe atrophy; AD, Alzheimer’s disease. †: Amyloid+ n=19

481

16

Table 2. Performance for the full model, factor groups and individual factors for detecting 482 amyloid positivity. 483

Amyloid+ relative to amyloid-

AUC (95% CI) Composite DSI 0.78 (0.65–0.91) Demographic 0.54 (0.37–0.70) Sex/Female ↓ 0.48 (0.35–0.60) Age ↑ 0.45 (0.28–0.61) Education (years) ↓ 0.59 (0.43–0.75) Cardiovascular 0.60 (0.46–0.75) Body mass index ↓ 0.65 (0.50–0.79) High blood pressure ↑ 0.49 (0.37–0.61) APOE (e4 carrier) ↑ 0.69 (0.56–0.82) Cognition 0.65 (0.49–0.81) mNTB Total ↓ 0.55 (0.38–0.72) mNTB Memory ↑ 0.54 (0.38–0.70) mNTB Processing speed ↓ 0.57 (0.41–0.73) mNTB Executive function ↓ 0.69 (0.53–0.84) MRI 0.75 (0.61–0.89) Volumes 0.72 (0.57–0.88) Total cortex ↓ 0.73 (0.59–0.88) Total gray matter ↓ 0.72 (0.57–0.88) Cerebellum cortex ↓ 0.69 (0.54–0.84) Thalamus proper ↓ 0.70 (0.55–0.85) Caudate ↓ 0.65 (0.49–0.81) Putamen ↓ 0.71 (0.56–0.87) Pallidum ↓ 0.61 (0.45–0.77) Brain Stem ↓ 0.61 (0.45–0.77) Hippocampus ↓ 0.70 (0.54–0.86) Amygdala ↓ 0.69 (0.53–0.85) Accumbens area ↓ 0.75 (0.62–0.89) Ventral diencephalon ↓ 0.68 (0.53–0.83) Cerebrospinal fluid ↓ 0.61 (0.44–0.78) Optic chiasm ↓ 0.60 (0.41–0.78) Total corpus callosum ↓ 0.62 (0.45–0.79) Visual MTA rating (Scheltens) ↑ 0.71 (0.59–0.84) AD-signature cortical thickness ↓ 0.65 (0.48–0.82) Upward arrow indicates the amyloid+ group having a larger mean value. AUC values (95% confidence interval) are from 100x5-cross-validation. Key: AUC, area under the receiver operating characteristic curve; DSI, disease state index; APOE, Apolipoprotein E; mNTB, modified neuropsychological test battery; MRI, magnetic resonance imaging; MTA, medial temporal lobe atrophy; AD, Alzheimer’s disease. 484

17

Table 3. Classifier statistics for selected DSI cutoff levels for detecting amyloid positivity. 485

DSI cutoff Sensitivity (%) Specificity (%) PPV (%) NPV (%)

RPP (%) 0.0 100 0 42 (42–42) 100 0.1 100 1 (0–1) 42 (42–42) 100 100 (99–100) 0.2 100 11 (10–12) 45 (45–45) 100 93 (93–94) 0.3 99 (98–99) 26 (24–27) 49 (49–50) 98 (96–99) 85 (84–86) 0.4 87 (86–89) 49 (47–50) 57 (55–58) 87 (85–88) 66 (65–68) 0.5 69 (67–71) 69 (68–71) 65 (63–66) 77 (76–79) 47 (45–48) 0.6 39 (37–41) 89 (88–90) 74 (71–77) 68 (67–69) 23 (22–24) 0.7 17 (16–19) 96 (95–97) 77 (74–81) 62 (62–63) 9 (9–10) 0.8 12 (11–13) 99 (99–100) 94 (92–97) 61 (61–62) 5 (5–6) 0.9 1 (1–2) 100 (100–100) 100 (100–100) 59 (58–59) 1 (0–1) 1.0 0 100 58 (58–58) 0

Values are mean values (95% confidence interval) from 100x5-cross-validation. Values with no 486 confidence interval are exact values. 487

Key: DSI, Disease state index; PPV, positive predictive value; NPV, negative predictive value, RPP 488 rate of positive prediction. 489

490 491

18

Table 4. Added value of different groups of factors in terms of model performance. In each 492 scenario (highlighted in bold type), factor groups are added one-by-one to a base model. In 493 the MRI subgroup analysis, factor groups are removed one-by-one from a base model. 494

AUC (95% CI) Demographic and Cardiovascular 0.56 (0.41–0.72)

and additionally Cognition 0.62 (0.46–0.77) APOE (e4 carrier) 0.69 (0.56–0.83) Visual MTA rating 0.66 (0.51–0.81) MRI all modalities 0.71 (0.56–0.85) Demographic, Cardiovascular, and Cognition 0.62 (0.46–0.77)

and additionally APOE 0.71 (0.56–0.85) Visual MTA rating 0.66 (0.51–0.82) MRI all modalities 0.72 (0.58–0.87) Demographic, Cardiovascular, Cognition, and APOE 0.71 (0.56–0.85)

and additionally Visual MTA rating 0.75 (0.62–0.89) MRI all modalities (complete model) 0.78 (0.65–0.91) APOE 0.69 (0.57–0.81)

and additionally Visual MTA rating 0.81 (0.69–0.92) MRI all modalities 0.82 (0.71–0.93) MRI subgroup analysis All MRI modalities 0.76 (0.62–0.90)

without Volumes 0.72 (0.57–0.86) Visual MTA rating 0.74 (0.60–0.89) AD-signature cortical thickness 0.76 (0.62–0.89) All MRI modalities and APOE 0.82 (0.71–0.93)

without Volumes 0.79 (0.66–0.91) Visual MTA rating 0.82 (0.71–0.93) AD-signature cortical thickness 0.83 (0.71–0.95) AUC values (95% confidence interval) are 100x5-cross-validated. Key: MRI, magnetic resonance imaging; AUC, area under the receiver operating characteristic curve; APOE, Apolipoprotein E; MTA, medial temporal lobe atrophy; AD, Alzheimer’s disease. 495

IV

Association of peripheral insulin resistance and other markers of type 2 diabetesmellitus with brain amyloid deposition in healthy individuals at risk of dementia

Pekkala T, Hall A, Mangialasche F, Kemppainen N, Mecocci P, Ngandu T, Rinne J O,Soininen H, Tuomilehto J, Kivipelto M and Solomon A

Submitted to journal for publication

Association of peripheral insulin resistance and other markers of type 2 diabetes mellitus with brain amyloid deposition in healthy individuals at risk of dementia

Timo Pekkalaa, Anette Halla, Francesca Mangialascheb,c, Nina Kemppainend,e, Patrizia Mecoccif, Tiia Ngandub,g, Juha O. Rinned,e, Hilkka Soininena,h, Jaakko Tuomilehtog,i,j, Miia Kivipeltoa,b,k,l, Alina Solomona,b

aInstitute of Clinical Medicine/Neurology, University of Eastern Finland, Kuopio, Finland bDivision of Clinical Geriatrics, Center for Alzheimer Research, NVS, Karolinska Institutet, Stockholm, Sweden cAging Research Center, NVS, Karolinska Institutet and Stockholm University, Stockholm, Sweden dTurku PET Centre, University of Turku, Turku, Finland eDivision of clinical neurosciences, Turku University Hospital, Turku, Finland fDepartment of Medicine, Institute of Gerontology and Geriatrics, University of Perugia, Perugia, Italy gPublic Health Promotion Unit, Finnish Institute for Health and Welfare, Helsinki, Finland hNeurocenter, Neurology, Kuopio University Hospital, Kuopio, Finland iDepartment of Public Health, University of Helsinki, Helsinki, Finland jNational School of Public Health, Madrid, Spain kInstitute of Public Health and Clinical Nutrition, University of Eastern Finland, Kuopio, Finland lAgeing Epidemiology Research Unit, School of Public Health, Imperial College London, London, United Kingdom. Running title: Association of IR with amyloid deposition *Corresponding author: Anette Hall Neurology, Institute of Clinical Medicine, University of Eastern Finland P.O. Box 1627 70211 Kuopio Finland Email: [email protected] Phone: +358-50-5392167

Abstract:

We explored the association of type 2 diabetes related blood markers with brain amyloid accumulation on

PiB-PET scans in 41 participants from the FINGER PET sub-study. We built a logistic regression model for

brain amyloid status with 12 plasma markers of glucose and lipid metabolism, controlled for diabetes and

APOE ε4 carrier status. Lower levels of insulin, insulin resistance index (HOMA-IR), C-peptide, and

plasminogen activator (PAI-1) were associated with amyloid positive status. Associations were significant at

the 90% confidence level after adjusting for multiple testing. No association was found between amyloid

status and fasting glucose or HbA1c.

Keywords: amyloid beta, positron emission tomography (PET), plasminogen activator (PAI-1), type 2

diabetes, apolipoprotein E (APOE).

Abstract word count: 97

Introduction

The association of type 2 diabetes (DM2) and insulin resistance (IR) with formation of beta-amyloid plaques

in Alzheimer’s disease (AD) is still unclear. DM2 is a risk factor of AD dementia [1], but evidence on

associations with AD-type pathology has been mixed. IR, an indicator of a pre-diabetic state and a hallmark

of DM2, has been associated with amyloid accumulation in middle-aged [2] and late-middle-aged [3], but

not in older [4] cognitively healthy individuals. A study of older individuals reported no relation of lifelong

IR exposure with either in vivo or post-mortem measures of brain amyloid pathology [5]. The mechanisms

linking amyloid accumulation and insulin resistance are unclear, and hypotheses include e.g. dysregulated

hormonal signaling and “type 3 diabetes” of the brain [6]. We explored the association of peripheral blood

markers of IR and DM2 with amyloid accumulation on PET scans in older individuals at risk for dementia,

but without dementia or substantial cognitive decline.

Methods

The study population included 41 participants in the Finnish Geriatric Intervention Study to Prevent

Cognitive Impairment and Disability (FINGER) exploratory PET sub-study who had available data on IR and

DM2-related blood markers (fasting blood glucose, insulin, HbA1c, and a 12-item Bio-Plex Pro Human

Diabetes assay). The FINGER main study [7] and PET sub-study [8] have previously been described in detail.

Briefly, FINGER investigated the cognitive benefits of a 2-year multidomain lifestyle intervention versus

regular health advice in 1260 individuals aged 60–77 years from the general population. Inclusion criteria

were elevated dementia risk based on Cardiovascular Risk Factors, Aging and Dementia (CAIDE) score [9],

and cognitive performance at mean level or slightly lower than expected for age according to Finnish

population norms for the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) test. 11C-

Pittsburgh compound B (PiB) PET scans were conducted at the Turku PET Centre, and visually assessed as

amyloid positive or negative by two-party consensus agreement [8]. The FINGER study was approved by the

Coordinating Ethics Committee of the Hospital District of Helsinki and Uusimaa. All participants gave

written informed consent at the screening and baseline visits, and for the exploratory neuroimaging sub-

study.

In the FINGER study, venous blood samples were taken at baseline in fasting status and using EDTA tubes.

Plasma aliquots were stored at -80C until analysis. Twelve markers related to glucose and lipid metabolism

were analysed using the multiplex suspension array system Bio-Plex Luminex® 200 instrument, (Bio-Rad

Laboratories, Hercules, CA, USA), with the Bio-Plex Pro Human Diabetes 10-plex panel (C-peptide; ghrelin;

GIP: gastric inhibitory polypeptide; GLP-1: glucagon-like peptide-1; glucagon; insulin; leptin; PAI-1; resistin;

visfatin) and 2-plex panel (adiponectin, adipsin) (Bio-Rad Laboratories, Hercules, CA, USA). The assays were

performed in one batch, and samples preparation and setting of the system running protocol were done

following the manufacturer’s instructions (www.Biorad.com).

In brief, plasma samples (10 µl) were first diluted (1:400 for adipsin adiponectin; 1:4 for all the other

compounds) using Serum Based Diluent provided by the manufacturer. Assay beads (50 µl) were

transferred in the 96-well plates and washed twice with wash buffer. Then, standards and plasma samples

(50 µl) were added to the appropriate wells. Plates were incubated for 1 h in a dark room, with mild

agitation at room temperature. The fluid was then removed by vacuum and wells were washed three times

with wash buffer. Detection antibodies (25 µl) were added to each well, and plates were incubated for 30

min in a dark room, with mild agitation at room temperature. The fluorescent conjugate Streptavidin-

Phycoerythrin (60 µl) was then added to each well and plates incubated for 10 minutes at room

temperature. Finally, plates were washed with wash buffer and assay beads were resuspended in assay

buffer (125 µl in each well), and plate reading was done with the Bio-Plex Luminex® 200 instrument.

All samples and standards were run in duplicate and were measured as pg/ml. Quality controls were

performed according to the manufacturer guidelines to ensure accuracy of measurements. The intra-assay

coefficient of variability (CV) for all compounds ranged from 1 to 2%. The detection limits (in pg/ml) were

the following: 32.7 for Adiponectin; 17 for Adipsin; 14.5 for C-peptide; 1.2 for ghrelin; 0.8 for GIP; 5.3 for

GLP-1; 4.9 for glucagon; 1 for insulin; 3.1 for leptin; 2.2 for PAI-1; 1.3 for resistin; 37.1 for visfatin). All

sample results below the lower limit of quantitation were classified as missing data.

After the plate reading, the results files were generated using Bio-Plex Manager software 4 (Bio-Rad

Laboratories, Hercules, CA, USA).

Data from the FINGER baseline visit, before the start of the intervention, was used in this study. The

homeostatic model assessment of insulin resistance index (HOMA-IR) was calculated based on fasting blood

insulin and glucose measures. APOE genotype was determined as described previously [10]. We built a

logistic regression model for brain amyloid status based on log-transformed diabetes markers, and diabetes

status (yes/no) and APOE ε4 carrier status (yes/no) as confounders. We corrected for multiple comparisons

using the false discovery rate (FDR) method [11]. All analyses were performed using MATLAB R2017b

(function mnrfit).

Results

The participants’ mean age (SD) was 71.1 (5.0) years, 51.2% were female, 14.6% had diabetes, 29.3% were

APOE ε4 carriers, and 39.0% had amyloid positive PiB-PET scans. Table 1 summarizes population

characteristics by amyloid status. APOE ε4 prevalence was significantly higher in the amyloid positive group

(56.3% vs. 12.0%, p=0.003). No statistically significant differences in age, sex, body mass index (BMI) or

diabetes status were observed. IR and DM2-related markers by amyloid status are shown in Table 2. Insulin

and plasminogen activator inhibitor-1 (PAI-1) concentrations were significantly lower in amyloid positive

compared with negative individuals (p<0.05). C-peptide and HOMA-IR were also lower in the amyloid

positive groups, but differences were significant only at the 90% confidence level. Other markers showed

no significant between-group differences.

In the logistic regression model adjusted for diabetes status and APOE genotype (Table 2), higher levels of

insulin, HOMA-IR, C-peptide, and PAI-1 were significantly associated with lower odds of amyloid positivity.

After FDR correction, these four markers were significant only at the 90% confidence level. Models with

either BMI, age, or sex as additional confounders showed a similar pattern, including after FDR correction.

Discussion

Results from the FINGER PET exploratory sub-study suggested slightly better insulin homeostasis and

related markers in amyloid positive older individuals at risk for dementia, but without dementia or

substantial cognitive impairment. Although DM2 has been indicated as a risk factor for dementia and AD

[1], it is not fully clear whether the underlying mechanisms are amyloid-related. Peripheral IR has been

linked to IR in the brain, which may affect Aβ pathology through e.g. neuroinflammatory pathways or

competitive cleavage of insulin and amyloid by the same enzyme [6]. However, previous studies

investigating peripheral IR and brain amyloid in CSF or on PET scans in cognitively healthy individuals have

reported no associations at older ages, and mixed findings in middle-aged populations [2-4,12]. Ekblad et

al. [4] found that midlife HOMA-IR, taken 15 years previous to the PET -scan, associated with greater brain

amyloid accumulation in elderly individuals without dementia, in both carriers and noncarriers of APOE e4

genotype, but the same association was not detected at the time of the scan in late-life. No associations

have also been reported in people with mild cognitive impairment or AD [4].

Peripheral IR and DM2 may also contribute to cognitive decline via vascular-related mechanisms.

Interestingly, an autopsy study in the 85+ age group showed less amyloid pathology and more

cerebrovascular pathology in people with diabetes, who also had an increased risk of AD dementia [13]. It

was suggested that, in people with diabetes and vascular pathology, less amyloid pathology may be needed

to trigger the onset of dementia. FINGER study participants had cognitive performance at the mean level or

slightly lower than expected for age, and they were also at risk for dementia based on the CAIDE score

including age, sex, education, BMI, systolic blood pressure, total cholesterol, and physical activity [9]. The

cognitive performance of a participant could be hypothesized to be determined by a similar interplay of

amyloid- and vascular-related mechanisms, e.g. poorer insulin homeostasis and higher vascular-related risk

in amyloid negative individuals. It is also possible that FINGER participants with lower IR represent a

selected group that has maintained better insulin homeostasis despite elevated cardiovascular risk. The

mechanisms linking IR and amyloid accumulation may be different in this specific group.

To our knowledge, this is the first study investigating the association of peripheral blood PAI-1 level and in-

vivo brain amyloid markers. Higher PAI-1 level seemed to be protective against amyloid accumulation,

although not significantly after correction for multiple comparisons. PAI-1 in CSF has in prior studies been

reported to have no association with AD status [14,15]. PAI-1 downregulates the activity of the protein-

cleaving plasmin system, and it is considered a risk factor for atherosclerosis due to its prothrombotic

effect. In the population-based Framingham Offspring Study higher fasting insulin level was associated,

among other things, with increased PAI-1 levels [16]. In the brain, however, PAI-1 and the plasmin system

may interact with amyloid fibrils and possibly affect plaque formation [17], or be directly neuroprotective

[18]. A previous study showed that increases in blood levels of PAI-1 were associated with WM integrity

loss in stroke-free, cognitively normal individuals aged 50-65 years and also reported an association of PAI-

1 with lower performance in speed or visuomotor coordination [19]. It is unclear if the association

suggested in the present study was mediated through effects in the brain, or if amyloid load was reduced

due to effects in the cardiovascular system.

The main strength of this study is the assessment of a comprehensive assay of IR and DM2-related markers

in relation to brain amyloid accumulation on PiB-PET, which has not been previously done. However, the

small sample size is a key limitation, restricting statistical power and the ability to control for other

potentially confounding factors. Our exploratory study adds to the growing amount of data on the

associations of IR and DM2-related markers with AD-related pathology. Future studies in larger populations

and with longitudinal data should further investigate these associations, taking into account e.g. APOE

genotype, degree of vascular pathology, type of DM2 treatment and level of glycemic control.

Acknowledgements

The study was funded by European Research Council grant 804371; Academy of Finland, Finnish Social

Insurance Institution, Alzheimer’s Research and Prevention Foundation, Alzheimer Association, Yrjö

Jahnsson Foundation, Juho Vainio Foundation (Finland); Swedish Research Council, Alzheimerfonden,

Region Stockholm ALF and NSV, Center for Innovative Medicine (CIMED) at Karolinska Institutet, Knut and

Alice Wallenberg Foundation, Stiftelsen Stockholms sjukhem, Konung Gustaf V:s och Drottning Victorias

Frimurarstiftelse (Sweden); Joint Program of Neurodegenerative Disorders – prevention (MIND-AD). J.O.

Rinne was funded by the Sigrid Juselius Foundation, Finnish State Research Funding, Academy of Finland

(grant 310962).

We thank all participants and members of the FINGER study group for their cooperation in data collection

and management.

Disclosure Statement

The authors have no conflict of interest to report.

References

[1] Cheng, G., Huang, C., Deng, H., Wang, H., 2012. Diabetes as a risk factor for dementia and mild

cognitive impairment: a meta-analysis of longitudinal studies. Intern Med J 42, 484-91.

[2] Ekblad LL, Johansson J, Helin S, Viitanen M, Laine H, Puukka P, Jula A, Rinne JO. Midlife insulin

resistance, APOE genotype, and late-life brain amyloid accumulation. Neurology. 2018 Mar

27;90(13):e1150-e1157.

[3] Willette AA, Johnson SC, Birdsill AC, Sager MA, Christian B, Baker LD, Craft S, Oh J, Statz E, Hermann

BP, Jonaitis EM, Koscik RL, La Rue A, Asthana S, Bendlin BB. Insulin resistance predicts brain amyloid

deposition in late middle-aged adults. Alzheimers Dement. 2015 May;11(5):504-510.e1.. Alzheimers

Dement 11, 504-510.e1.

[4] Laws SM, Gaskin S, Woodfield A, Srikanth V, Bruce D, Fraser PE, Porter T, Newsholme P, Wijesekara

N, Burnham S, Doré V, Li QX, Maruff P, Masters CL, Rainey-Smith S, Rowe CC, Salvado O, Villemagne

VL, Martins RN, Verdile G., 2017. Insulin resistance is associated with reductions in specific cognitive

domains and increases in CSF tau in cognitively normal adults. Sci Rep 7, 9766.

[5] Thambisetty M, Jeffrey Metter E, Yang A, Dolan H, Marano C, Zonderman AB, Troncoso JC, Zhou Y,

Wong DF, Ferrucci L, Egan J, Resnick SM, O'Brien RJ., 2013. Glucose intolerance, insulin resistance,

and pathological features of Alzheimer disease in the Baltimore Longitudinal Study of Aging. JAMA

Neurol 70, 1167-72.

[6] de la Monte, S. M., 2017. Insulin Resistance and Neurodegeneration: Progress Towards the

Development of New Therapeutics for Alzheimer's Disease. Drugs 77, 47-65.

[7] Ngandu T, Lehtisalo J, Solomon A, Levälahti E, Ahtiluoto S, Antikainen R, Bäckman L, Hänninen T, Jula

A, Laatikainen T, Lindström J, Mangialasche F, Paajanen T, Pajala S, Peltonen M, Rauramaa R,

Stigsdotter-Neely A, Strandberg T, Tuomilehto J, Soininen H, Kivipelto M., 2015. A 2 year multidomain

intervention of diet, exercise, cognitive training, and vascular risk monitoring versus control to

prevent cognitive decline in at-risk elderly people (FINGER): a randomised controlled trial. Lancet

385, 2255-63.

[8] Kemppainen N, Johansson J, Teuho J, Parkkola R, Joutsa J, Ngandu T, Solomon A, Stephen R, Liu Y,

Hänninen T, Paajanen T, Laatikainen T, Soininen H, Jula A, Rokka J, Rissanen E, Vahlberg T, Peltoniemi

J, Kivipelto M, Rinne JO.., 2018. Brain amyloid load and its associations with cognition and vascular

risk factors in FINGER study. Neurology 90, e206-e213.

[9] Kivipelto, M., Ngandu, T., Laatikainen, T., Winblad, B., Soininen, H., Tuomilehto, J., 2006. Risk score

for the prediction of dementia risk in 20 years among middle aged people: a longitudinal, population-

based study. Lancet Neurol 5, 735-41.

[10] Stephen R, Liu Y, Ngandu T, Rinne JO, Kemppainen N, Parkkola R, Laatikainen T, Paajanen T, Hänninen

T, Strandberg T, Antikainen R, Tuomilehto J, Keinänen Kiukaanniemi S, Vanninen R, Helisalmi S,

Levälahti E, Kivipelto M, Soininen H, Solomon A., 2017. Associations of CAIDE Dementia Risk Score

with MRI, PIB-PET measures, and cognition. J Alzheimers Dis 59, 695-705.

[11] Benjamini, Y., Hochberg, Y., 1995. Controlling the False Discovery Rate: A Practical and Powerful

Approach to Multiple Testing. J R Stat Soc Series B Stat Methodol 57, 289-300.

[12] Westwood S, Liu B, Baird AL, Anand S, Nevado-Holgado AJ, Newby D, Pikkarainen, M, Hallikainen M,

Kuusisto J, Streffer JR, Novak G, Blennow K, Andreasson U, Zetterberg H, Smith U, Laakso M, Soininen

H, Lovestone S. The influence of insulin resistance on cerebrospinal fluid and plasma biomarkers of

Alzheimer's pathology. Alzheimers Res Ther. 2017 Apr 26;9(1):31.

[13] Ahtiluoto S, Polvikoski T, Peltonen M, Solomon A, Tuomilehto J, Winblad B, Sulkava R, Kivipelto M.,

2010. Diabetes, Alzheimer disease, and vascular dementia: a population-based neuropathologic

study. Neurology 75, 1195-202.

[14] Martorana A, Sancesario GM, Esposito Z, Nuccetelli M, Sorge R, Formosa A, Dinallo V, Bernardi G,

Bernardini S, Sancesario G., 2012. Plasmin system of Alzheimer's disease patients: CSF analysis. J

Neural Transm (Vienna) 119, 763-9.

[15] Leung YY, Toledo JB, Nefedov A, Polikar R, Raghavan N, Xie SX, Farnum M, Schultz T, Baek Y, Deerlin

VV, Hu WT, Holtzman DM, Fagan AM, Perrin RJ, Grossman M, Soares HD, Kling MA, Mailman M,

Arnold SE, Narayan VA, Lee VM, Shaw LM, Baker D, Wittenberg GM, Trojanowski JQ, Wang LS., 2015.

Identifying amyloid pathology-related cerebrospinal fluid biomarkers for Alzheimer's disease in a

multicohort study. Alzheimers Dement (Amst) 1, 339-348.

[16] Meigs JB, Mittleman MA, Nathan DM, Tofler GH, Singer DE, Murphy-Sheehy PM, Lipinska I,

D'Agostino RB, Wilson PW. Hyperinsulinemia, hyperglycemia, and impaired hemostasis: the

Framingham Offspring Study. JAMA. 2000 Jan 12;283(2):221-8.

[17] Bi Oh, S., Suh, N., Kim, I., Lee, J.-Y., 2015. Impacts of aging and amyloid-β deposition on plasminogen

activators and plasminogen activator inhibitor-1 in the Tg2576 mouse model of Alzheimer's disease.

Brain Res 1597, 159-67.

[18] Cho, H., Joo, Y., Kim, S., Woo, R.-S., Lee, S. H., Kim, H.-S., 2013. Plasminogen activator inhibitor-1

promotes synaptogenesis and protects against Aβ1-42-induced neurotoxicity in primary cultured

hippocampal neurons. Int J Neurosci 123, 42-9.

[19] Miralbell J, Soriano JJ, Spulber G, López-Cancio E, Arenillas JF, Bargalló N, Galán A, Barrios MT,

Cáceres C, Alzamora MT, Pera G, Kivipelto M, Wahlund LO, Dávalos A, Mataró M. Structural brain

changes and cognition in relation to markers of vascular dysfunction. Neurobiol Aging. 2012

May;33(5):1003.e9-17.

Table 1. Baseline population characteristics by amyloid status.

Amyloid

negative

Amyloid

positive p-value

N 25 16

Sex (% female) 52.0% 50.0% 0.914

APOE ε4 carrier

(%) 12.0% 56.3% 0.003

Age 70.2 (5.6) 72.4 (3.0) 0.234

BMI 28.1 (3.8) 26.1 (2.8) 0.080

Diabetic (%) 12.0% 18.8% 0.570

Table 2. Associations of IR and DM2-related markers with amyloid status.

Mean concentration (SD) Logistic regression model

Marker

Amyloid

negative

Amyloid

positive p-value†

B‡

(95% confidence interval) p-value

Adjusted

p-value

C-Peptide (10^3 pg/ml) 1.31 (0.61) 0.95 (0.45) 0.056 -5.70 (-10.35 – -1.05) 0.016 0.072

Ghrelin (10^3 pg/ml) 1.57 (0.51) 1.55 (0.39) 0.936 +0.11 (-6.07 – 6.29) 0.972 0.998

GIP (10^3 pg/ml) 0.29 (0.12) 0.29 (0.16) 0.479 -1.53 (-5.39 – 2.32) 0.436 0.689

GLP-1 (10^3 pg/ml) 0.59 (0.11) 0.58 (0.08) 0.224 +0.01 (-8.79 – 8.81) 0.998 0.998

Glucagon (10^3 pg/ml) 1.07 (0.23) 1.00 (0.17) 0.530 -2.14 (-11.30 – 7.01) 0.646 0.808

Insulin (10^3 pg/ml) 0.27 (0.17) 0.17 (0.09) 0.036 -4.54 (-8.26 – -0.81) 0.017 0.072

Leptin (10^3 pg/ml) 7.55 (4.79) 6.06 (5.15) 0.145 -1.63 (-4.07 – 0.81) 0.191 0.569

PAI-1 (10^3 pg/ml) 5.31 (1.32) 4.16 (0.62) 0.003 -13.3 (-23.9 – -2.6) 0.015 0.072

Resistin (10^3 pg/ml) 2.22 (0.56) 2.03 (0.46) 0.316 -3.67 (-10.13 – 2.79) 0.266 0.569

Visfatin (10^3 pg/ml) 4.83 (2.25) 4.43 (2.30) 0.224 -2.03 (-6.77 – 2.71) 0.401 0.689

Adiponectin (10^6 pg/ml) 5.45 (3.44) 6.03 (3.82) 0.548 -0.25 (-2.30 – 1.79) 0.808 0.932

Adipsin (10^6 pg/ml) 1.21 (0.46) 1.45 (0.89) 0.925 +1.08 (-2.06 – 4.22) 0.500 0.689

fP-Glucose (mmol/l) 5.92 (0.88) 6.30 (1.17) 0.303 +4.80 (-9.33 – 18.93) 0.505 0.689

B-HbA1c (mmol/mol) 36.7 (4.1) 37.3 (4.6) 0.957 +15.0 (-10.9 – 40.9) 0.258 0.569

HOMA-IR (mmol·mU/L²) 2.06 (1.24) 1.33 (0.71) 0.060 -4.52 (-8.31 – -0.74) 0.019 0.072

†: Significance tested using the Mann-Whitney U test

‡: Coefficient of log-transformed values

Key: GIP Gastric inhibitory polypeptide; GLP-1 Glucagon-like peptide-1; PAI-1 Plasminogen activator inhibitor-1;

HbA1c Glycated hemoglobin

DIS

SE

RT

AT

ION

S | T

IMO

PE

KK

AL

A | M

UL

TIM

OD

AL

PR

ED

ICT

ION

OF

DE

ME

NT

IA A

ND

BR

AIN

PA

TH

OL

OG

Y | N

o 563

uef.fi



ISBN 978-952-61-3379-9ISSN 1798-5706



TIMO PEKKALA

MULTIMODAL PREDICTION OF DEMENTIA AND BRAIN PATHOLOGY

Dementia causes a considerable burden on individuals and societies, and interventions

at earlier stages should be developed. In this thesis dementia and related neuropathology are predicted in elderly cognitively healthy individuals in order to identify high-risk

individuals for interventions and to enrich targetable pathologies in trial populations. Also, the study investigates the associations

of markers of early type 2 diabetes and brain amyloid deposition, a hallmark of

Alzheimer’s disease.

TIMO PEKKALA

epublications.uef.fi · DISSERTATIONS | TIMO PEKKALA | MULTIMODAL PREDICTION OF DEMENTIA AND BRAIN...

Documents

Transcript of epublications.uef.fi · DISSERTATIONS | TIMO PEKKALA | MULTIMODAL PREDICTION OF DEMENTIA AND BRAIN...