University of Groningen

Towards tailored elderly carePeters, Lilian

DOI:10.1016/j.jpsychores.2013.02.003

IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite fromit. Please check the document version below.

Document VersionPublisher's PDF, also known as Version of record

Publication date:2015

Link to publication in University of Groningen/UMCG research database

Citation for published version (APA):Peters, L. (2015). Towards tailored elderly care: with self-assessment measures of frailty and casecomplexity. University of Groningen. https://doi.org/10.1016/j.jpsychores.2013.02.003

CopyrightOther than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of theauthor(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).

The publication may also be distributed here under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license.More information can be found on the University of Groningen website: https://www.rug.nl/library/open-access/self-archiving-pure/taverne-amendment.

Take-down policyIf you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediatelyand investigate your claim.

Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons thenumber of authors shown on this cover page is limited to 10 maximum.

Download date: 22-04-2022

77

Chap

ter 4

Prediction modeling and risk assessment using gender, morbidity and the Groningen Frailty Indicator in the oldest old population: the Leiden 85-plus Study

L.L. Peters, A.H. van Houwelingen, H. Boter, W.L. Zijlema, J. Gussekloo, A.J.M. de Craen, J.W. Blom, J.P.J. Slaets, E. Buskens, H. Burger, W.P.J. den Elzen

Submitted

78

ABSTRACT

Objectives Care planning requires valid (frailty) instruments to predict outcome in the oldest old population. The aim of this study is to evaluate assessment models based on demographics, morbidity and/or the Groningen Frailty Indicator (GFI).

Design Prospective follow-up study.

Study population Community-dwelling elderly persons aged 86-years who participated in the Leiden 85-plus Study.

Measurements Cox proportional hazards models were developed to assess the associations between predictors and mortality, hospitalization, and functional decline. Potential predictors were included in the models if p-value was ≤ 0.15. Final models were evaluated on overall performance, discriminative power (predictive accuracy on the individual level) and calibration (predictive accuracy at group level). Clinical risk assessment was based on the model with the best overall performance.

Results Prospective data of 391 elderly participants were analyzed. During the four years of follow-up, 122 persons died, 251 were hospitalized, and 352 experienced functional decline. Gender and morbidity appeared significant predictors of mortality and hospitalization. Morbidity remained the sole predictor for functional decline. The GFI-total score was associated with all poor outcomes. The model including the statistically significant predictors gender, morbidity and the GFI-total score showed best results of overall performance and calibration. However the discriminative power of this model remained poor. Using clinical risk assessment, the calculated risks at four years follow-up were 13-95% for mortality, 46-90% for hospitalization and 82-99% for functional decline. The risk scores derived showed accurate predictions during follow-up for all poor outcomes.

Conclusion Though the accuracy of the GFI for making predictions at individual elderly level appeared limited, the GFI is a valuable tool for risk assessment of poor outcomes at group levels in the oldest old population.

79

Chap

ter 4

4.1 INTRODUCTION

Life expectancy is increasing worldwide, but not all life years gained are free of disease or loss of functioning1. As a result the demand for (health) care for elderly is increasing exponentially. Thus the demographic transition challenges both healthcare systems and society2. To sustain health, older persons at risk of poor outcomes should be identified and selected for (cost-)effective preventive interventions3,4. The selection of high-risk older adults based on only chronological age disregards the fact that the speed of ageing shows considerable inter-individual variation5-7. Selection on morbidity is also not recommended as elderly persons differ in terms of severity of illness, functional status, prognosis and risk of adverse events even if they are diagnosed with the same disease or pattern of conditions8. Therefore, the concept of frailty has been introduced in primary and geriatric care as an indicator of risk of poor outcomes (e.g. mortality, institutionalization, hospitalization, and functional decline)9-12.

A valid screening tool to identify frail older adults may benefit clinical care and society. Next to clinical judgment such a tool may be a useful aid in tailoring preventive interventions to elderly people at increased risk of poor outcomes. From a societal perspective valid frailty measures facilitate more efficient planning and use of limited resources if (cost-)effective interventions are offered to high-risk people only. However, the concept of frailty has been debated since its introduction more than a decade ago. Various measures have been developed. Current instruments differ in perspectives on frailty, and therefore different domains are included to measure the construct13,14. Due to this difference in perspective, a systematic review found that the prevalence of frailty may appear to vary widely, i.e., from 4 to 59% in community dwelling elderly populations14. By and large, the concept frailty has been defined rather narrow comprising the physical domain, i.e., a deficit model consisting of an accumulation of impairments and conditions to create a Frailty Index15,16, or comprising a much broader phenotype associated with an underlying physiologic state of multisystem and energy dysregulation15,16.More recently, frailty has been denoted as a multidimensional construct comprising several domains (e.g. physical, psychological, social and cognitive domains)13,14,17,18.

The Groningen Frailty Indicator (GFI) includes the latter domains and the self-assessment version has shown good psychometric properties19-21. However, scarce evidence of the predictive validity is available on the GFI22-25. A one-year follow-up study showed that a community-dwelling population (mean age 77 years) with a GFI-score of four or higher had an increased risk of mortality, hospital admission, and developing functional decline22. Likewise, a five-year follow-up study including a general elderly

80

population of eleven European countries (mean age 65 years) corroborated the predictive value of the GFI on mortality23,24. In a clinical population the GFI predicted quality of life one year later25. However, the number of these follow-up studies is limited and the results might not be representative for the increasing population of the oldest old.

Moreover, some health professionals and researchers have suggested that the identification of frail elderly participants using an instrument is too time consuming and consequently unfeasible in primary care practices. Therefore, the main objective of the present study was to evaluate the predictive validity of five types of prediction models on mortality, hospitalization and functional decline in a four-year follow-up period in an oldest old population.

81

Chap

ter 4

4.2 METHODS

4.2.1 Study populationThe present study was embedded in the Leiden 85-plus Study; a population-based prospective follow-up study of inhabitants of the city of Leiden, the Netherlands. All inhabitants of the 1912–1914 birth cohorts were contacted in the month of their 85th birthday between September 1997 and September 1999 and were asked to participate. There were no further exclusion criteria. In total 705 elderly participants were eligible to participate, 14 subjects died prior to enrollment, and 92 subjects refused to participate. 599 elderly participants (87%) were included at baseline and were annually visited at their homes (Figure 1, flowchart). All participants gave written informed consent for the study including the use of data from their medical records for additional analysis, following explanation of the study requirements and assurance of confidentiality and anonymity26. The Medical Ethical Committee of the Leiden University Medical Center approved the study and the informed consent procedure (file number: P97/04).

For the present study elderly participants were included if they were 86 years and older and completed all items of the GFI.

4.2.2 Predictors At baseline, the following demographic variables were obtained through interviews; gender, living situation, and education level. In addition, morbidity was ascertained through general practitioners and pharmacy registries: e.g. myocardial infarction, cerebrovascular disease, diabetes mellitus, chronic obstructive pulmonary disorder, cancer, and arthritis.

At the time of recruitment, the GFI had not yet been developed. Therefore, for the present study we used items that were assessed in de Leiden 85-plus Study and served as proxies for the items of the self-report version of the GFI. The GFI comprises 15 items and assesses four domains of functioning: physical (mobility functions, general health, vision, hearing, weight loss, morbidity), cognition, social (emotional isolation), and psychological (emotional problems, feelings of anxiety; Appendix I). In addition, ‘weight loss’ was the calculated difference between the weights in kg, measured at age 85 and age 86. As weight difference was available at age 86 (instead of 85), the GFI items assessed at age 86 were used as the baseline in this study. The GFI-total score is the number of positive responses to questions about the presence (yes/no) of a problem or dependency. The items were summed to calculate the total score (range 0-15)6.

82

Figure 1 Flowchart

83

Chap

ter 4

4.2.3 Poor outcomesDuring follow-up, mortality data until December 2003 were obtained from the municipal registry. Data on hospitalization were annually provided by the general practitioners up to the age of 90 years. In the same follow-up period functional decline was annually measured using the Groningen Activity Restriction Scale (GARS). The GARS comprises 18 items and is a valid and reliable measure of (instrumental) activities of daily living27. GARS total scores range from 18 to 72 points, with higher scores indicating more dependency in activities. During the four year follow-up period the GARS was annually assessed, and an increase of four or more during this period was defined as functional decline28. The analyses of both functional decline and first hospital admission after participating in the study included death as an endpoint as well. A respondent was marked in the database as lost to follow-up if - for practical reasons- no data were collected about hospitalization or functional decline.

4.2.4 Statistical analysesBaseline characteristics were analyzed using descriptive statistics. Differences between participating and excluded older persons were evaluated with the Mann-Whitney test, Pearson Chi-Square test and Fisher’s Exact test, where appropriate. Median scores, including interquartile ranges (IQR), were calculated for the GFI-total score as the data were not normally distributed. Cox proportional hazards models were used to estimate the associations between the predictors and time to each poor outcome with censoring at lost to follow-up or at the end of follow-up. The associations were expressed as hazard ratios (HR). Time to poor outcome was defined as time to the midpoint of the year in which the outcome had occurred, except for mortality for which we used the exact date of death. The multivariate models included the GFI as a continuous score, which is preferred as dichotomization of variables reduces the power of the analyses29. Though, the separate GFI-items were entered as dichotomous variables, this corresponds with the scorings method of these items.

In the first prediction model gender and morbidity were entered as predictors. Though, age is a predictor that also could be easily to distract from a patient file, it could not be added to the model since all respondents had the same age in the present study. The second model included the GFI-total score as the sole predictor variable. Subsequently, a third model was constructed by extending model 2 with gender and the presence of morbidities as predictor variables using a forward selection method. We used a p-value of ≤ 0.15, to decide whether to retain potential predictors in the (multivariate) model. Arbitrarily, values for alpha greater than 0.05 are frequently used to drop variables

84

in prediction modeling to limit the bias in the predictor coefficients30,31. In the fourth ‘full’ model all separate GFI-items were entered as predictors in

the model. Next, this prediction model was adjusted with gender and morbidity predictors.

4.2.5 Model evaluationModels were evaluated in terms of overall performance, calibration, and discrimination.

4.2.6 Overall performanceThe prediction models were evaluated with -2log likelihood tests to compare the relative overall performance of nested models, i.e. gender and morbidity based models with and without GFI. Lower values on the -2log likelihood test indicated a better model fit.

4.2.7 Calibration Calibration is the extent of agreement between predicted risk and observed risk. This evaluation was performed graphically by plotting the observed risks against the predicted risk in deciles for each given follow-up period32. Ideally, the plot shows a 45○ gradient32. For survival data the observed risk is the Kaplan–Meier (cumulative) risk estimate over a given period, e.g. a one-year or two-year risk. To calculate the mean predicted absolute cumulative risks over a certain time of follow-up, the baseline survival function from the Cox model was estimated first. The baseline survival function S0(t) is the time dependent 1- cumulative risk of the outcome when the linear predictor (LP) is zero. The LP is defined as ß1*X1+ ß2*X2+ …, with the X denoting the predictor values (i.e. GFI-scores, gender, morbidity) and the ß’s representing the regression coefficients from the Cox regression. Because the survival function is given by S(t) = S0(t) exp (LP)

the baseline survival function [S0(t)] can be calculated as33:

S0(t) = S(t) 1/exp (LP)

By applying this formula to each participant we calculated the 1 year, 2 year, 3 year and 4 years cumulative baseline survival for each poor outcome. Using the LP of each participant, i.e. the part of the model representing the individual deviation from the baseline survival probability, we calculated the model based predicted 1, 2, 3, and 4 years risks as follows:

1- S(1 year) = S0(1year) exp (LP), 1- S(2 year) = S0(2 year) exp (LP), et cetera

85

Chap

ter 4

4.2.8 Discriminative validity Harrell’s concordance -statistic (c-statistic) was calculated to assess the overall discriminative power of the models. The c-statistic for censored data is an overall measure of discriminatory power comparable with the area under the ROC curve34,35. Its value is the probability that, given a random pair of subjects of whom one had the outcome event, the predicted survival time is larger for the participant who lived longer. The value has been classified as follows; fail (0.51-0.60), poor (0.61-0.70), fair (0.71-0.80), good (0.81-0.90) and excellent (0.91-1.00)35.

4.2.9 Adjustment for overfitting and construction of a practical formula for risk assessmentRegression coefficients and c-statistics were adjusted for overfitting by using bootstrapping techniques for each prediction model separately36. A linear shrinkage factor was calculated from the bootstrap results and applied to the regression coefficients36,37. The model’s predictive performance after correction for overfitting is more close to the performance that can be expected when the model is applied to future similar populations.

Next, a practical formula based on the adjusted linear predictor of the best performing model was constructed. This was done by dividing the ßs of the predictors by the smallest beta observed and rounding them to the nearest integer. These integers represented predictor loads. Subsequently a total linear predictor score for each subject was calculated by adding the predictor loads multiplied by the individual predictor values, e.g. one for male gender.

Statistical analyses were performed using SPSS Statistics 22.0 (SPSS inc. Chicago, Illinois) and bootstrapping using. S-plus 8.1 for Windows (Tibco Software Inc.).

86

4.3 RESULTS

Out of a total of 599 baseline participants, we included 391 elderly participants who were 86 years of age and who completed all GFI items. For 208 elderly adults the data could not be analyzed as 47 elderly persons died before they reached the age of 86 years, 39 declined to participate, and for 122 elderly subjects a GFI-score could not be calculated (Figure 1, flowchart). Compared with those excluded, participants were more often males, living independently, and had a higher level of education (p<0.05).

At baseline, 254 (65%) subjects were female and 49 (13%) participants lived in a home for the elderly or nursing home (Table 1). Of the population, two thirds (n=264) were diagnosed with at least one morbidity. The participants had a median GFI-score of 3 (IQR 1-5). During the four-years follow-up period 122 participants died, 251 participants were hospitalized and 352 experienced functional decline. The annually mortality rate was calculated with the Kaplan-Meier Survival tables, these rates were 5%, 9%, 12% and, 9%, respectively.

Table 1 Baseline characteristics of the study sample (n =391)

Sex, n (%)Female 254 (65)

Education, n (%)Primary school or lower 228 (58)Secondary school or higher 163 (42)

Living situation, n (%)Independent-living 232 (59)Assisted-living residences 110 (28)Home for the elderly/nursing home 49 (13)

Morbidity n (%)(Osteo-)arthritis 135 (35)Cancer 83 (22)Diabetes mellitus 52 (13)Chronic Obstructive Pulmonary Disease 44 (11)Myocardial infarction 39 (10)Cardiovascular accident 30 (8)

Assessments, median (IQR)1

Frailty (Groningen Frailty Indicator) 3 (1-5)(i)ADL2 (Groningen Activity Restriction Scale) 29 (23-37)

1 Interquartile range (IQR)2 (Instrumental) activities of daily living

87

Chap

ter 4

4.3.1 Overall performance The overall performances of all final prediction models are shown in Table 2. Gender and morbidity (model 1) appeared significant mutually independent predictors of mortality and hospitalization, whereas, morbidity was the only statistically significant predictor of functional decline. The GFI-total score (model 2) was statistically significantly associated with mortality, hospitalization, and functional decline (p≤0.11). Model 3 (including predictors gender, morbidity and GFI-total score) appeared to have superior overall performance compared to model 1 (i.e. gender and morbidity) and showed a statistically significant improvement of overall performance for mortality (p≤0.001) and functional decline (p=0.05), however not for hospitalization (p=0.12).

In the fourth model, all items of the GFI were entered as separate predictors and the items which were statistically significantly associated with the poor outcome (p≤0.15) are presented in Table 2. Item 8 (During the last 6 months have you lost a lot of weight unwillingly?) was a significant predictor for all poor outcomes (p< 0.07). Item 1 (are you able to carry out shopping independently?) remained a statistically significant predictor for mortality and functional decline (p<0.07). Another item of the physical domain, item 5 (Which mark do you give yourself for physical fitness?) was a statistically significant predictor for mortality and hospitalization (p<0.01). Item 14 (Have you recently felt downhearted or sad?) of the psychological domain was a statistically significant predictor for hospitalization. GFI item 4 (Are you able going to the toilet independently?), and item 10 (Do you have any complaints about your memory?) remained significant predictors for functional decline. Similar associations with poor outcomes were observed by adding gender and morbidity to the fourth model.

88

Table 2 Overall performance and hazard ratios of the univariable and multivariable Cox proportional hazard models to predict 4-years mortality, hospitalization, and functional decline

Mortality Hospitalization Functional Decline-2 log

likelihood HR p- value-2 log

likelihood HR p- value-2 log

likelihood HR p- valuePrediction model 1 1394 2796 3832

Gender1 1.96 ≤0.001 1.36 0.02 - - Morbidity2 1.57 0.03 1.39 0.02 1.34 0.04

Prediction model 2 1404 2614 3832

GFI-total score 1.02 ≤0.001 1.05 0.11 1.05 0.03

Prediction model 3 13823 26003 38263

GFI-total score 1.14 ≤0.001 1.04 0.11 1.05 0.04Gender1 2.11 ≤0.001 1.43 ≤0.001 1.18 0.12Morbidity2 1.44 0.07 1.36 0.02 1.20 0.09

Prediction model 4 1377 2589 3821GFI 1 1.33 0.07 1.03 0.81 1.19 0.07GFI 4 0.82 0.65 0.79 0.51 1.49 0.15GFI 5 1.57 ≤0.001 1.31 0.01 1.10 0.26GFI 8 2.02 ≤0.001 1.69 ≤0.001 1.35 0.07GFI 10 0.95 0.72 1.11 0.24 1.13 0.10GFI 14 0.85 0.32 0.80 0.04 0.94 0.49

Prediction model 5 1359 2577 3816GFI 1 1.34 0.08 0.98 0.90 1.18 0.11GFI 5 1.58 ≤0.001 1.36 ≤0.001 1.11 0.24GFI 8 2.03 ≤0.001 1.71 ≤0.001 1.31 0.11GFI 10 0.94 0.63 1.12 0.23 1.14 0.09GFI 14 0.86 0.35 0.78 0.03 0.94 0.48Gender1 1.72 ≤0.001 1.27 0.01 1.11 0.20Morbidity2 1.26 0.15 1.24 0.04 1.14 0.11

1 Gender: 0 is female and 1 is male2 Morbidity: 0 is no morbidity and 1 is at least one of the following diseases; myocardial infarction, cerebrovascular disease, diabetes mellitus, chronic obstructive pulmonary disorder, cancer, arthritis3 P-values for -2log likelihood tests in comparison with model 1, for mortality (p=0.001), functional decline (p=0.05), and hospitalization (p=0.12) respectively.

4.3.2 CalibrationThe model including GFI-total score, gender, and morbidity appeared the best calibrated model based on the greatest agreement between the predicted and observed probabilities for the prediction of poor outcomes, irrespective of follow-up period. Figures 2a-c show the calibration plots of 4 years predictions of poor outcomes. The calibration plots for 1, 2 and 3 years of follow-up showed similar results.

89

Chap

ter 4

Figure 2a Predicted probability mortality

Obs

erve

d ra

te

mor

talit

yO

bser

ved

rate

ho

spita

lizati

onO

bser

ved

rate

fu

nctio

nal d

eclin

e

Predicted probability hospitalization

Predicted probability functional decline

Figure 2a-c Calibration plots of the multivariable Cox proportional hazards model (GFI-total score, gender and morbidity) for the prediction of mortality (2a), hospitalization (2b), and functional decline (2c) at 4 years of follow up. The dotted line represents the line of identity, i.e. perfect calibration.

Figure 2b

Figure 2c

90

4.3.3 Discriminating validity The discriminating validity of all models of all outcomes was fail to poor with c-statistics ranging from 0.54-0.64 (Table 3).

Table 3 Discriminating validity of univariable and multivariable Cox proportional hazard models to predict 4-years mortality, hospitalization, and functional decline

Prediction models

Harrell’s c–statisticMortality

Harrell’s c–statistic

Hospitalization

Harrell’s c–statistic

Functional declineModel 1 Gender1 and/or morbidity2 0.60 0.56 0.54Model 2 GFI-total score 0.57 0.54 0.54Model 3 GFI-total score, gender1 and morbidity2 0.64 0.59 0.55Model 4 GFI separate items 0.59 0.57 0.54Model 5 GFI separate items, gender1 and morbidity2 0.62 0.59 0.561 Gender: 0 is female and 1 is male2 Morbidity: 0 is no morbidity and 1 is at least one of the following diseases; myocardial infarction, cerebrovascular disease, diabetes mellitus, chronic obstructive pulmonary disorder, cancer, arthritis

4.3.4 Overfitting adjustment and practical formulas for risk stratificationAfter correction for overfitting, the mean shrinkage factors for the prediction models 1 to 5 were: 1.00, 0.985, 0.945, 0.664 and 0.691, respectively. The prediction model including the GFI-total score, gender and morbidity was superior regarding calibration and discrimination. Therefore the prediction formulas were based on this model for each follow-up period (Table 4). The practical formulas for the linear predictor score were for mortality: 1*GFI+6*gender+3*morbidity; for hospitalization: 1*GFI+9*gender+8*morbidity; and, for functional decline: 1*GFI+4*gender+3*morbidity. These formulas represent the following: GFI-total score, gender (female [0] vs. male [1]); and morbidity (no chronic disorder [0] vs. ≥1 chronic disorder [1]). This indicates that, for example, males with a GFI-total score of 1 and no morbidity have a mortality formula score of 7. Figure 3a (grey line) shows their mortality risks of 5%, 12%, 21% and 29% at 1, 2, 3, and 4 years, respectively. In addition, males with a GFI-total score of 8 and one or more chronic disorders have a formula score of 17 and a mortality risk of 16%, 38%, 57%, and 70% at 1, 2, 3, and 4 years, respectively. In addition, stratified by follow-up period the association between the linear predictor scores and mortality, hospitalization and functional decline were plotted in graphs (Figures 3a-c).These figures can be used

91

Chap

ter 4

to stratify clinical risk scores for elderly participants in classes of four-year risks using the linear predictors’ formulas for the poor outcomes. These graphs provide for four years prediction the following risk ranges: mortality (13-95%), hospitalization (46-90%) and, functional decline (82-99%).

Table 4 Multivariable predictors of mortality, hospitalization and functional decline at 1,2,3 and 4 years follow-up

Prediction formula Predictor loadMortality1 1-CRm^e 0.13(GFI)+0.74(gender)+ 0.36(morbidity)-0.90

GFI-total score4 1Gender5 6Morbidity6 3

Hospitalization2 1- CRh^e 0.04(GFI)+0.36(gender)+ 0.31(morbidity)-0.47

GFI-total score4 1Gender5 9Morbidity6 8

Functional decline3 1- CRfd^e 0.06(GFI)+0.27(gender)+ 0.18(morbidity)-0.32

GFI-total score 4 1Gender5 4Morbidity6 3

1 Cumulative risk for the prediction of mortality (CRm) for 1, 2, 3 or 4 years were respectively: 0.95, 0.87, 0.79 and 0.712 Cumulative risk for the prediction of hospitalization (CRh) for 1, 2, 3 or 4 years were respectively: 0.79, 0.64, 0.49 and 0.373 Cumulative risk for the prediction of functional decline (CRfd) for 1, 2, 3 or 4 years were respectively: 0.47, 0.25, 0.15, and 0.094 Groningen Frailty Indicator total score, range 0-155 Gender: 0 is female and 1 is male6 Morbidity: 0 is no morbidity and 1 is at least one of the following diseases; myocardial infarction, cerebrovascular disease, diabetes mellitus, chronic obstructive pulmonary disorder, cancer, and arthritis

92

Figure 3a

Figure 3a-c Risk assessment for mortality (3a), hospitalization (3b) and functional decline (3c) over 1, 2, 3 and 4 years using a linear predictor score based on GFI-total score, gender and chronic diseases

Figure 3b

Figure 3c

Legends

1 year prediction

2 years prediction

3 years prediction

4 years prediction

Linear predictor funtional decline

Pred

icte

d ris

k fo

r fun

tiona

l dec

line

Linear predictor mortality

Pred

icte

d ris

k fo

r mor

talit

y

Pred

icte

d ris

k fo

r hos

pita

lizati

onLinear predictor hospitalization

93

Chap

ter 4

4.4 DISCUSSION

The aim of this study was to evaluate the predictive performance of five models that predict poor outcomes in the oldest old in the general population. Gender and morbidity (model 1) appeared significant predictors of mortality and hospitalization, whereas, morbidity was a sole predictor for functional decline. The GFI-total score (model 2) was statistically significantly associated with mortality, hospitalization, and functional decline. In previous studies the GFI-total score was also associated with mortality and functional decline, however not with hospitalization as these studies used a more conservative p-value of 0.05 to remain predictors in the model, instead of our applied p-value of 0.1522,23. The third model including the significant predictors GFI-total score, gender and morbidity showed the best overall performance. This best evaluated model predicted risk on a group level as the predicted risks of the outcomes agreed well with the observed risks. Though the GFI-total score contributed to the risk estimates, some other predictors like gender and morbidity were slightly stronger predictors of poor outcomes. Our study showed that male gender and having at least one disease were associated with higher risks of poor outcomes. Other studies found that age is an important predictor of poor outcomes. However, it may not be the best option, as the ageing process does not occur at the same pace among people of the same calendar age5-7. Since, all participants in the present study were of the same age it was not possible to take age into account, though age is a potentially relevant predictor it was not a sensible predictor in the current analyses. However, the predictors in the risk models showed more accurate predictions for mortality than the predicted age specific mortality rates from Statistics Netherlands, which are based on chronological age only38.

This is the first study that explored the contribution of the separate items of the GFI, with and without gender and morbidity variables, in the prediction of poor outcomes. The results showed that several items of the physical domain of the GFI were significant predictors for poor outcomes. Furthermore, one item of the cognitive domain remained a significant predictor of functional decline. In the prediction of hospitalization also one item of the psychological domain was a significant predictor. In contrast, items of the social domains did not substantially contribute to the prediction of mortality, hospitalization or functional decline. Our results are in contrast with findings from earlier research, which showed that psychological factors were predictive of mortality39. Our findings may add to the debate of frailty being a construct comprising physical functioning only versus frailty as a multidimensional construct13,14. Therefore, we recommend further studies to assess the contributions of the separate GFI items in

94

the prediction of poor outcomes with and without other characteristics, like age. A limitation of the present study is that we used a proxy of the GFI self- assessment

version because this instrument was not available at the start of the Leiden 85-plus Study. Four items of the GARS were included in the GFI proxy. These GARS items were also used to calculate functional decline. However, this procedure did not substantially affect the results since the GARS data selected to calculate the (baseline) GFI-score were different from those used to calculate functional decline (follow-up). Moreover, the relatively small study population increased the risk of overfitting. Therefore, we used bootstrapping techniques for over- optimism of the model’s performance in the derivation data. Bootstrapping techniques have shown superiority over other approaches to address these problems such as split-sample or cross validation methods40. The present results were found in relatively healthy 86 years old people since nursing home residents were less likely to participate. This might explain the lower yearly incidence mortality rates in the present study (5-12%), than rates observed in the general oldest old population in the Netherlands (14-19%)38. In addition, the data of the Leiden 85-plus Study were collected a decade ago, and possibly healthcare practice patterns regarding hospital admission have changed over time. Therefore, future studies are needed to demonstrate the external validity and usefulness of the prediction scores in current samples of the oldest old.

Our study has several strengths. We included an age group which is often excluded from clinical trials, preventive interventions and is sometimes undertreated41. Therefore, studying this population is valuable and can have implications for (preventive) treatment options in this age group. We developed models that can assist in stratifying groups of elderly persons according to their absolute 1 to 4 year risk of mortality, hospitalization and functional decline combining the GFI-total score, gender and the presence of morbidity. The predicted four-year risks of poor outcomes in this homogeneous oldest old population showed considerable variation, i.e., mortality ranged from 13 to 95%; hospitalization from 46 to 90% and functional decline from 82 to 99%. Given that the models are well calibrated, the large variance in model based risk predictions indicates that individual characteristics are potent determinants of risk and may be helpful for the purpose of stratification, in particular according to the risk of mortality.

Our models provide a better understanding of the risk of poor outcomes in the oldest old population. Group risk assessment with these models may enable healthcare management and authorities to allocate care more appropriately to the oldest old. For example, care as usual could be provided to groups of 86-year olds with low scores on the practical prediction formulas, while groups with higher scores might benefit from

95

Chap

ter 4

more extensive monitoring by healthcare professionals. However, for making individual predictions the GFI models seemed inadequate as the discriminative power and predictive accuracy of the models were not sufficient.

Regarding the discriminative power of the prediction models, GFI model 2 (GFI-total score) showed fail discriminative power for mortality. Whereas a European study showed slightly better results: the GFI-total score had poor to fair discriminative power to predict all-cause mortality between a two and five years follow-up period24. GFI model 3 (GFI-total score, gender, morbidity), and model 1 (gender and/or morbidity) both showed similar results on the c-statistics. However, it is unlikely that both models identify the same older participants since the GFI identifies older participants with dependencies on several domains (i.e. physical, cognitive, psychological and social), whereas morbidity only reflects the problems on a physical domain. Importantly, similar diagnostic accuracy does not imply similar clinical impact. The entire chain of diagnosis and care has to be assessed prior to making adequate policy advice42.

Though Model 1 can be easily implemented in clinical practice, this is not recommended as the identified elderly persons differ in terms of severity of illness, functional status, prognosis and risk of adverse events even if they are diagnosed with the same disease or pattern of conditions8. Nevertheless, the multivariate model including GFI, gender and, morbidity is beneficial for public health, hospitals, and policy makers as it can be used to identify groups of community-dwelling elderly persons who are at risk of poor outcomes instead of an individual risk. Accordingly, specific management strategies can be implemented to provide elderly care more appropriately (e.g. the allocation and provisions of hospital beds).

4.4.1 ConclusionIn conclusion, our study showed that the prediction model including gender, morbidity and GFI-total score is a valuable tool for risk assessment of poor outcomes within the next four years on a group level in the oldest old population. The discriminative power of the models was insufficient to allow individual predictions. The use of GFI-based risk models may however allow clinicians to standardize elderly long-care on a group level to manage their oldest old population more efficiently.

96

4.5 REFERENCES

1. Sole-Auro A, Guillen M, Crimmins EM. Health care usage among immigrants and native-born elderly populations in eleven european countries: Results from SHARE. Eur J Health Econ. 2012;13(6):741-754.

2. Bergman H, Ferrucci L, Guralnik J, et al. Frailty: An emerging research and clinical paradigm - issues and controversies. Journals of Gerontology - Series A Biological Sciences and Medical Sciences. 2007;62(7):731-737.

3. Chang JT, Morton SC, Rubenstein LZ, et al. Interventions for the prevention of falls in older adults: Systematic review and meta-analysis of randomised clinical trials. BMJ. 2004;328(7441):680.

4. Rockwood K, Mitnitski A, Song X, Steen B, Skoog I. Long-term risks of death and institutionalization of elderly people in relation to deficit accumulation at age 70. J Am Geriatr Soc. 2006;54(6):975-979.

5. Steverink N, Slaets J, Schuurmans H, Van Lis M. Measuring frailty: Developing and testing the GFI (groningen frailty indicator). Gerontologist. 2001;41(special issue I):236-237.

6. Schuurmans H, Steverink N, Lindenberg S, Frieswijk N, Slaets JP. Old or frail: What tells us more? J Gerontol A Biol Sci Med Sci. 2004;59(9):962-965.

7. Slaets JP. Vulnerability in the elderly: Frailty. Med Clin North Am. 2006;90(4):593-601.

8. Guiding principles for the care of older adults with multimorbidity: an approach for clinicians. Guiding principles for the care of older adults with multimorbidity: An approach for clinicians: American geriatrics society expert panel on the care of older adults with multimorbidity. J Am Geriatr Soc. 2012;60(10):E1-E25.

9. Fried LP, Tangen CM, Walston J, et al. Frailty in older adults: Evidence for a phenotype. J Gerontol A Biol Sci Med Sci. 2001;56(3):146-156.

10. Rockwood K, Stadnyk K, MacKnight C, McDowell I, Hebert R, Hogan DB. A brief clinical instrument to classify frailty in elderly people. Lancet. 1999;353(9148):205-206.

11. Woods NF, LaCroix AZ, Gray SL, et al. Frailty: Emergence and consequences in women aged 65 and older in the women’s health initiative observational study. J Am Geriatr Soc. 2005;53(8):1321-1330.

12. Mitnitski AB, Mogilner AJ, Rockwood K. Accumulation of deficits as a proxy measure of aging. ScientificWorldJournal. 2001;1:323-336.

13. de Vries NM, Staal JB, van Ravensberg CD, Hobbelen JSM, Olde Rikkert MGM, Nijhuis-van der Sanden MWG. Outcome instruments to measure frailty: A systematic review. Ageing Research Reviews. 2011;10(1):104-114.

14. Collard RM, Boter H, Schoevers RA, Oude Voshaar RC. Prevalence of frailty in community-dwelling older persons: A systematic review. J Am Geriatr Soc. 2012;60(8):1487-1492.

15. Fried LP, Ferrucci L, Darer J, Williamson JD, Anderson G. Untangling the concepts of disability, frailty, and comorbidity: Implications for improved targeting and care. J Gerontol A Biol Sci Med Sci. 2004;59(3):255-263.

16. Morley JE, Vellas B, van Kan GA, et al. Frailty consensus: A call to action. J Am Med Dir Assoc. 2013;14(6):392-397.

17. Walston J, Hadley EC, Ferrucci L, et al. Research agenda for frailty in older adults: Toward a better understanding of physiology and etiology: Summary from the american geriatrics Society/National institute on aging research conference on frailty in older adults. J Am Geriatr Soc. 2006;54(6):991-1001.

18. Gobbens RJJ, Luijkx KG, Wijnen-Sponselee MT, Schols JMGA. Towards an integral conceptual model of frailty. Journal of Nutrition, Health and Aging. 2010;14(3):175-181.

19. Peters LL, Boter H, Buskens E, Slaets JP. Measurement properties of the groningen frailty indicator in home-dwelling and institutionalized elderly people. J Am Med Dir Assoc. 2012;13(6):546-551.

20. Metzelthin SF, Daniels R, van Rossum E, de Witte L, van den Heuvel WJ, Kempen GI. The psychometric properties of three self-report screening instruments for identifying frail older people in the community. BMC Public Health. 2010;10:176-184.

97

Chap

ter 4

21. Bielderman A, van der Schans CP, van Lieshout MR, et al. Multidimensional structure of the groningen frailty indicator in community-dwelling older people. BMC Geriatr. 2013;13(1):86.

22. Daniels R, van Rossum E, Beurskens A, van den Heuvel W, de Witte L. The predictive validity of three self-report screening instruments for identifying frail older people in the community. BMC Public Health. 2012;12(1):69.

23. Theou O, Brothers TD, Pena FG, Mitnitski A, Rockwood K. Identifying common characteristics of frailty across seven scales. J Am Geriatr Soc. 2014(62):901-906.

24. Theou O, Brothers TD, Mitnitski A, Rockwood K. Operationalization of frailty using eight commonly used scales and comparison of their ability to predict all-cause mortality. J Am Geriatr Soc. 2013;61(9):1537-1551.

25. Jansen JC, Abma EM, Hegeman JH, Nieboer P, ten Duis HJ, Slaets JPJ. Kwetsbaarheid als indicator voor herstel na een fractuur. Ned Tijdschr Traum. 2011(3):51-57.

26. der Wiel AB, van Exel E, de Craen AJ, et al. A high response is not essential to prevent selection bias: Results from the leiden 85-plus study. J Clin Epidemiol. 2002;55(11):1119-1125.

27. Kempen GI, Miedema I, Ormel J, Molenaar W. The assessment of disability with the groningen activity restriction scale. conceptual framework and psychometric properties. Soc Sci Med. 1996;43(11):1601-1610.

28. Kempen GI, Ranchor AV, van Sonderen E, van Jaarsveld CH, Sanderman R. Risk and protective factors of different functional trajectories in older persons: Are these the same? J Gerontol B Psychol Sci Soc Sci. 2006;61(2):P95-101.

29. Royston P, Altman DG, Sauerbrei W. Dichotomizing continuous predictors in multiple regression: A bad idea. Stat Med. 2006;25(1):127-141.

30. Royston P, Moons KG, Altman DG, Vergouwe Y. Prognosis and prognostic research: Developing a prognostic model. BMJ. 2009;338:b604.

31. Steyerberg EW, Eijkemans MJ, Habbema JD. Stepwise selection in small data sets: A simulation study of bias in logistic regression analysis. J Clin Epidemiol. 1999;52(10):935-942.

32. Vergouwe Y, Steyerberg EW, Eijkemans MJ, Habbema JD. Validity of prognostic models: When is a model clinically useful? Semin Urol Oncol. 2002;20(2):96-107.

33. Spijker J, de Graaf R, Ormel J, Nolen WA, Grobbee DE, Burger H. The persistence of depression score. Acta Psychiatr Scand. 2006;114(6):411-416.

34. Gönen M, Heller G. Concordance probability and discriminatory power in proportional hazards regression. Biometrika. 2005;92:965-970.

35. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982;143(1):29-36.

36. Efron B, Tibshirani R. An introduction to the bootstrap. monographs on statistics and applied probability. New York: Chapman & Hall; 1993.

37. Steyerberg EW, Eijkemans MJ, Harrell FE,Jr, Habbema JD. Prognostic modelling with logistic regression analysis: A comparison of selection and estimation methods in small data sets. Stat Med. 2000;19(8):1059-1079.

38. Statistics Netherlands. www.cbs.nl. Accessed October 15th, 2014.

39. Puts MTE, Lips P, Ribbe M.W.’. The effect on frailty on residential/nursing home admission in the netherlands independent on chronic diseases and functional limitations. European Journal of Ageing. 2005(2):264-274.

40. Harrell FE,Jr, Lee KL, Mark DB. Multivariable prognostic models: Issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. 1996;15(4):361-387.

41. Habicht DW, Witham MD, McMurdo ME. The under-representation of older people in clinical trials: Barriers and potential solutions. J Nutr Health Aging. 2008;12(3):194-196.

42. Schaafsma JD, van der Graaf Y, Rinkel GJ, Buskens E. Decision analysis to complete diagnostic research by closing the gap between test characteristics and cost-effectiveness. J Clin Epidemiol. 2009;62(12):1248-1252.

98

Appendix I Groningen Frailty Indicator proxy items in the Leiden 85+ study

GFI Items Items derived from the Leiden 85+ studyPhysical domain1 Are you able to carry out these tasks single-handedly and without any help?

(the use of help resources such as walking stick, walking frame or wheelchair is considered to be independent)Shopping

1 Can you fully independently, do the shopping? (GARS item 18)

2 Walking around outside (around the house or the neighbours) 2 Can you fully independently walk outdoors? (GARS item 10)3 Dressing and undressing 3 Can you fully independently dress yourself? (GARS item 1)4 Going to the toilet 4 Can you fully independently get on and off the toilet? (GARS 6)5 What mark do you give yourself for physical fitness? (scale 0-10) 5 How do you judge your own general health?6 Do you experience problems in daily life because of poor vision? 6 Are you able to read the paper and are you able to recognize a face

on a distance of 4 meters (with glasses)? 7 Do you experience problems in daily life because of being hard of hearing? 7 Are you able to understand everything in a dialogue?8 During the last 6 months have you lost a lot of weight unwillingly?

(3 kg in 1 month or 6 kg in 2 months)8 [Difference between measured weight on the age of 85 years and 86 years]

9 Do you take 4 or more different types of medicine? 9 [Data from pharmacy registries]Cognitive domain10 Do you have any complaints about your memory? 10 Do you experience difficulties with your memory?Social domain11 Do you sometimes experience emptiness around yourself? 11 I experience a general sense of emptiness12 Do you sometimes miss people around yourself? 12 I miss having people around me13 Do you sometimes feel abandoned? 13 I often feel rejectedPsychological domain14 Have you recently felt downhearted or sad? 14 Have you recently experienced emotional problems, like anxiety,

depression, irritation or melancholy?15 Have you recently felt nervous or anxious? 15 Are you afraid that something bad will happen to you?

Scoring GFI Scoring items Leiden 85+ study

Items 1-4 Yes=0; No=1 Items 1-4 Yes, without trouble=0; Yes with some troubles=0; Yes with much trouble=0; No=1

Item 5 0-6=1; 7-10=0 Item 5 Excellent=0; very good=0; good=0; fair=1; poor=1;Items 6-9 No=0; Yes=1 Item 6-7 Yes=0; Yes, without trouble=0; Yes with some troubles=0;

Yes with much trouble=1; No=1Item 10 No=0; Sometimes=0; Yes=1 Item 10 Yes, but not problematic1; Yes problematic=1;

Yes very problematic=1; No=0Items 11-15 No=0; Sometimes=1; Yes=1 Item 11-13 Yes =1; No=0

Item 14 No not at all=0; Yes, a little bit=1; Yes, somewhat=1; Yes, quite a bit=1; Yes, very much=1

Item 15 No=0; Yes=1

99

Chap

ter 4

Appendix I Groningen Frailty Indicator proxy items in the Leiden 85+ study

GFI Items Items derived from the Leiden 85+ studyPhysical domain1 Are you able to carry out these tasks single-handedly and without any help?

(the use of help resources such as walking stick, walking frame or wheelchair is considered to be independent)Shopping

1 Can you fully independently, do the shopping? (GARS item 18)

2 Walking around outside (around the house or the neighbours) 2 Can you fully independently walk outdoors? (GARS item 10)3 Dressing and undressing 3 Can you fully independently dress yourself? (GARS item 1)4 Going to the toilet 4 Can you fully independently get on and off the toilet? (GARS 6)5 What mark do you give yourself for physical fitness? (scale 0-10) 5 How do you judge your own general health?6 Do you experience problems in daily life because of poor vision? 6 Are you able to read the paper and are you able to recognize a face

on a distance of 4 meters (with glasses)? 7 Do you experience problems in daily life because of being hard of hearing? 7 Are you able to understand everything in a dialogue?8 During the last 6 months have you lost a lot of weight unwillingly?

(3 kg in 1 month or 6 kg in 2 months)8 [Difference between measured weight on the age of 85 years and 86 years]

9 Do you take 4 or more different types of medicine? 9 [Data from pharmacy registries]Cognitive domain10 Do you have any complaints about your memory? 10 Do you experience difficulties with your memory?Social domain11 Do you sometimes experience emptiness around yourself? 11 I experience a general sense of emptiness12 Do you sometimes miss people around yourself? 12 I miss having people around me13 Do you sometimes feel abandoned? 13 I often feel rejectedPsychological domain14 Have you recently felt downhearted or sad? 14 Have you recently experienced emotional problems, like anxiety,

depression, irritation or melancholy?15 Have you recently felt nervous or anxious? 15 Are you afraid that something bad will happen to you?

Scoring GFI Scoring items Leiden 85+ study

Items 1-4 Yes=0; No=1 Items 1-4 Yes, without trouble=0; Yes with some troubles=0; Yes with much trouble=0; No=1

Item 5 0-6=1; 7-10=0 Item 5 Excellent=0; very good=0; good=0; fair=1; poor=1;Items 6-9 No=0; Yes=1 Item 6-7 Yes=0; Yes, without trouble=0; Yes with some troubles=0;

Yes with much trouble=1; No=1Item 10 No=0; Sometimes=0; Yes=1 Item 10 Yes, but not problematic1; Yes problematic=1;

Yes very problematic=1; No=0Items 11-15 No=0; Sometimes=1; Yes=1 Item 11-13 Yes =1; No=0

Item 14 No not at all=0; Yes, a little bit=1; Yes, somewhat=1; Yes, quite a bit=1; Yes, very much=1

Item 15 No=0; Yes=1

100