Health Care Applications Health Care Applications of Data Miningof Data Mining

Mollie R. Poynton PhD, APRNMollie R. Poynton PhD, APRNAssistant Professor, University of Utah College of NursingAssistant Professor, University of Utah College of Nursing

Adjunct Assistant Professor, University of Utah School of MediciAdjunct Assistant Professor, University of Utah School of Medicine, ne, Department of Biomedical InformaticsDepartment of Biomedical Informatics

February 8, 2007February 8, 2007

Where you could be if you werenWhere you could be if you weren’’t t herehere……

Standing in line at the Department of Motor Vehicles

Getting a Getting a ““well checkwell check”” from your pediatric nurse from your pediatric nurse practitionerpractitioner

Or maybeOr maybe……

Even betterEven better……

Giving birth

TodayToday’’s Agendas Agenda

Coping with Health Care DataCoping with Health Care DataData Warehousing & Data MiningData Warehousing & Data MiningApplications in Health CareApplications in Health CareExample: pharmacokinetic and Example: pharmacokinetic and pharmacodynamicpharmacodynamic modeling of modeling of remifentanilremifentanil

Large amounts of dataLarge amounts of data

11stst problem: storageproblem: storage–– Solution: storage now inexpensiveSolution: storage now inexpensive

Large amounts of dataLarge amounts of data

11stst problem: storageproblem: storage–– Solution: storage now inexpensiveSolution: storage now inexpensive

22ndnd problem: optimize data for analysisproblem: optimize data for analysis–– Solution: data warehousingSolution: data warehousing

Accessing data from applicationsAccessing data from applications

1970s 1970s –– mainframe/ not easily accessedmainframe/ not easily accessed1980s 1980s –– distributed database management distributed database management systems, still some difficulty pulling datasystems, still some difficulty pulling data1990s 1990s –– advent of enterprise data advent of enterprise data warehouses (EDW), stores of data warehouses (EDW), stores of data optimized for analyticsoptimized for analytics

What is a Data Warehouse?What is a Data Warehouse?

A collection of integrated subjectA collection of integrated subject--oriented oriented data extracted from the enterprise source data extracted from the enterprise source systemssystemsEach unit of data is relevant to some Each unit of data is relevant to some moment in timemoment in timeA data warehouse contains atomic data A data warehouse contains atomic data and lightly summarized dataand lightly summarized data

UHC Data Warehouse UHC Data Warehouse ArchitectureArchitecture

Data WarehouseIntegration

Engine

PeopleSoft

Billing

Radiology Reports

Clinical Documentation

Lab Results

Patient Demographics

ETL

Process

Epic Olympus/ Cerner

•Hybrid InmonArchitecture

•Oracle Backend

• Data stored at atomic level

• Long term data store

•Non-volatile -All updates are tracked

• Real-time clinical data

•Batch financial data

•Over 200 Contributor Systems

UHC Data Warehouse UHC Data Warehouse ArchitectureArchitecture

Data WarehouseIntegration

Engine

PeopleSoft

Billings

Radiology Reports

Clinical Documentation

Results

Patient Demographics

ETL

Process

Epic Olympus/Cerner

•Data marts are organized by subject area.

•Data marts are optimized to support reporting and analysis Visits

DMFinancial

DM

DM - A DM - B

GL DM

What data does the EDW What data does the EDW contain?contain?

Patient and Visit DataPatient and Visit DataOperational DataOperational DataClinical DataClinical DataFinancial DataFinancial DataResearch DataResearch Data

Which representsWhich represents……–– One Terabyte of dataOne Terabyte of data–– Over 500+ million recordsOver 500+ million records

Circumstance of Data ExplosionCircumstance of Data Explosion

Inexpensive storageInexpensive storageGrowth of data warehousing technologyGrowth of data warehousing technology““Data dumpingData dumping””Data volumes = terabytesData volumes = terabytes

Large amounts of dataLarge amounts of data

11stst problem: storageproblem: storage–– Solution: storage now inexpensiveSolution: storage now inexpensive

22ndnd problem: optimize data for analysisproblem: optimize data for analysis–– Solution: data warehousingSolution: data warehousing

33rdrd problem: extract useful information problem: extract useful information and knowledge from data storesand knowledge from data stores–– Solution: ???Solution: ???

Why is such large, complex data Why is such large, complex data problematic?problematic?

Not readily used by consumers of data/ Not readily used by consumers of data/ information/ knowledgeinformation/ knowledgeData quality issuesData quality issuesInIn--memory processing often not feasiblememory processing often not feasible–– Will this problem be solved soon?Will this problem be solved soon?Combinatorial explosion encounteredCombinatorial explosion encounteredTraditional statistical techniques may not Traditional statistical techniques may not be sufficiently flexible to model complex be sufficiently flexible to model complex patternspatterns

Data/ Knowledge DisparityData/ Knowledge Disparity

Molecular biology: millions of genotypes, but Molecular biology: millions of genotypes, but only thousands of associated phenotypesonly thousands of associated phenotypesPharmaceuticals: Enormous increase in Pharmaceuticals: Enormous increase in ““leadsleads””, but stunted productivity., but stunted productivity.Nursing: Warehousing of data from clinical Nursing: Warehousing of data from clinical information systems information systems –– improved nursing care?improved nursing care?Human medical data in North America and Human medical data in North America and Europe = thousands of terabytes Europe = thousands of terabytes –– how much how much has our clinical knowledge increased?has our clinical knowledge increased?

Large amounts of dataLarge amounts of data

11stst problem: storageproblem: storage–– Solution: storage now inexpensiveSolution: storage now inexpensive

22ndnd problem: optimize data for analysisproblem: optimize data for analysis–– Solution: data warehousingSolution: data warehousing

33rdrd problem: extract useful knowledge problem: extract useful knowledge from data storesfrom data stores–– Solution: knowledge discovery in databases/ Solution: knowledge discovery in databases/

data miningdata mining

Gaining Knowledge from Big DataGaining Knowledge from Big Data

Queries/ summariesQueries/ summariesOLAP (online analytic processing)OLAP (online analytic processing)VisualizationVisualizationStatistical methodsStatistical methodsMachine learning methodsMachine learning methods

The Knowledge Discovery in Databases (KDD) Process

From: Fayyad, Piatetsky-Shapiro, & Smyth, 1996

Why Data Mining/ KDD?Why Data Mining/ KDD?

Data explosion/ highly dimensional health Data explosion/ highly dimensional health care databasescare databasesGrowth of data warehousesGrowth of data warehousesNeed for technologies that accelerate Need for technologies that accelerate modeling of clinical knowledge.modeling of clinical knowledge.Exciting potential for discovering useful Exciting potential for discovering useful information and knowledge from existing information and knowledge from existing data storesdata stores

Data warehousing v. Data miningData warehousing v. Data mining

Data warehousing creates a single point of Data warehousing creates a single point of query, a single resource for data from query, a single resource for data from heterogeneous sources.heterogeneous sources.Data mining (aka knowledge discovery in Data mining (aka knowledge discovery in databases) extracts useful knowledge/ databases) extracts useful knowledge/ patterns from the data (predictive models, patterns from the data (predictive models, trends, etctrends, etc……))Data warehousing Data warehousing facilitatesfacilitates data mining.data mining.

Applications of Data Mining in Applications of Data Mining in Health CareHealth Care

Modeling Health OutcomesModeling Health OutcomesModeling clinical knowledge for decision Modeling clinical knowledge for decision support systemssupport systemsBioinformaticsBioinformaticsPharmaceutical researchPharmaceutical researchBusiness IntelligenceBusiness Intelligence

2000 2001 2002 2003 2004 2005 2006

Number of items

0

50

100

150

200

250

300

350

Year

Number of MEDLINE items

Number of items 82 118 135 183 253 290 343

2000 2001 2002 2003 2004 2005 2006

“data mining” OR “knowledge discovery”

Application: Hospital QI/QAApplication: Hospital QI/QA

Infection ControlInfection Control–– Surveillance systems developed using these Surveillance systems developed using these

techniques shown more sensitive and specific.techniques shown more sensitive and specific.¹¹Care Pathways/ GuidelinesCare Pathways/ Guidelines

–– Prompt use of clinical pathways for appropriate Prompt use of clinical pathways for appropriate patientspatients²²

Application: Predicting Patient Application: Predicting Patient OutcomesOutcomes

Predicting smoking cessation statusPredicting smoking cessation status3,43,4

Predicting prePredicting pre--term laborterm labor55

Predicting pneumonia for diagnostic Predicting pneumonia for diagnostic decision supportdecision support6,76,7

Predicting cancer prognosisPredicting cancer prognosis88

Application: Genomic MedicineApplication: Genomic Medicine

DataData--intensive area of researchintensive area of researchData mining methods are hypothesisData mining methods are hypothesis--generatinggeneratingGenerates Generates ““leadsleads”” or new directions for or new directions for pharmaceutical developmentpharmaceutical developmentDiscover relationships from published Discover relationships from published literatureliterature99

Other ApplicationsOther Applications

Pharmacokinetic and Pharmacokinetic and pharmacodynamicpharmacodynamicmodeling modeling –– RemifentanilRemifentanil1010

Decision Support for Pathology/ RadiologyDecision Support for Pathology/ Radiology–– Classifying tumors/ lesions, identifying Classifying tumors/ lesions, identifying

abnormalitiesabnormalities11,1211,12

Special characteristics of health care dataSpecial characteristics of health care data1313

1.1. HeterogeneityHeterogeneity2.2. Lack of canonical formLack of canonical form3.3. Poor mathematical characterizationPoor mathematical characterization4.4. Sensitive/ private nature of health care dataSensitive/ private nature of health care data5.5. High dimensionalityHigh dimensionality6.6. VolumeVolume

Heterogeneity of Data in Health CareHeterogeneity of Data in Health Care

Health care concepts often Health care concepts often lack canonical formlack canonical form

Concepts lack a single, preferred notation that Concepts lack a single, preferred notation that encompasses all equivalent conceptsencompasses all equivalent conceptsEx: chest pain, radiates to left armEx: chest pain, radiates to left arm

CP CP --> L arm> L armChest pressure with radiation to left armChest pressure with radiation to left armChest pressure radiating to armChest pressure radiating to arm*(We are making great progress here with *(We are making great progress here with

standardized standardized vocabsvocabs, , snomedsnomed, , NIC/NOC/NIC/NOC/NANDA,etcNANDA,etc……))

Poor mathematical characterizationPoor mathematical characterization

Concepts in healthcare lack formal Concepts in healthcare lack formal structure (like that found in basic structure (like that found in basic sciences).sciences).No standard representations that can be No standard representations that can be entered into formulas/ models (does mild entered into formulas/ models (does mild depression = 0.5 moderate depression?). depression = 0.5 moderate depression?). Significant need for canonical form Significant need for canonical form ––standardized vocabularies, so that we can standardized vocabularies, so that we can identify equivalent concepts.identify equivalent concepts.

Privacy IssuesPrivacy Issues

Both patient and provider!!Both patient and provider!!Follow HIPAA guidelinesFollow HIPAA guidelines……Theoretical possibility of reTheoretical possibility of re--identificationidentificationSensitive Sensitive ““predictionspredictions””Important to weight risks v. benefitsImportant to weight risks v. benefits

Population pharmacokinetic and Population pharmacokinetic and pharmacodynamicpharmacodynamic models of models of

remifentanilremifentanil in healthy volunteers in healthy volunteers using machine learning methodsusing machine learning methods

Our teamOur teamS. KangS. Kang, University of , University of InjeInje, School of Health , School of Health

AdministrationAdministrationM. PoyntonM. Poynton, Informatics Program, University of Utah , Informatics Program, University of Utah

College of Nursing and College of Nursing and University of Utah School of Medicine, Department of University of Utah School of Medicine, Department of

Biomedical InformaticsBiomedical InformaticsG. NohG. Noh, , AsanAsan Medical Center, University of Ulsan Medical Center, University of Ulsan

College of Medicine, College of Medicine, Clinical Pharmacology and Therapeutics, Anesthesiology Clinical Pharmacology and Therapeutics, Anesthesiology

and Pain Medicineand Pain MedicineK. KimK. Kim, , SanggyeSanggye Paik Hospital, Paik Hospital, InjeInje University College University College

of Medicine, of Medicine, Anesthesiology and Pain MedicineAnesthesiology and Pain MedicineH. LeeH. Lee, , MokdongMokdong Hospital, Hospital, EwhaEwha WomansWomans University University

College of Medicine, Anesthesiology and Pain College of Medicine, Anesthesiology and Pain MedicineMedicine

D. KimD. Kim, , DankookDankook University College of Medicine, University College of Medicine, Department of Anesthesiology and Pain MedicineDepartment of Anesthesiology and Pain MedicineK. K. BaeBae, , AsanAsan Medical Center, University of Ulsan College Medical Center, University of Ulsan College

of Medicine, of Medicine, Clinical Pharmacology and TherapeuticsClinical Pharmacology and TherapeuticsO. LinaresO. Linares, University of Utah, Bioengineering, University of Utah, BioengineeringS. KernS. Kern, University of Utah College of Pharmacy, , University of Utah College of Pharmacy, Pharmaceutics/Pharmaceutical Chemistry, Pharmaceutics/Pharmaceutical Chemistry,

AnesthesiologyAnesthesiology

Terms:Terms:

Pharmacokinetics (PK) v. Pharmacokinetics (PK) v. PharmacodynamicsPharmacodynamics (PD)(PD)

The process by which The process by which a drug is absorbed, a drug is absorbed, distributed, distributed, metabolized and metabolized and eliminated.eliminated.

The action or The action or effect(seffect(s) ) of a drug.of a drug.

Other termsOther terms

EEG (electroencephalogram)EEG (electroencephalogram)–– measures the electrical activity of the brain/ CNSmeasures the electrical activity of the brain/ CNS

CNS (central nervous system)CNS (central nervous system)–– brain and spinal cordbrain and spinal cord

ApENApEN (approximate entropy)(approximate entropy)–– a parameter measured by electroencephalography a parameter measured by electroencephalography –– can be used as a surrogate measure of CNS activitycan be used as a surrogate measure of CNS activity

ANN (artificial neural network)ANN (artificial neural network)NONMEMNONMEM–– Dominant software package used for pharmacokinetic Dominant software package used for pharmacokinetic

and and pharmacodynamicpharmacodynamic modeling modeling –– builds mixed builds mixed nonlinear effects modelsnonlinear effects models

remifentanilremifentanil

Purpose: anesthesia during surgical Purpose: anesthesia during surgical procedures (along with a muscle relaxant)procedures (along with a muscle relaxant)OpioidOpioid mumu--receptor agonist (similar to receptor agonist (similar to FentanylFentanyl))CNS depressantCNS depressantVery rapid achievement of peak effectVery rapid achievement of peak effectVery short duration of actionVery short duration of action

BackgroundBackground

Some basic investigation of machine Some basic investigation of machine learning methods for PK and PD modeling, learning methods for PK and PD modeling, but dominant method = mixed nonlinear but dominant method = mixed nonlinear effects model (NONMEM).effects model (NONMEM).14,1514,15

Noh data set excellent test bed for Noh data set excellent test bed for comparisoncomparison

An ideal data setAn ideal data set……

30 healthy volunteers infused with 30 healthy volunteers infused with 11--8 8 μμg/kg/min for g/kg/min for 20 minutes or continuous infusion 3 20 minutes or continuous infusion 3 μμg/kg/min.g/kg/min.Unusually broadUnusually broad spectrum of pharmacokinetic and spectrum of pharmacokinetic and pharmacodynamicpharmacodynamic parametersparametersIdeal data set for examining relative performance of Ideal data set for examining relative performance of machine learning methods and dominant method, machine learning methods and dominant method, mixed nonmixed non--linear effects model (NONMEM), for linear effects model (NONMEM), for predicting individual blood concentrationspredicting individual blood concentrations

ObjectivesObjectives

1.1. Compare accuracy of ANN and NONMEM in Compare accuracy of ANN and NONMEM in predicting predicting remifentanilremifentanil blood concentration.blood concentration.1010

2.2. Determine whether an artificial neural network Determine whether an artificial neural network can predict overshoot of can predict overshoot of ApEnApEn (approximate (approximate entropy) during recovery from profound entropy) during recovery from profound remifentanilremifentanil effect.effect.1010

3.3. Compare performance of multiple methods in Compare performance of multiple methods in predicting predicting remifentanilremifentanil blood concentration: blood concentration: NONMEM, ANN, SVM, ensemble methods.NONMEM, ANN, SVM, ensemble methods.1616

Background Background ObjObj. 2. 2

EEG effect/ CNS suppression of EEG effect/ CNS suppression of remifentanilremifentanil thought to be the same on thought to be the same on administration and recoveryadministration and recoveryNoh and colleagues observed an Noh and colleagues observed an ““overshootovershoot”” of CNS depression not of CNS depression not predicted by NONMEM.predicted by NONMEM.

Overshoot of EEG Overshoot of EEG ApEnApEn

EEG (electroencephalogram) DataEEG (electroencephalogram) Data

Abundant! (Measurements q 10Abundant! (Measurements q 10--20 20 seconds)seconds)24,509 measurements of 24,509 measurements of ApENApEN(approximate entropy)(approximate entropy)Does NONMEM require such a reduction of Does NONMEM require such a reduction of measurements that important information measurements that important information is lost?is lost?

ObjectivesObjectives

1.1. Compare accuracy of ANN and NONMEM in Compare accuracy of ANN and NONMEM in predicting predicting remifentanilremifentanil blood concentration.blood concentration.1010

2.2. Determine whether an artificial neural network Determine whether an artificial neural network can predict overshoot of can predict overshoot of ApEnApEn (approximate (approximate entropy) during recovery from profound entropy) during recovery from profound remifentanilremifentanil effect.effect.1010

3.3. Compare performance of multiple methods in Compare performance of multiple methods in predicting predicting remifentanilremifentanil blood concentration: blood concentration: NONMEM, ANN, SVM, ensemble methods.NONMEM, ANN, SVM, ensemble methods.1616

We hypothesized thatWe hypothesized that……

1.1. The ANN PK model would predict blood The ANN PK model would predict blood concentrations of concentrations of remifentanilremifentanil with with equivalent or greater accuracy than a equivalent or greater accuracy than a NONMEM PK model.NONMEM PK model.

2.2. An ANN PD model including all instances An ANN PD model including all instances of of ApENApEN (n = 24,509) with corresponding (n = 24,509) with corresponding predicted blood concentrations of predicted blood concentrations of remifentanilremifentanil would predict would predict ApEnApEnovershoot.overshoot.

Our strategyOur strategy

1.1. Build and compare PK models using Build and compare PK models using NONMEM & ANN.NONMEM & ANN.

2.2. If ANN as accurate as NONMEM, build a If ANN as accurate as NONMEM, build a PD model with ANN predictionsPD model with ANN predictions……

ANN Models: PK & PDANN Models: PK & PD

PK modelsPK models

PD Models: NONMEM v. ANNPD Models: NONMEM v. ANN

Predicting Predicting ApEnApEn overshoot with an ANNovershoot with an ANN

Conclusions from Conclusions from remifentanilremifentanil studystudy

1.1. Reduction of Reduction of pharmacodynamicpharmacodynamic data in order to make data in order to make PD modeling with NONMEM computationally feasible PD modeling with NONMEM computationally feasible resulted in inaccurate representation of REAL clinical resulted in inaccurate representation of REAL clinical response to response to remifentanilremifentanil..

2.2. An ANN PD model was able to predict An ANN PD model was able to predict ApENApEN overshoot overshoot during recovery from profound during recovery from profound remifentanilremifentanil effect.effect.

3.3. Overall, the predictive accuracy of the ANN PK model Overall, the predictive accuracy of the ANN PK model was better than that of the NONMEM PK model, but the was better than that of the NONMEM PK model, but the ANN PK model tended to ANN PK model tended to underpredictunderpredict the lower range the lower range of measured concentrations of of measured concentrations of remifentanilremifentanil..

ObjectivesObjectives

1.1. Compare performance of ANN and NONMEM in Compare performance of ANN and NONMEM in predicting predicting remifentanilremifentanil concentration, using concentration, using basic pharmacokinetic parameters.basic pharmacokinetic parameters.1010

2.2. Determine whether an artificial neural network Determine whether an artificial neural network can predict overshoot of can predict overshoot of ApEnApEn (approximate (approximate entropy).entropy).1010

3.3. Compare performance of multiple methods in Compare performance of multiple methods in predicting predicting remifentanilremifentanil concentration: concentration: NONMEM, ANN, SVM, ensemble methods.NONMEM, ANN, SVM, ensemble methods.1616

ReferencesReferences1.1. BrossetteBrossette SE, Sprague AP, Jones WT, Moser SA. A data mining system SE, Sprague AP, Jones WT, Moser SA. A data mining system for infection control surveillance. for infection control surveillance. Methods Methods InfInf Med. Med. Dec 2000;39(4Dec 2000;39(4--5):3035):303--310.310.2.2. LagorLagor C, C, AronskyAronsky D, D, FiszmanFiszman M, M, HaugHaug PJ. Automatic identification of PJ. Automatic identification of patients eligible for a pneumonia guideline: comparing the diagnpatients eligible for a pneumonia guideline: comparing the diagnostic ostic accuracy of two decision support models. accuracy of two decision support models. MedinfoMedinfo. . 2001;10(Pt 1):4932001;10(Pt 1):493--497.497.3.3. Poynton MR, McDaniel AM. Classification of smoking cessation staPoynton MR, McDaniel AM. Classification of smoking cessation status tus with a with a backpropagationbackpropagation neural network. neural network. J J BiomedBiomed Inform. Inform. Mar 24 2006.Mar 24 2006.4.4. Poynton M. Poynton M. Classification of smoking cessation status with machine Classification of smoking cessation status with machine learning methodslearning methods. Bloomington, IN: Graduate School, Indiana University; . Bloomington, IN: Graduate School, Indiana University; 2005.2005.5.5. Goodwin LK, Goodwin LK, IannacchioneIannacchione MA, Hammond WE, Crockett P, Maher S, MA, Hammond WE, Crockett P, Maher S, Schlitz K. Data mining methods find demographic predictors of prSchlitz K. Data mining methods find demographic predictors of preterm eterm birth. birth. NursNurs Res. Res. NovNov--Dec 2001;50(6):340Dec 2001;50(6):340--345.345.6.6. AronskyAronsky D, Chan KJ, D, Chan KJ, HaugHaug PJ. Evaluation of a computerized PJ. Evaluation of a computerized diagnostic decision support system for patients with pneumonia: diagnostic decision support system for patients with pneumonia: study study design considerations. design considerations. J Am Med Inform Assoc. J Am Med Inform Assoc. SepSep--Oct 2001;8(5):473Oct 2001;8(5):473--485.485.7.7. AronskyAronsky D, D, FiszmanFiszman M, Chapman WW, M, Chapman WW, HaugHaug PJ. Combining decision PJ. Combining decision support methodologies to diagnose pneumonia. support methodologies to diagnose pneumonia. Proc AMIA Proc AMIA SympSymp. . 2001:122001:12--16.16.8.8. QuarantaQuaranta V, Weaver AM, Cummings PT, Anderson AR. Mathematical V, Weaver AM, Cummings PT, Anderson AR. Mathematical modeling of cancer: the future of prognosis and treatment. modeling of cancer: the future of prognosis and treatment. ClinClin ChimChim ActaActa. . Jul 24 2005;357(2):173Jul 24 2005;357(2):173--179.179.9.9. McDonald DM, Chen, H., Su, H., Marshall, B.B. . Extracting gene McDonald DM, Chen, H., Su, H., Marshall, B.B. . Extracting gene pathway relations using a hybrid grammar: the Arizona Relation Ppathway relations using a hybrid grammar: the Arizona Relation Parser. . arser. . Bioinformatics. Bioinformatics. 2004;20(18):33702004;20(18):3370--3378.3378.

References (cont).References (cont).10.10. Kang S, Poynton M, Kim K, et al. Population pharmacokinetic and Kang S, Poynton M, Kim K, et al. Population pharmacokinetic and pharmacodynamicpharmacodynamic models of models of remifentanilremifentanil in Korean healthy volunteers in Korean healthy volunteers using NONMEM and neural network analysis. using NONMEM and neural network analysis. British Journal of Clinical British Journal of Clinical Pharmacology. Pharmacology. In press.In press.11.11. Wei L, Yang Y, Nishikawa RM, Jiang Y. A study on several machineWei L, Yang Y, Nishikawa RM, Jiang Y. A study on several machine--learning methods for classification of malignant and benign cluslearning methods for classification of malignant and benign clustered tered microcalcificationsmicrocalcifications. . IEEE Trans Med Imaging. IEEE Trans Med Imaging. Mar 2005;24(3):371Mar 2005;24(3):371--380.380.12.12. DreiseitlDreiseitl S, S, OhnoOhno--Machado L, Kittler H, Machado L, Kittler H, VinterboVinterbo S, S, BillhardtBillhardt H, Binder H, Binder M. A comparison of machine learning methods for the M. A comparison of machine learning methods for the diagnsosisdiagnsosis of of pigmented skin lesions. pigmented skin lesions. Journal of Biomedical Informatics. Journal of Biomedical Informatics. 2001;34:282001;34:28--36.36.13.13. CiosCios KJ, Moore GW. Uniqueness of medical data mining. KJ, Moore GW. Uniqueness of medical data mining. ArtifArtif IntellIntellMed. Med. SepSep--Oct 2002;26(1Oct 2002;26(1--2):12):1--24.24.14.14. DurisovaDurisova L, L, DedikDedik, L. New mathematical methods in pharmacokinetic , L. New mathematical methods in pharmacokinetic modeling. modeling. Basic & Clinical Pharmacology & Toxicology. Basic & Clinical Pharmacology & Toxicology. 2005;96:3352005;96:335--342.342.15.15. TolleTolle KM, Chen H, Chow H. Estimating drug/ plasma concentration KM, Chen H, Chow H. Estimating drug/ plasma concentration levels by applying neural networks to pharmacokinetic data sets.levels by applying neural networks to pharmacokinetic data sets. Decision Decision Support Systems. Support Systems. 2000;2000:1392000;2000:139--151.151.16.16. Poynton MR, Kang, S.H., Noh, G.J., Kim, K., Lee, H., Kim, D., Poynton MR, Kang, S.H., Noh, G.J., Kim, K., Lee, H., Kim, D., BaeBae, K., , K., Linares, O., Kern, S. Machine learning methods applied to pharmaLinares, O., Kern, S. Machine learning methods applied to pharmacokinetic cokinetic modeling: A multimodeling: A multi--method comparison. Under review.method comparison. Under review.

Informatics at University of UtahInformatics at University of Utah

University of Utah College of Nursing, University of Utah College of Nursing, Nursing Informatics Program Nursing Informatics Program http://www.nurs.utah.edu/programs/informatics/index.htmhttp://www.nurs.utah.edu/programs/informatics/index.htm

University of Utah School of Medicine, University of Utah School of Medicine, Department of Biomedical Informatics Department of Biomedical Informatics http://http://uuhsc.utah.edu/medinfouuhsc.utah.edu/medinfo//