ModelChain:DecentralizedPrivacy-PreservingHealthcarePredictiveModelingFrameworkonPrivateBlockchainNetworksTsung-TingKuo,PhD,1Chun-NanHsu,PhD,1andLucilaOhno-Machado,MD,PhD1,21HealthSystemDepartmentofBiomedicalInformatics,UniversityofCaliforniaSanDiego,LaJolla,CA2DivisionofHealthServicesResearch&Development,VASanDiegoHealthcareSystemAbstractCross-institutionalhealthcarepredictivemodelingcanaccelerateresearchandfacilitatequalityimprovementinitiatives,andthusisimportantfornationalhealthcaredeliverypriorities.Forexample,amodelthatpredictsriskofre-admissionforaparticularsetofpatientswillbemoregeneralizableifdevelopedwithdatafrommultipleinstitutions.Whileprivacy-protectingmethodstobuildpredictivemodelsexist,mostarebasedonacentralizedarchitecture,whichpresentssecurityandrobustnessvulnerabilitiessuchassingle-point-of-failure(andsingle-point-of-breach)andaccidentalormaliciousmodificationofrecords.Inthisarticle,wedescribeanewframework,ModelChain,toadaptBlockchaintechnologyforprivacy-preservingmachinelearning.Eachparticipatingsitecontributestomodelparameterestimationwithoutrevealinganypatienthealthinformation(i.e.,onlymodeldata,noobservation-leveldata,areexchangedacrossinstitutions).Weintegrateprivacy-preservingonlinemachinelearningwithaprivateblockchainnetwork,applytransactionmetadatatodisseminatepartialmodels,anddesignanewproof-of-informationalgorithmtodeterminetheorderoftheonlinelearningprocess.WealsodiscussthebenefitsandpotentialissuesofapplyingBlockchaintechnologytosolvetheprivacy-preservinghealthcarepredictivemodelingtaskandtoincreaseinteroperabilitybetweeninstitutions,tosupporttheNationwideInteroperabilityRoadmapandnationalhealthcaredeliveryprioritiessuchasPatient-CenteredOutcomesResearch(PCOR).IntroductionCross-institutioninteroperablehealthcarepredictivemodelingcanadvanceresearchandfacilitatequalityimprovementinitiatives,forexample,bygeneratingscientificevidenceforcomparativeeffectivenessresearch,1acceleratingbiomedicaldiscoveries,2andimprovingpatient-care.3Forexamp...
