人工智能论文20230415 2023PhD Princeton Probabilistic models for structured biomedical data.pdfVIP

Probabilisticmodelsforstructured

biomedicaldata

AndrewJones

ADissertation

PresentedtotheFaculty

ofPrincetonUniversity

inCandidacyfortheDegree

ofDoctorofPhilosophy

RecommendedforAcceptance

bytheDepartmentof

ComputerScience

Adviser:BarbaraE.Engelhardt

January2023

©CopyrightbyAndrewJones,2022.

AllRightsReserved

Abstract

Modernbiomedicaldatasets—frommolecularmeasurementsofgeneexpressiontopathologyimages—

holdpromisefordiscoveringnewtherapeuticsandprobingbasicquestionsaboutthebehaviorof

cells.Thoughtfulstatisticalmodelingofthesecomplex,high-dimensionaldataiscrucialtoelucidate

robustscientificfindings.Acommonassumptionindataanalysisthatthedatasamplesareinde-

pendentandidenticallydistributed.However,thisassumptionisnearlyalwaysviolatedinpractice.

Thisisespeciallytrueinthesettingofbiomedicaldata,whichoftenexhibitsomeamountofstruc-

ture,suchassubgroupsofpatients,cells,ortissuetypesorothercorrelationstructureamongthe

samples.

Inthisbodyofwork,Iproposedataanalysisandexperimentaldesignframeworkstoaccount

forseveraltypesofhighly-structuredbiomedicaldata.Theseapproaches,whichtaketheformof

Bayesianmodelsandassociatedinferencealgorithms,arespecificallytailoredfordatasetswithgroup

structure,multipledatamodalities,andspatialorganizationofsamples.

Inthefirstlineofwork,Iproposeamodelforcontrastivedim