Clinical Prediction Models A Practical Approach to Development Validation and Updating by Ewout W. Steye.pdf

Clinical Prediction Models A Practical Approach to Development Validation and Updating by Ewout W. Steye.pdf

《Clinical Prediction Models A Practical Approach to Development Validation and Updating by Ewout W. Steye.pdf》由会员分享，可在线阅读，更多相关《Clinical Prediction Models A Practical Approach to Development Validation and Updating by Ewout W. Steye.pdf（574页珍藏版）》请在冰豆网上搜索。

StatisticsforBiologyandHealthEwoutW.SteyerbergClinicalPredictionModelsAPracticalApproachtoDevelopment,Validation,andUpdatingSecondEditionStatisticsforBiologyandHealthSeriesEditorsMitchellGail,DivisionofCancerEpidemiologyandGenetics,NationalCancerInstitute,Rockville,MD,USAJonathanM.Samet,DepartmentofEpidemiology,SchoolofPublicHealth,JohnsHopkinsUniversity,Baltimore,MD,USAB.Singer,DepartmentofStatistics,UniversityofCaliforniaatBerkeley,Berkeley,CA,USAMoreinformationaboutthisseriesathttp:

/W.SteyerbergClinicalPredictionModelsAPracticalApproachtoDevelopment,Validation,andUpdatingSecondEdition123EwoutW.SteyerbergDepartmentofBiomedicalDataSciencesLeidenUniversityMedicalCenterLeiden,TheNetherlandsISSN1431-8776ISSN2197-5671（electronic）StatisticsforBiologyandHealthISBN978-3-030-16398-3ISBN978-3-030-16399-0（eBook）https:

/doi.org/10.1007/978-3-030-16399-01stedition:

SpringerScience+BusinessMedia,LLC20092ndedition:

SpringerNatureSwitzerlandAG2019Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpartofthematerialisconcerned,specificallytherightsoftranslation,reprinting,reuseofillustrations,recitation,broadcasting,reproductiononmicrofilmsorinanyotherphysicalway,andtransmissionorinformationstorageandretrieval,electronicadaptation,computersoftware,orbysimilarordissimilarmethodologynowknownorhereafterdeveloped.Theuseofgeneraldescriptivenames,registerednames,trademarks,servicemarks,etc.inthispublicationdoesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevantprotectivelawsandregulationsandthereforefreeforgeneraluse.Thepublisher,theauthorsandtheeditorsaresafetoassumethattheadviceandinformationinthisbookarebelievedtobetrueandaccurateatthedateofpublication.Neitherthepublishernortheauthorsortheeditorsgiveawarranty,expressedorimplied,withrespecttothematerialcontainedhereinorforanyerrorsoromissionsthatmayhavebeenmade.Thepublisherremainsneutralwithregardtojurisdictionalclaimsinpublishedmapsandinstitutionalaffiliations.ThisSpringerimprintispublishedbytheregisteredcompanySpringerNatureSwitzerlandAGTheregisteredcompanyaddressis:

Gewerbestrasse11,6330Cham,SwitzerlandForAleida,Matthijs,LaurensandSuzanneFormyfatherWimPrefaceThefirsteditionofthisbookwasmadeduringtheyears20052007.Sincethenquitesomenewdevelopmentshavetakenplace,bothinthegeneralscientificdirectionthatpredictionresearchistakingandspecifictechnicalinnovations.Thesedevelopmentshavebeenaddressedasfaraspossibleinthesecondedition.Manynewreferenceshavebeenadded.Somedetailedmaterialhasbeenmovedfromprinttotheweb.Manyfigureshavebeenredrawnincolorforbetterclarityandattrac-tiveness.Inall,manychangeshavebeenmadetonearlyeverychapter.Predictionmodelsareimportantinwidelydiversefields,includingmedicine,physics,engineering,meteorology,andfinance.Predictionmodelsarebecomingmorerelevantinthemedicalfieldwiththeincreaseinbiologicalknowledgeonpotentialpredictorsofoutcome,e.g.,from“omics”（includinggenomics,tran-scriptomics,proteomics,glycomics,metabolomics）.Also,theBigDataeraimplieswewillhaveincreasingaccesstolargevolumesofroutinelycollecteddata.Thenumberofapplicationsforpredictionmodelswillincrease,e.g.,withtargetedearlydetectionofdisease,andindividualizedapproachestodiagnostictestingandtreatment.Wearemovingtoaneraofpersonalizedevidence-basedmedicinethatasksforanindividualizedapproachtosharedmedicaldecision-making.Evidence-basedmedicinehasacentralplaceformeta-analysistosummarizeresultsfromran-domizedcontrolledtrials;predictionmodelssummarizetheeffectsofpredictorstoprovideindividualizedpredictionsoftheabsoluteriskofadiagnosticorprognosticoutcome.Predictionmodelsandrelatedalgorithmswillincreasinglyformthebasisforpersonalizedevidence-basedmedicineandindividualizeddecision-making.WhyReadThisBook?

Mymotivationforworkingonthefirstandsecondeditionsofthisbookstemsprimarilyfromthefactthatthedevelopmentandapplicationsofpredictionmodelsareoftensuboptimalinmedicalpublications.Withthisbook,Ihopetocontributetoviibetterunderstandingofrelevantissuesandgivepracticaladviceonbettermodelingstrategiesthanarenowadaysused.Issuesincludethefollowing:

（a）Betterpredictivemodelingissometimesreadilypossible,e.g.,alargedatasetwithhigh-qualitydataisavailable,butallcontinuouspredictorsaredichot-omized,whichisknowntohaveseveraldisadvantages.（b）Smallsamplesareused:

Studiesareunderpowered,implyingunreliableanswerstodifficultquestionssuchas“Whicharethemostimportantpredictorsinthispredictionproblem?

”Theproblemofsmallsamplesizeisaggravatedbydoingacompletecaseanalysiswhichdiscardsinformationfromnearlycompleterecords.Statisticalimputationmethodsarenowadaysavailabletoexploitallavail-ableinformation,especially“multipleimputations.”Predictorsareomittedthatshouldreasonablyhavebeenincludedbasedonsubjectmatterknowledge.Analystsrelytoomuchonthelimiteddatathattheyhaveavailableintheirdataset,insteadofwiselycombininginformationfromseveralsources,suchasmedicalliteratureandexpertsinthefield.Stepwiseselectionmethodsareabundantwhenresearchersapplyregressionmodeling,whilethesemethodsaresuboptimal,especiallyinsmalldatasets.Modelingapproachesareusedthatrequirehighernumbers.Data-hungrytechniques,suchasneuralnetworkmodeling,machinelearningorartificialintelligencetechniques,shouldnotbeusedinsmalldatasets.Noattemptsaremadetowardsvalidation,orvalidationisdoneinefficiently.Forexample,asplit-sampleapproachisfollowed,leadingtoasmallersampleformodeldevelopmentandasmallersampleformodelvalidation.Bettermethodsarenowadaysavailableandshouldbeusedfarmoreoften,specificallybootstrapresampling.（c）Claimsareexaggerated:

Often,weseestatementssuchas“theindependentpredictorswereidenti-fied”;inmanyinstances,suchfindingsarepurelyexploratoryandmaynotbereproducible;theymaylargelyrepresentnoise.Modelsarenotinternallyvalid,withoveroptimisticexpectationsofmodelperformanceinnewpatients.Onemodernmachinelearningmethodwithafancynameisclaimedasbeingsuperiortoamoretraditionalregressionapproach,whilenocon-vincingevidenceispresented,andasuboptimalmodelingstrategywasfollowedfortheregressionmodel.Faircomparisonsbetweenwell-usedstatisticalmethodsandmachinelearningmethodsarerequired.Researchersareinsufficientlyawareofoverfitting,implyingthattheirapparentfindingsaremerelycoincidental.viiiPreface（d）Poorgeneralizability:

Ifmodelsarenotinternallyvalid,wecannotexpectthemtogeneralizetonewpatients.Modelsaredevelopedforeachlocalsituation,discardingearlierfindingsoneffectsofpredictorsandearliermodels;aframeworkforcontinuousimprovementandupdatingofpredictionmodelsisrequired.Inthisbook,Isuggestmanysmallimprovementsinmodelingstrategies.Combined,theseimprovementsshouldleadtobetterdevelopment,validation,andupdatingofpredictionmodels.IntendedAudienceReadersshouldhaveabasicknowledgeofbiostatistics,especiallyregressionanalysis,butnostrongbackgroundinmathematicsisrequired.Thenumberofformulasisdeliberatelykeptsmall.Thefocusisonconceptsinpredictionresearch,whicharealsorelevanttocomputerscientistsanddatascientistsworkingonpre-dictioninthefieldofPredictiveAnalytics.Usually,abottom-upapproachisfollowedinteachingregressionanalysistechniques,startingwiththerequiredtypeofdata,modelassumptions,estimationmethods,andbasicinterpretation.Thisbookismoretop-down:

giventhatwewanttopredictanoutcome,howcanwebestutilizeregressionandrelatedtechniques?

Threelevelsofreadersareenvisioned:

（a）Thecoreintendedaudienceisformedbyepidemiologistsandappliedbio-statisticianswhowanttodevelop,validate,orupdateapredictionmodel.Bothstudentsandprofessionalsshouldfindpracticalguidanceinthisbook,espe-ciallybytheproposedsevenstepstodevelopavalidmodel（PartII）.（b）Thesecondgroupisformedbyclinicians,policy-makers,andhealthcareprofes-sionalswhowanttojudgeastudythatpresentsorvalidatesapredictionmodel.Thisbookshouldaidtheminacriticalappraisal,providingexplanationsoftermsandconceptsthatarecommoninpublicationsonpredictionmodels.Theyshouldtrytoreadchaptersofparticularinterest,orreadthemaintextofthechapters.Theycanskiptheexamplesandmoretechnicalsections（indicatedwith*）.（c）Thethirdgroupincludesmoretheoreticalresearchers,suchas（bio）statisticiansandcomputerscientists,whowanttoimprovethemethodsthatweuseinpredictionmodels.Theymayfindinspirationforfurthertheoreticalworkandsimulationstudiesinthisbook.Manyofthemethodsinpredictionmodelingarenotfullydevelopedyet,andcommonsenseorintuitionunderliessomeoftheproposedapproachesinthisbook.Improvementsarewelcome!

PrefaceixOtherSourcesManyexcellenttextbooksexistonregressionanalysistechniques,buttheseusuallydonothaveafocusonmodelingstrategiesforprediction.ThemainexceptionisFrankHarrellsbook“RegressionModelingStrategies”.Hebringsadvancedbio-statisticalconceptstopracticalapplication,supportedbythermspackageforR.Harrellsbookmay,however,betooadvancedforclinicalandepidemiologicalresearchers.ThisalsoholdsfortheHastie,Tibshirani,andFriedmanquitethoroughtextbook“TheElementsofStatisticalLearning”.Thesebooksareveryusefulforamorein-depthdiscussionofstatisticaltechni