基于SPM2动态因果模型操作练习

DynamicCausalModeling(DCM)APracticalPerspectiveOllieHulmeBarrieRoulstonZekilabDisclaimerThefollowingspeakershaveneverusedDCM.Anyimpressionofexpertiseorexperienceisentirelyaccidental.Structure•1.QuickrecaponwhatDCMcandoforyou.•2.WhattothinkaboutwhendesigningaDCMexperiment•3.HowtodoDCM.Whatbuttonstopressetc.ARe-capforDummiesYoucanaskdifferenttypesofquestionsaboutbrainprocessing.QuestionsofWhereQuestionsofHowFunctionalSpecializationisaquestionofWhere?•Whereinthebrainisacertaincognitive/perceptualattributeprocessed?•WhataretheRegionallyspecificeffects•yournormalSPManalysis(GLM)FunctionalIntegrationisaquestionofHOWExperimentallydesignedinputHowdoesthesystemwork?Whataretheinter-regionaleffects?Howdothecomponentsofthesysteminteractwitheachother?MODEL-FREEMODEL-DEPENDENTHypothesisdrivenDCM!Functionalconnectivity=thetemporalcorrelationbetweenspatiallyremoteareasEffectiveconnectivity=theinfluenceoneareaexertsoveranother2CategoriesofFunctionalintegrationanalysisPPIDCMoverview•DCMallowsyoumodelbrainactivityattheneuronallevel(whichisnotdirectlyaccessibleinfMRI)takingintoaccounttheanatomicalarchitectureofthesystemandtheinteractionswithinthatarchitectureunderdifferentconditionsofstimulusinputandcontext.•Themodelledneuronaldynamics(z)aretransformedintoarea-specificBOLDsignals(y)byahemodynamicforwardmodel(λ).TheaimofDCMistoestimateparametersattheneuronallevelsothatthemodelledBOLDsignalsaremostsimilartotheexperimentallymeasuredBOLDsignals.PlanningaDCM-compatiblestudy•Experimentaldesign:–preferablymulti-factorial(e.g.atleast2x2)StaticMovingNoattentAttent.1.SensoryinputfactorAtleastonefactorthatvariesthesensoryinput…changingthestimulus…aperturbationtothesystem2.ContextualfactorAtleastonefactorthatvariesthecontextinwhichtheperturbationoccurs.Oftenattentionalfactor,orchangeincognitivesetetc.PlanningaDCM-compatiblestudy•TRshouldbeasshortaspossible2seconds•PossiblecorrectionsforlongerTR’s1.slice-timing2.Restrictmodeltoproximateregions.Theclosertheyarealongzaxisthelowerthetemporaldiscrepancy12sliceacquisitionvisualinput•TimingproblemsinDCM:Duetothesequentialacquisitionofmultipleslicestherewillbetemporalshiftsbetweenregionaltimeserieswhichlieindifferentslices.Thiscausestimingmisspecification.AtshortTR’sthisisnottoomuchofaproblemsincetheinformationintheresponsevariableispredominantlycontainedintherelativeamplitudesandshapesofhemodynamicresponseratherthantheirtimings.ConsequentlyDCMisrobustagainsttimingerrorsupto1second•Hypothesisandmodel:–definespecificapriorihypotheses….–DCMisnotexploratory!Specifyyourhypothesesaspreciselyaspossible.Thisrequiresneurobiologicalexpertise(thefunpart)…readlotsofpapers!Lookforconvergentevidencefrommultiplemethodologiesanddisciplines.Anatomyisyourfriend.ParietalareasV5HypothesisAattentionmodulatesV5directlyV1HypothesisBAttentionmodulateseffectiveconnectivitybetweenPPCtoV5Definingyourhypothesis+Whenattendingtomotion…….+1.Whichparametersdoyouthinkaremostrelevant?Whichparametersrepresentmyhypothesis?WhicharethemostrelevantintrinsicanatomicalConnections?Whicharethemostrelevantchangesineffectiveconnectivity/connectionstrength?Whicharetherelevantsensoryinputs?2.Definingcriteriaforinference:single-subjectanalysis:Whatstatisticalthreshold?Whatcontrasts?groupanalysis:Which2nd-levelmodel?Pairedt-testforparameteraparameterb,One-samplet-test:parametera0rmANOVA(incaseofmultiplesessionspersubject)3.EnsurethatthemodelyougenerateisabletotestyourhypothesesThemodelshouldincorporateeverycomponentofthehypothesisParietalareasV5DirectinfluenceV1PulvinarIndirectinfluenceDCMcannotdistinguishbetweendirectandindirect!Hypothesesofthisnaturecannotbetested4.EvaluatewhetherDCMcanansweryourquestionCanDCMdistinguishbetweenyourhypotheses?Incaseof1.Specifyyourmainhypothesisanditscompetinghypothesesaspreciselyaspossibleusingconvergentevidencefromtheempiricalandtheoreticalliterature2.ThinkspecificallyabouthowyourexperimentwilltestthehypothesisandwhetherthehypothesisissuitableforDCMtotest.3.Klaasemphasisesthatyoushould‘Testyourmodelbeforeconductingtheexperimentusingsyntheticdata.Simulationisthekey!’4.DCMistricky,asktheexpertsduringthedesignstage.Theyareveryhelpful.ADCMin5easysteps…1.Specifythedesignmatrix2.DefinetheVOIs3.Enteryourchosenmodel4.Lookattheresults5.ComparemodelsSpecifydesignmatrix•NormalSPMregressors-nomotion,noattention-motion,noattention-nomotion,attention-motion,attention•DCManalysisregressors-nomotion(photic)-motion-attentionDefiningVOIs•Singlesubject:chooseco-ordinatesfromappropriatecontrast.e.g.V5frommotionvs.nomotion•RFX:DCMperformedat1stlevel,butdefinegroupmaximumforareaofinterest,theninsinglesubjectfindnearestlocalmaximumtothisusingthesamecontrastandaliberalthreshold(e.g.P0.05,uncorrected).DCMbutton‘specify’NB:inorder!Canselect:-effectsofeachcondition-intrinsicconnections-contrastofconnectionsOutputLatent(intrinsic)connectivity(A)Modulationofconnections(B)PhoticAttentionMotionInput(C)ComparingmodelsSeewhatmodelbestexplainsthedata,e.g.OriginalModelAttentionmodulatesV1toV5AlternativeModelAttentionmodulatesV5?DCMbutton