<%BANNER%>

Detecting effective connectivity in neural time series

Permanent Link: http://ufdc.ufl.edu/UFE0042776/00001

Material Information

Title: Detecting effective connectivity in neural time series
Physical Description: 1 online resource (132 p.)
Language: english
Creator: BOYKIN,ERIN RAE
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: EFFECTIVE -- GRANGER -- NEURAL
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Time series inference techniques for analyzing the directionality of interactions in dynamic systems have become popular tools in the neuroscience community for identifying the effective connectivity structure of brain networks. Commonly applied techniques include Granger causality (GC), partial directed coherence (PDC), and phase dynamics modeling (PDM). The goal of my work is to investigate the limitations of several of these time series inference techniques as they are used to analyze effective connectivity in neural time series. Specifically, I investigate the effects of differences in noise variance between time series and find that PDC and PDM are particularly vulnerable to such noise variance differences. I propose and demonstrate the use of a novel modified PDM technique, called generalized PDM (GPDM), which is similar to a previously described modified PDC technique, and is more robust to noise variance differences between time series. Next, I develop a method to assess the ability of various time series inference techniques to accurately determine the true effective connectivity of a given network of coupled, narrow-band neuronal oscillators. My method is based on decision tree classifiers, trained using several time series features that can be computed solely from experimentally recorded data. I show that the classifiers constructed in my study provide a framework for determining whether a particular time series inference technique is likely to produce incorrect results when applied to a given time series recorded from narrow-band oscillators. Finally, I consider the application of time series inference techniques to simulated data from neural mass models (NMMs). While popular time series inference techniques work well for a restrictive class of NMMs, I show that these techniques fail when applied to a more general NMM that captures entire regions of interest (ROI) in the brain. I propose a novel time series inference technique based on GPDM and empirical mode decomposition (EMD), which I call EMD-GPDM, to determine effective connectivity in such NMM-ROI systems and demonstrate its performance.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by ERIN RAE BOYKIN.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Khargonekar, Pramod.
Local: Co-adviser: Ogle, William.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2012-04-30

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042776:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042776/00001

Material Information

Title: Detecting effective connectivity in neural time series
Physical Description: 1 online resource (132 p.)
Language: english
Creator: BOYKIN,ERIN RAE
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: EFFECTIVE -- GRANGER -- NEURAL
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Time series inference techniques for analyzing the directionality of interactions in dynamic systems have become popular tools in the neuroscience community for identifying the effective connectivity structure of brain networks. Commonly applied techniques include Granger causality (GC), partial directed coherence (PDC), and phase dynamics modeling (PDM). The goal of my work is to investigate the limitations of several of these time series inference techniques as they are used to analyze effective connectivity in neural time series. Specifically, I investigate the effects of differences in noise variance between time series and find that PDC and PDM are particularly vulnerable to such noise variance differences. I propose and demonstrate the use of a novel modified PDM technique, called generalized PDM (GPDM), which is similar to a previously described modified PDC technique, and is more robust to noise variance differences between time series. Next, I develop a method to assess the ability of various time series inference techniques to accurately determine the true effective connectivity of a given network of coupled, narrow-band neuronal oscillators. My method is based on decision tree classifiers, trained using several time series features that can be computed solely from experimentally recorded data. I show that the classifiers constructed in my study provide a framework for determining whether a particular time series inference technique is likely to produce incorrect results when applied to a given time series recorded from narrow-band oscillators. Finally, I consider the application of time series inference techniques to simulated data from neural mass models (NMMs). While popular time series inference techniques work well for a restrictive class of NMMs, I show that these techniques fail when applied to a more general NMM that captures entire regions of interest (ROI) in the brain. I propose a novel time series inference technique based on GPDM and empirical mode decomposition (EMD), which I call EMD-GPDM, to determine effective connectivity in such NMM-ROI systems and demonstrate its performance.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by ERIN RAE BOYKIN.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Khargonekar, Pramod.
Local: Co-adviser: Ogle, William.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2012-04-30

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042776:00001


This item has the following downloads:


Full Text

PAGE 1

DETECTINGEFFECTIVECONNECTIVITYINNEURALTIMESERIESByERINR.BOYKINADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOLOFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENTOFTHEREQUIREMENTSFORTHEDEGREEOFDOCTOROFPHILOSOPHYUNIVERSITYOFFLORIDA2011

PAGE 2

c2011ErinR.Boykin 2

PAGE 3

TomyparentsforgivingmetheirloveandDNA 3

PAGE 4

ACKNOWLEDGMENTS Mysincerestthankstomycommitteemembersfortheirtimeandguidance.Inparticular,IthankDr.Ogleforgivingmethefreedomtopursuemyresearchinterestsandforrekindlingmycuriosityinandloveofscience,Dr.Talathi(Sachin)forhisunaggingenthusiasmwhilehelpingmepursuemyacademicgoalsandforencouragingmetoaimhighinmyefforts,Dr.Khargonekarforhisinsightfulfeedbackandforsharinghisincrediblebreadthofknowledge,andDr.Principeforwideningmyresearchperspectivesandforhavingcondenceinmyabilities.Iamhonoredandhumbledtohavetheopportunitytoworkwithagroupofsuchtalentedandesteemedprofessors.IamalsodeeplyindebtedtomyearlyacademicmentorsincludingDr.RobertHolt,Dr.DougJordan,Dr.JieHan,andDr.JoseFortes.Fortheirusefuldiscussionsandconstructivecriticisms,IthankDr.MingzhouDing,Dr.AlexCadotte,andDr.PaulCarney.IhavebeenfortunateenoughtohavethecompanyofanumberofcaringanddelightfullabmatesduringmyacademiccareerincludingXinFu,NanRattanatamrong,SelviKadirvel,GirishVenkatasubramanian,PhilBarish,MattVernon,YingXu,andShanLi.ManythankstomyfellowgraduatestudentcollaboratorsincludingSohanSeth,KateBehrmanandSashaMikheyev.Fortheirunparalleledfriendshipandnever-endingwordsofencouragement,IthankmyFISI's:Gabs,Jenn,Jess,Kelly,Kristine,Monica,andNicole.Iowethedeepestdebtofgratitudetomyparentswhohavesupportedmeemotionally,nancially,andeveryotherwayimaginable.Finally,wordscanhardlydojusticetotheloveandappreciationthatIhaveformycaringandbrillianthusband,Patrick.Hehastrulylledtheunforgivingminutewithsixtyseconds'worthofdistancerun. 4

PAGE 5

TABLEOFCONTENTS page ACKNOWLEDGMENTS .................................. 4 LISTOFTABLES ...................................... 8 LISTOFFIGURES ..................................... 9 ABSTRACT ......................................... 11 CHAPTER 1INTRODUCTION ................................... 13 1.1ConnectivityConcepts ............................. 13 1.2TimeSeriesInferenceTechniques ...................... 15 1.3Assumptions .................................. 16 1.4KeyContributions ................................ 17 2GENERALMETHODS ................................ 20 2.1AutoregressiveModeling ............................ 21 2.1.1VectorAutoregressiveModeling .................... 23 2.1.2SpectralPropertiesofVARModels .................. 24 2.2GrangerCausality ............................... 25 2.2.1PairwiseGrangerCausality ...................... 25 2.2.1.1TimedomainpairwiseGrangercausality ......... 25 2.2.1.2SpectraldomainpairwiseGrangercausality ........ 27 2.2.1.3Exampleapplications .................... 30 2.2.2ConditionalGrangerCausality ..................... 32 2.2.2.1TimedomainconditionalGrangercausality ........ 32 2.2.2.2SpectraldomainconditionalGrangercausality ...... 34 2.2.2.3PartitionmatrixGrangercausality ............. 36 2.2.2.4Exampleapplications .................... 38 2.3PartialDirectedCoherence .......................... 41 2.3.1GeneralizedPartialDirectedCoherence ............... 43 2.3.2ExampleApplications ......................... 44 2.4PhaseDynamicsModeling .......................... 44 2.4.1PairwisePhaseDynamicModeling .................. 46 2.4.2MultivariatePhaseDynamicModeling ................ 47 2.4.3ExampleApplications ......................... 48 2.5RecentAdvancesandApplications ...................... 49 2.6SignicanceTesting .............................. 50 5

PAGE 6

3CHARACTERIZATIONANDMITIGATIONOFTHEEFFECTSOFNOISEVARIANCEONTIMESERIESINFERENCETECHNIQUES ........... 63 3.1NoiseVarianceandLinearTimeSeriesInferenceTechniques ....... 63 3.1.1TheImpactofNoiseVarianceonLinearTimeSeriesInferenceTechniques ............................... 63 3.1.2MitigatingtheImpactofNoiseVarianceonLinearTimeSeriesInferenceTechniques .......................... 65 3.2NoiseVarianceandPhaseDynamicsModeling ............... 66 3.2.1TheImpactofNoiseVarianceonPhaseDynamicsModeling .... 66 3.2.2ANewTechniqueforMitigatingtheImpactofNoiseVarianceonPhaseDynamicsModeling ....................... 67 4DETECTINGEFFECTIVECONNECTIVITYINNETWORKSOFCOUPLEDNEURONALOSCILLATORS ............................ 74 4.1Neuron,SynapseandNetworkModels .................... 76 4.1.1NeuronModel .............................. 76 4.1.2SynapseModels ............................ 77 4.1.3NetworkModel ............................. 77 4.2ClassierConstructionMethodology ..................... 78 4.2.1DenitionofTimeSeriesFeatures .................. 78 4.2.2DecisionTreeConstruction ...................... 79 4.2.3ResultingDecisionTrees ........................ 82 4.2.4ExtensiontoLargerNetworks ..................... 86 4.3RecommendationsforDecisionTreeClassierUseinNetworksofCoupledNeuronalOscillators .............................. 87 5DETERMININGEFFECTIVECONNECTIVITYINNERUALMASSMODELNETWORKS ..................................... 96 5.1NeuralMassModels .............................. 100 5.2DeterminingEffectiveConnectivityinNetworksofNeuralMassModelPopulations ................................... 104 5.3DeterminingEffectiveConnectivityinNetworksofNeuralMassModelRegionsofInterest ............................... 106 5.3.1DeterminingEffectiveConnectivityViaGrangerCausality,PartialDirectedCoherenceandPhaseDynamicsModeling ........ 106 5.3.2EMD-GPDM:ANovelTechniqueforDeterminingEffectiveConnectivityinNetworksofNeuralMassModelRegionsofInterest ................................. 107 5.4EmpiricalModeDecomposition ........................ 109 6CONCLUSIONS ................................... 121 6

PAGE 7

REFERENCES ....................................... 123 BIOGRAPHICALSKETCH ................................ 132 7

PAGE 8

LISTOFTABLES Table page 4-1Testerrorofclassiers ................................ 90 4-2Decisiontreesensitivityandspecicityvalues ................... 90 4-3Likelihoodthattimeseriesinferencetechniquescorrectlyidentifycausality ... 90 5-1Neuralmassmodelbasalparametervaluesconsistentacrossallpopulationtypes ......................................... 113 5-2Neuralmassmodelbasalparametervaluesthatdifferamongpopulationtypes 113 5-3SensitivityandspecicityvaluesofGC,GPDC,andPDMwhenappliedtonetworksofthreecoupledNMMpopulations ................... 113 5-4SensitivityandspecicityvaluesofGC,GPDC,andPDMwhenappliedtonetworksoftwocoupledNMMROIs ........................ 113 5-5SensitivityandspecicityvaluesofEMD-GPDMandEMD-PDMwhenappliedtonetworksoftwocoupledNMMROIs ....................... 114 8

PAGE 9

LISTOFFIGURES Figure page 2-1Twoexampletimeseriesgeneratedfromstochastic,autoregressivemodels .. 52 2-2Spectraldomain,pairwiseGrangercausalityresultwhenappliedtostochastic,autoregressivemodels ................................ 53 2-3Twoexampletimeseriesgeneratedfromstochastic,vanderPoloscillators .. 54 2-4Spectraldomain,pairwiseGrangercausalityresultwhenappliedtostochastic,vanderPoloscillators ................................ 55 2-5Spectraldomain,conditionalGrangercausalityresultwithoutpartitionmatrixmethod ........................................ 56 2-6Spectraldomain,conditionalGrangercausalityresultwithpartitionmatrixmethodappliedtoARmodel ............................ 57 2-7Spectraldomain,conditionalGrangercausalityresultwithpartitionmatrixmethodappliedtovanderPoloscillators ..................... 58 2-8Partialdirectedcoherenceappliedtothreewhitenoiseprocesses ....... 59 2-9Generalizedpartialdirectedcoherenceappliedtothreewhitenoiseprocesses 60 2-10PartialdirectedcoherenceappliedtothreetimeseriesgeneratedfromstochasticvanderPoloscillators .......................... 61 2-11Generalizedpartialdirectedcoherenceappliedtothreetimeseriesgeneratedfromstochastic,vanderPoloscillators ....................... 62 3-1Probabilitythatpartialdirectedcoherenceidentiescausalityinasystem ... 69 3-2Probabilitythatgeneralizedpartialdirectedcoherenceidentiescausalityinasystem ........................................ 70 3-3ProbabilitythatGrangercausalityidentiescausalityinasystem ........ 71 3-4ResultsofthephasedynamicsmodelingtechniqueappliedtotwocoupledvanderPoloscillators ................................ 72 3-5ResultsofthegeneralizedphasedynamicsmodelingtechniqueappliedtotwocoupledvanderPoloscillators ......................... 73 4-1ThenetworkstructureofcoupledMorris-Lecarneuronsconsideredinmystudy 91 4-2Flowchartillustratingtheexperimentaldesignofmystudy ............ 92 4-3DecisiontreesconstructedforGrangercausalityacrossthreedifferentcouplingtypesconsidered ................................... 93 9

PAGE 10

4-4Compositedecisiontreesforalltimeseriesinferencetechniques ........ 94 4-5Thetrajectoriesoftwoinstantaneousphasetimeseriesfromtwodatasets ... 95 5-1Asystemdiagramofasinglepopulationneuralmassmodel ........... 114 5-2Powerspectraldensitiesofneuralmassmodelpopulations ........... 115 5-3Twoneuralmassmodelpopulationnetworks ................... 116 5-4Twocoupledneuralmassmodelregionsofinterest ................ 117 5-5Powerspectraldensityofasingleneuralmassmodelregionofinterest .... 117 5-6Anexampleneuralmassmodelpopulationnetwork ............... 118 5-7Anexampleneuralmassmodelregionofinterestsignalandtwoconstituentintrinsicmodefunctions ............................... 119 5-8Theempiricalmodedecomposition-generalizedpartialdirectedcoherencetechnique ....................................... 120 10

PAGE 11

AbstractofDissertationPresentedtotheGraduateSchooloftheUniversityofFloridainPartialFulllmentoftheRequirementsfortheDegreeofDoctorofPhilosophyDETECTINGEFFECTIVECONNECTIVITYINNEURALTIMESERIESByErinR.BoykinMay2011Chair:PramodP.KhargonekarMajor:ElectricalandComputerEngineeringTimeseriesinferencetechniquesforanalyzingthedirectionalityofinteractionsindynamicsystemshavebecomepopulartoolsintheneurosciencecommunityforidentifyingtheeffectiveconnectivitystructureofbrainnetworks.CommonlyappliedtechniquesincludeGrangercausality(GC),partialdirectedcoherence(PDC),andphasedynamicsmodeling(PDM).Thegoalofmyworkistoinvestigatethelimitationsofseveralofthesetimeseriesinferencetechniquesastheyareusedtoanalyzeeffectiveconnectivityinneuraltimeseries.Specically,IinvestigatetheeffectsofdifferencesinnoisevariancebetweentimeseriesandndthatPDCandPDMareparticularlyvulnerabletosuchnoisevariancedifferences.IproposeanddemonstratetheuseofanovelmodiedPDMtechnique,calledgeneralizedPDM(GPDM),whichissimilartoapreviouslydescribedmodiedPDCtechnique,andismorerobusttonoisevariancedifferencesbetweentimeseries.Next,Idevelopamethodtoassesstheabilityofvarioustimeseriesinferencetechniquestoaccuratelydeterminethetrueeffectiveconnectivityofagivennetworkofcoupled,narrow-bandneuronaloscillators.Mymethodisbasedondecisiontreeclassiers,trainedusingseveraltimeseriesfeaturesthatcanbecomputedsolelyfromexperimentallyrecordeddata.Ishowthattheclassiersconstructedinmystudyprovideaframeworkfordeterminingwhetheraparticulartimeseriesinferencetechniqueis 11

PAGE 12

likelytoproduceincorrectresultswhenappliedtoagiventimeseriesrecordedfromnarrow-bandoscillators.Finally,Iconsidertheapplicationoftimeseriesinferencetechniquestosimulateddatafromneuralmassmodels(NMMs).WhilepopulartimeseriesinferencetechniquesworkwellforarestrictiveclassofNMMs,IshowthatthesetechniquesfailwhenappliedtoamoregeneralNMMthatcapturesentireregionsofinterest(ROI)inthebrain.IproposeanoveltimeseriesinferencetechniquebasedonGPDMandempiricalmodedecomposition(EMD),whichIcallEMD-GPDM,todetermineeffectiveconnectivityinsuchNMM-ROIsystemsanddemonstrateitsperformance. 12

PAGE 13

CHAPTER1INTRODUCTIONThroughoutitshistory,thestudyofneurosciencehasfocusedextensivelyontheissueofneuralconnectivity.ThisthemecanbeobservedintheearlyworksofGustavFritschandEduardHitzigwhodemonstratedinthelate1800'sthatelectricalstimulationofdifferentareasofthecerebrumcausedinvoluntarycontractionsofspecicmusclesindogs( FritschandHitzig 1870 ).ItwasalsoastrongmotivatingfactorintheworkofWilderPeneld,aneurosurgeonwhostimulatedthebrainsofhisepilepticpatientswithelectricalprobesinordertoobservethepatients'responses.Peneld'sstudiesenabledhimtocreatemapsillustratingtheconnectionsbetweensensoryandmotorcorticesofthebrainandthevariouslimbsandorgansofthebody,whicharestillusedtoday( PeneldandJasper 1954 ).Morerecently,theHumanConnectomeProject,sponsoredbytheNationalInstitutesofHealth,hasspurredeffortstoconstructacomprehensivemapofneuralconnectionsinthehumanbrain,termedthehumanconnectome( Spornsetal. 2005 ).Traditionalmethodsofbrainresearch,suchasaxonaltracing,histologicaldissection,anddegenerationarenotidealforuseintheseefforts,however,duetotheirinvasivenature( Morecraftetal. 2009 ).Instead,researchersarefocusedprimarilyonusingnon-invasiveimagingandbrainrecordingtechnologiessuchasfunctionalmagneticresonanceimaging(fMRI),electroencephalograms(EEG),magnetoencephalograms(MEG),andpositronemissiontomography(PET)( Kandeletal. 2000 ).Withthesetechnologiescontributingtotherapidgrowthinrecordeddatafromthebrain,neuroscientistshaverecognizedtheneedforstatisticaltoolsthatcanuncoverbrainconnectivityfromtimeseriesofneuralactivity. 1.1ConnectivityConceptsInthepastdecade,manysuchstatisticaltools,referredtohereastimeseriesinferencetechniques,havebeendevelopedoradaptedforthispurpose.Timeseries 13

PAGE 14

inferencetechniquescanbeappliedtotimeseriesrecordingsofneuralactivityatmultiplescales,frompatchclamprecordingswhichrecordtheringofasingleneuron,toEEGrecordingswhichcapturetheelectriceldsgeneratedbypopulationsofneurons,tofMRIswhichmeasuresthehemodynamicresponseofthebraintolarge-scaleneuralactivity( Kandeletal. 2000 ).Ingeneral,timeseriesinferencetechniquesdonotrevealtheanatomicalconnectivityofneurons,cellpopulations,orbrainregions,termedstructuralconnectivity( Sporns 2010 ; Spornsetal. 2005 ).Rather,theyreectwhatisknownasthefunctionaloreffectiveconnectivityofthebrain.Functionalconnectivityisdenedasastatisticaldependencebetweenneuronalunits.Timeseriesinferencetechniquesthatmeasurecorrelation/covariance,spectralcoherenceorphaselockingbetweenpairsoftimeseriesarecommontoolsfordeterminingfunctionalconnectivity( Malsburg 1994 ).Thesemeasuresarehighlytimedependentandaremodulatedbyfactorssuchassensorystimulationandanorganism'sinternalstate.Also,becausetheyonlyconsiderstatisticalindependenceorthelacktherefore,theyarepurelyundirectedmeasuresanddonotmakeanyexplicitreferencetothephysical,anatomicalstructureofthebrainortothecausalrelationshipsamongneuronalunits( Friston 1994 ).Instead,suchdirected,causalrelationshipsarecapturedbyeffectiveconnectivity.Effectiveconnectivity,denedasthecausalinuenceoneneuronalunitsexertsoveranother,canbedeterminedthroughtimeseriesinferencetechniques( BaccalaandSameshima 2001 ; Granger 1969 ; Malsburg 1994 ; Schreiber 2000 ),statisticalmodeling( Fristonetal. 2003 ; Kline 2010 )orphysicalstimulationandablationexperiments( Massiminietal. 2005 ; McCulloch 1945 ).Likefunctionalconnectivity,effectiveconnectivityisdynamic,uctuatingwithchangesinexternalstimuliorinternalstate( Friston 1994 ; Sporns 2010 ).Acompleteandthoroughunderstandingofthebrainrequiresbothstructuralandfunctional/effectiveconnectivityinformation.Evenifthehumanconnectome,thecompletewiringdiagramofthebrain,isconstructed,therewillstillbeaneedfor 14

PAGE 15

temporalinformationregardingpathwaysthatarecriticalinlearninganddiseasesstatesaswellastheplasticityofthesepathwaysovertime.Sincethesequestionscannotbeaddressedwithstructuralinformationalone,thereisanevidentneedforfunctional/effectiveconnectivitydatatohelpcompleteourunderstandingofhowthebrainfunctionstemporally( Sporns 2010 ). 1.2TimeSeriesInferenceTechniquesWiththisneedinmind,timeseriesinferencetechniquesformeasuringfunctionalconnectivityhavebeenexploredextensivelyandappliedtomanyexperimentallyrecordedneuraldatasets( Allumetal. 1982 ; Buckneretal. 2009 ; Malsburg 1994 ; Rogersetal. 2007 ; Uddinetal. 2009 ; Valdes-Sosaetal. 2005 ).Timeseriesinferencetechniquesformeasuringeffectiveconnectivity,ontheotherhand,whilegrowinginpopularity,haveonlyrecentlygainedacceptanceasatechniqueforanalyzingneuraltimeseries.Infact,manyofthesemeasureshaveonlybeendevelopedorproposedforuseonneuraldatawithinthepasttwodecades( BaccalaandSameshima 2001 ; Baccalaetal. 1998 ; KaminskiandBlinowska 1991 ).Ofthethreebroadapproachestodeterminingeffectiveconnectivity(timeseriesinference,statisticalmodelingandphysicalstimulation/ablation)onlytimeseriesinferencetechniquesareconsidereddata-drivenanddonotrequireeitherapredenedanatomicalmodelofthesystemunderstudy,asstatisticalmodelingtechniquesrequire,orphysicalinterventioninthenervoussystem,asstimulation/ablationexperimentsrequire.Suchalowbarriertoentryhasmadethesedata-driventimeseriesinferencetechniquesverypopularfordeterminingeffectiveconnectivity.However,beingfairlynewtechniquesontheneurosciencescene,theirlimitationshavenotyetbeenfullyexplored.Forthisreason,mystudyfocusesonanalyzingseveraltimeseriesinferencetechniquesformeasuringeffectiveconnectivity.ThisclassoftechniquesencompassesawidevarietyofmethodsincludinglinearapproachessuchasGrangercausality( Geweke 1982 1984 ; Granger 1969 ),partialdirectedcoherence(PDC)( Baccalaand 15

PAGE 16

Sameshima 2001 ),andthedirectedtransferfunction(DTF)( KaminskiandBlinowska 1991 ),nonlinearapproachesdesignedfornarrow-bandoscillatorydatasuchasphasedynamicsmodeling(PDM)( RosenblumandPikovsky 2001 ),aswellasnonlinearapproachessuchaskernel-Grangercausality( Marinazzoetal. 2008 )andapproachesbasedoninformationtheorysuchastransferentropy( Schreiber 2000 )andconditionalindependence( SethandPrincipe 2010 ).Whileanargumenthasbeenmadethatthenonlinearnatureofneuraldatarequiresnonlineartechniquesformeasuringeffectiveconnectivity( Marinazzoetal. 2008 ; Schreiber 2000 ),lineartechniqueshavebeenwidelyappliedtoexperimentaldata( Brovellietal. 2004 ; Cadotteetal. 2010 ; Chenetal. 2006 ; Havliceketal. 2010 ; Liaoetal. 2010a ; Satoetal. 2009 )and,furthermore,havebeenshowncomputationallytoproducecorrectresultsevenonnonlineardata( Winterhalderetal. 2005 2007 ).Becauseofthis,andinlightoftheirpopularity,Iexaminetheapplicabilityofseverallinearandonesemi-nonlineartimeseriesinferencetechniqueformeasuringeffectiveconnectivity. 1.3AssumptionsAllofthetechniquesIconsiderherearebasedontheconceptoftemporalprecedenceandassuch,relyonseveralkeyassumptions.Therstconcernsthefactthatthesetechniquesoperateindiscretetime,whichmeansthatthesamplingofnaturallycontinuoussystemdynamicsmustoccuratsucharateastocapturethetimescalesoverwhichcausalinteractionsaremanifest( Florinetal. 2010 ; Granger 1969 ).Second,experimentallyrecordedtimeseriesshouldaccuratelypreservethetemporalrelationshipsoftheunderlyingsystem.Forthisreason,nonuniformdelaysintheresponseorrecordingofsystemvariablescanbedetrimentaltocorrectlydistinguishingcauseandeffect( Friston 2009a ; Sporns 2010 ).Finally,thesetechniquesassumethatthesetofvariablesmeasuredforanalysisisbothcompleteandnon-redundant,i.e.,allpotentiallycausalvariablesaremeasuredandnoneofthesevariablescontainduplicate,redundantinformation( Granger 1969 ; Seth 2005 ).Thissensitivitytounmeasured 16

PAGE 17

variables,alsocalledthehiddenunitproblem,isalimitationofvirtuallyalltimeseriesanalysistoolsandhasbeendiscussedatlengthelsewhere( Guoetal. 2008 ).Itisassumedinthefollowingthattheseassumptionsholdforthedataunderconsideration.Evenwiththepreviousassumptionsinplace,thetechniquesconsideredhereformeasuringeffectiveconnectivityhavebeenshowntoproduceincorrectresultswhenthetimeseriesbeinganalyzedarehighlysynchronized( Bezruchkoetal. 2003 ; Smirnovetal. 2007 )orhavedifferentnoisecharacteristics( Alboetal. 2004 ; Nalatoreetal. 2007 ; Winterhalderetal. 2005 ).Nevertheless,itisstandardpracticetoapplythesetechniquestotimeserieswithnoregardforwhetherthedatafallintooneoftheseproblematiccategories.Onereasonforthisisthatwhiletheselimitationshavebeenobservedintheory,littleworkhasbeendonetodenetheseconstraintsastheyapplytoexperimentaltimeseries.TheclosestexampleofsuchworkisthatdonebySmirnovetal.( Smirnovetal. 2007 )whocomparetheapplicationoftwotimeseriesinferencetechniques,PDCandPDM,toseveraldifferentneuronaloscillatormodels.TheirresultisalooseupperboundonthelevelofsynchronizationbetweentwocoupledneuronaloscillatorsabovewhichPDCandPDMgenerallygivesincorrectresults.Smirnovetal.'sstudywaslimitedinscope,however,sinceitonlydelineatedboundsonsynchronizationinthecontextofnarrow-bandneuraloscillators( Smirnovetal. 2007 ).Systematicstudiesdelineatingboundsnotonlyonsynchronizationbutothertimeseriesfeaturesinthecontextofvarioustypesofneuraltimeseriesisstilllacking. 1.4KeyContributionsItisthereforethegoalofmystudyto1)investigatethelimitationsofseveraltimeseriesinferencetechniquesastheyareusedtodetermineeffectiveconnectivityinavarietyofexperimentalneuraltimeseriesand2)proposenoveltechniquesormodicationstoexistingtechniquesthatovercometheobservedlimitations.Thespeciccontributionsofmyworkareasfollows: 17

PAGE 18

Contribution1:Iinvestigatetheeffectsofdifferencesinnoisevariancebetweentimeseriesonseveraltimeseriesinferencetechniques.Inlinewithpreviousndings,IdemonstratethatPDCishighlyaffectedbynoisevariancedifferences( Winterhalderetal. 2005 ).IalsoobservesimilarbehaviorfromthePDMtechnique,andingwhichhasnotbeenpreviouslydescribedintheliterature.AsamorerobustalternativetoPDM,IproposeamodiedPDMtechniquewhichItermgeneralizedPDM(GPDM)duetoitssimilaritytoamodiedPDCtechniqueknownasgeneralizedPDC(GPDC).Finally,usingasimulatedexample,IdemonstratethatGPDMisindeedlessaffectedbynoisevariancedifferencesbetweentimeseriesthanitspredecessor,PDM.Contribution2:Idevelopamethodtoassesstheabilityofvarioustimeseriesinferencetechniquestoaccuratelydeterminethetrueeffectiveconnectivityofagivennetworkofcoupled,narrow-bandneuronaloscillators.Theoutputsofsuchoscillatorsarerepresentativeofexperimentalrecordingsonthescaleofsingleneurons.Mymethodisbasedondecisiontreeclassierswhicharetrainedusingseveraltimeseriesfeaturesthatcanbecomputedsolelyfromexperimentallyrecordeddata.Ishowthattheclassiersconstructedinmystudyprovideaframeworkfordeterminingwhetheraparticulartimeseriesinferencetechniqueislikelytoproduceincorrectresultswhenappliedtoagiventimeseries.Asaresult,mymethodcanbeappliedbyuserswithoutanypriorknowledgeofnetworkstructureordynamicssinceallthetimeseriesfeaturesusedtodetermineapplicabilitycanbemeasureddirectlyfromthetimeseriesthemselves.Contribution3:Iinvestigatetheaccuracyofseveraltimeseriesinferencetechniquesastheyapplytoeffectiveconnectivitydeterminationinnetworksofneuralmassmodels(NMMs).TheseNMMsarerepresentativeofexperimentalrecordings,suchasEEGs,thataremadeatthescaleoflargeneuronalpopulations.Specically,IconsidertheaccuracyofGC,GPDCandPDMwhenappliedtoanNMMnetworkintroducedbyZavagliaetal.( Zavaglia 2008 ).Ibeginwithasimplecasewhereeach 18

PAGE 19

neuronalpopulationintheNMMnetworkexhibitsrhythmicactivityatasingledominantfrequency.UsingthisNMMnetwork,IshowthatGPDCisabletoaccuratelyrecoverthetrueconnectivityofthemodelwithhighprobability.Next,Iconsiderthecaseofinterconnectedneuronalregionsofinterest(ROIs),eachofwhichcontainsseveralNMMswithdifferentdominantrhythmicfrequencies( Ursinoetal. 2007 ; Zavagliaetal. 2006 ).ThismodelismorecomplexsinceeachROIexhibitsactivityoverarangeoffrequencyvalues.Inthiscase,Ishowthatnoneoftheaforementionedtimeseriesinferencetechniquesaccuratelyrecoversthetrueconnectivityofthemodel.InordertoaddresstheobservedinadequaciesofthetimeseriesinferencetechniquesthatIconsidered,IproposeanewtechniquebasedonGPDMandtheempiricalmodedecomposition(EMD)thatiscapableofaccuratelyrecoveringeffectiveconnectivityinanROI-basedNMM.Inthefollowing,IdescribetheapplicationofmyproposedtechniqueanddemonstrateitseffectivenessonROI-basedNMMs. 19

PAGE 20

CHAPTER2GENERALMETHODSTheprimarygoalofmyworkistoinvestigatethelimitationsofseveraltimeseriesinferencetechniquesastheyareusedtoanalyzeeffectiveconnectivityinavarietyofexperimentalneuraldatasets.Thetimeseriesinferencetechniquesconsideredhereincludethreelinearmethods:Grangercausality(GC),partialdirectedcoherence(PDC),andgeneralizedpartialdirectedcoherence(GPDC)aswellasonenonlinearmethod:phasedynamicsmodeling(PDM).Inthischapter,Idescribehowthesevarioustechniquesareappliedtotimeseries.Thisdescriptionincludesthemathematicalformulationsofthetechniques,examplesoftheiruse,andthedeterminationofsignicancelevels.Asmotivationfordescribingalltechniques,IbeginbyassumingthatIhavesimultaneouslyrecordedtimeseriesmeasurementsfromnnodes,whereeachnoderepresentstheobjectsofstudyandcanbeeitherasingleneuron,apopulationofneurons,oranentirebrainregiondependingonthescaleatwhichrecordinghastakenplace.ThesentimeseriesmeasurementscanberepresentedinvectorformasX(t)=[x1(t),x2(t),,xn(t)]T.Thegoalofapplyingatimeseriesinferencetechniquetothisdataistoestimatetheeffectiveconnectivitybetweenthesenodes,i.e.,determinewhethernodeiwithrecordedtimeseriesxi(t)causallyinuencesnodejwithrecordedtimeseriesxj(t).Thelineartimeseriesinferencetechniquesconsideredinthischapterareactuallyformulatedforthegeneralclassofstochasticprocesses.Tomaintaingenerality,IusethetermstochasticprocesswhendiscussingtheformulationofthesetechniquesandIusethetermtimeserieswhendiscussingtheapplicationofthesetechniquestoexperimentaltimeseriesdata.ThelineartechniquesthatIconsiderherearebasedontheconceptofcausalityoriginallyproposedbyNorbertWiener.Wienerreasonedthatifavariable,xi(t),is 20

PAGE 21

causallyinuencinganothervariablexj(t),thenthepredictionofxi(t)basedonpastvaluesofxj(t)andxi(t)willbemoreaccuratethanthatbasedonpastvaluesofxj(t)alone( Wiener 1956 ).ItwasnotuntilsomeyearslaterthatCliveGranger,workingintheeldofeconomics,rstformalizedthisconceptofcausalityinthecontextoflinearstochasticprocesses( Granger 1980 ).Asaresult,thetermWiener-GrangercausalityorsimplyGrangercausalityisnowusedtorefertothisconceptualizationofcausality.IusethetermcausalityheretorefertothisconceptofGrangercausality,whichisdistinguishedfromotherphilosophicaldenitionsofcausalitythathavebeenproposedintheliterature( Zellner 1979 ).Informalizinghistechnique,Grangerproposedusingautoregressive(AR)modelstoquantifytheimprovementgainedbyaddingxi(t)totheregressionofxj(t)( Granger 1980 ).FollowingGranger'soriginalformulation,threeofthefourtimeseriesinferencetechniquesconsideredhere,GC,PDC,andGPDC,requireX(t)torstbettoanARmodel.Thus,beforeIoutlinethemathematicalformulationsofthesetechniques,IbeginwithanoverviewofARmodelconstruction. 2.1AutoregressiveModelingAnautoregressive(AR)modelisalinearregressionformulathatisusedtopredictfuturevaluesofastochasticprocessbasedonpastvaluesoftheprocess.Consideringasinglestationary,zeromeanstochasticprocess,x(t),witht=1,,N,thebasicformulaofaunivariateARmodelofx(t)is( Boxetal. 2008 ; Hamilton 1994 ):x(t)=pXk=1a(k)x(t)]TJ /F5 11.955 Tf 11.95 0 Td[(k)+(t), (2)wherea(1),,a(p)arethemodelcoefcientsand(t)isawhitenoiseprocesswhichaccountsfortheresidualmodelerror.Themodelorder,denotedbyp,isindicativeofthenumberoftimelagsinthemodel.Forinstance,ifthepresentvalueofthestochasticprocess,x(t),dependsonitspastthreevalues,x(t)]TJ /F7 11.955 Tf 12.4 0 Td[(1),x(t)]TJ /F7 11.955 Tf 12.4 0 Td[(2),andx(t)]TJ /F7 11.955 Tf 12.4 0 Td[(3),thenp=3. 21

PAGE 22

FittinganappropriateARmodeltoastochasticprocessrequiresthespecicationofthemodelcoefcients,a(1),,a(p),thestandarddeviationofthenoiseprocess,denotedas,andthemodelorder,p.Valuesfora(1),,a(p)andcanbeobtainedbyrstmultiplyingbothsidesofEq.( 2 )byx(t)]TJ /F5 11.955 Tf 12.48 0 Td[(m)andtakingtheexpectedvalue.TheresultisthesetofYule-Walkerequations( Boxetal. 2008 ):(m)=pXk=1a(k)(m)]TJ /F5 11.955 Tf 11.95 0 Td[(k)+2m,0, (2)form0where(m)=E[x(t)x(t)]TJ /F5 11.955 Tf 12.62 0 Td[(m)]istheautocorrelationfunctionofx(t)andm,0istheKroneckerdeltafunctionwhichtakesthevaluezeroeverywhereexceptatm=0whereitisequaltoone.Sincem=0,,p,Eq.( 2 )yieldsatotalofp+1equationswhichcanbeusedtosolveforthepmodelcoefcientsplusthestandarddeviationofthenoiseprocess,.TheYule-Walkerestimatesofa(1),,a(p)andareobtainedbyreplacingthetheoreticalautocorrelations,(m),inEq.( 2 )byestimatedautocorrelations,^(m),whichcanbeobtainedfromtheobservedx(t).Determininganoptimalorder,p,foranARmodelrequiresbalancingthedesiretominimizemodelerrorwiththedesireformodelparsimony.Modelerror,reectedbythevarianceofthenoiseterm,monotonicallydecreasesasmodelorderincreases.Atthesametime,increasingthemodelorderwilleventuallyleadtoovertting,whichinturnleadstoinaccurateestimatesforthemodelcoefcients.Inordertoresolvethistrade-off,severalcriteriahavebeendevelopedformodelorderselectionwhichincorporatebothaminimizationofthenoisevarianceandapenaltyforlargemodelorders.OneexampleisAkaike'sinformationcriterion(AIC)whichisdenedas( Akaike 1974 ):AIC(p)=ln(2(p))+2p N. (2)AnotherisAkaike'snalpredictionerror(FPE),denedas( Akaike 1969 ):FPE(p)=2(p)N+p+1 N)]TJ /F5 11.955 Tf 11.95 0 Td[(p)]TJ /F7 11.955 Tf 11.95 0 Td[(1. (2) 22

PAGE 23

AthirdistheSchwarzBayesianinformationcriterion(BIC)( Schwarz 1978 ):BIC(p)=ln(2(p))+pln(N) N. (2)Alloftheseformulasareincreasingfunctionsof2(p)andincreasingfunctionsofp.Thevalueofpatwhicheachformulaisminimizedisconsideredtheoptimalmodelorderbasedontheparticularcriterioninquestion. 2.1.1VectorAutoregressiveModelingTheARformulapresentedinEq.( 2 )canalsobegeneralizedtocapturetheevolutionandinterdependenciesbetweenmultiplestochasticprocesses.Amodelinthisgeneralizedformisreferredtoasavectorautoregressive(VAR)model.AVARmodelofthestochasticprocessvector,X(t),isconstructedbyincludingforeachprocess,xi(t),anequationdescribingitsevolutionbasedonitsownpreviousvaluesalongwiththepreviousvaluesofallotherprocessesinthemodel.AssumingthateachstochasticprocessinX(t)iszeromeanandjointlystationary,theVARmodelofX(t)takesthefollowingform( Boxetal. 2008 ; Hamilton 1994 ; Lutkepohl 2005 ):X(t)=pXk=1A(k)X(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+(t). (2)ThisVARmodelisn-dimensionalandoforderp.Thepcoefcientmatrices,A(1),,A(p),areeachofdimensionnn.AssumingthatX(t)isadequatelyrepresentedbythisVARmodeloforderp,theterm(t),whichrepresentstheresidualmodelerror,isann1vectorofmutuallyuncorrelatedwhitenoiseprocesseswithcovariancematrix.AsintheunivariateARmodelcase,modelparametersaresolvedforbymultiplyingbothsidesofEq.( 2 )byXT(t)]TJ /F5 11.955 Tf 12.02 0 Td[(k)andtakingtheexpectedvalue.Theresultistheset 23

PAGE 24

ofYule-Walkermatrixequations:)]TJ /F7 11.955 Tf 6.94 0 Td[((m)=8>>><>>>:pPk=1A(k))]TJ /F7 11.955 Tf 6.94 0 Td[((m)]TJ /F5 11.955 Tf 11.96 0 Td[(k),ifm>0,pPk=1A(k))]TJ /F7 11.955 Tf 6.94 0 Td[(()]TJ /F5 11.955 Tf 9.29 0 Td[(k)+,ifm=0, (2)where)]TJ /F7 11.955 Tf 6.94 0 Td[((m)=E[X(t)XT(t)]TJ /F5 11.955 Tf 12.14 0 Td[(m)]istheautocorrelationfunctionofX(t).Thecoefcientmatrix,A(k),andthecovariancematrix,,ofthenoiseterm,areobtainedbysolvingEq.( 2 )viatheLevinson,WigginsandRobinson(LWR)algorithm( Morettin 1984 ; Whittle 1963 ).Modelorderisagainchosenbyminimizingoneofthefollowingmodelorderselectioncriterion( Lutkepohl 1985 ):AIC(p)=ln[det((p))]+2n2p N, (2)FPE(p)=det((p))N+np+1 N)]TJ /F5 11.955 Tf 11.96 0 Td[(np)]TJ /F7 11.955 Tf 11.96 0 Td[(1n, (2)BIC(p)=ln[det((p))]+n2pln(N) N, (2)wheredet()denotesdeterminant.Othermodelorderselectioncriterioncanalsobeusedforthispurpose( Lutkepohl 1985 ). 2.1.2SpectralPropertiesofVARModelsOnceaVARmodelisadequatelyestimatedforX(t)inthetimedomain,thismodelcanbeusedtoderivepropertiesofX(t)inthespectraldomain.Spectralfeaturesarederivedfromthecoefcientmatrix,A(k),andthenoisecovariancematrix,,oftheVARmodel.Specically,( Gardner 1988 ):A(f)=I)]TJ /F9 7.97 Tf 18.27 15.21 Td[(pXk=1A(k)e)]TJ /F8 7.97 Tf 6.59 0 Td[(2jfk (2) 24

PAGE 25

whereIisthen-dimensionalidentitymatrix,j=p )]TJ /F7 11.955 Tf 9.3 0 Td[(1andfisnormalizedfrequencyintheinterval[-0.5,0.5].Thetransformedcoefcientmatrix,A(f),isannnmatrixthatisnowafunctionoffrequency.Fromthetransformedcoefcientmatrixandthenoisecovariancematrix,,oftheVARmodel,onecanalsoobtainthetransfermatrix,denotedH(f),andthespectralmatrix,S(f),ofX(t).Thetransferfunctionmatrixwithnoise(t)asinputandX(t)asoutputisdenedas( Gardner 1988 ):H(f)=A)]TJ /F8 7.97 Tf 6.59 0 Td[(1(f), (2)whereA(f)isdenedinEq.( 2 ).Thespectralmatrixcanbeobtainedusingthefollowingformula( Boxetal. 2008 ):S(f)=H(f)H(f), (2)where*denotescomplexconjugatetranspose.Theserelationshipsformthebasisforspectraldomaintimeseriesinferencetechniques. 2.2GrangerCausality 2.2.1PairwiseGrangerCausalityPairwiseGrangercausality(GC)isatimeseriesinferencetechniquethatcanbeusedtomeasuretheeffectiveconnectivitybetweentwostochasticprocesses,denotedasxi(t)andxj(t).Thistechnique,sinceitisbasedonlinearARmodelsisitselflinearandparametric. 2.2.1.1TimedomainpairwiseGrangercausalityAssumingthatxi(t)andxj(t)arezeromean,individuallyandjointlystationarystochasticprocesses,univariateARmodelscanbettoeachprocesssuchthat:xi(t)=piXk=1a(k)xi(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+i,1(t), (2) 25

PAGE 26

xj(t)=pjXk=1b(k)xj(t)]TJ /F5 11.955 Tf 11.95 0 Td[(k)+j,1(t), (2)wherevar(i,1(t))=2iandvar(j,1(t))=2j.SettingX(t)=[xi(t),xj(t)]T,abivariateVARmodelcanalsobettoX(t)suchthat:X(t)=pXk=1A(k)X(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+(t), (2)where(t)=[i,2(t),j,2(t)]T.Thecovariancematrixof(t),denotedby,canbeexpandedas:=0B@iiijjijj1CA, (2)whereii=var(i,2),jj=var(j,2),andij=ji=cov(i,2,j,2).Ifxi(t)andxj(t)areindependent,ij=0,ii=2i,andjj=2j.FromEq.( 2 ),thevalueof2i(t)measurestheaccuracyoftheARpredictionofxi(t)basedonitsownpreviousvalues,whilethevalueofiirepresentstheaccuracyoftheVARpredictionofxi(t)basedonthepreviousvaluesofbothxi(t)andxj(t).AccordingtoGranger,thiscomparisongivesrisetothefollowingquanticationofcausalityfromxj(t)toxi(t)( Dingetal. 2006 ; Granger 1980 ):gcj!i=ln2i ii, (2)andsimilarlyfromxi(t)toxj(t):gci!j=ln2j jj. (2)Ifgcj!i=0,thenthereisnocausalityfromxj(t)toxi(t),whileifgcj!i>0,thencausalityfromxj(t)toxi(t)issaidtoexist.Thesameholdstrueforgci!j. 26

PAGE 27

2.2.1.2SpectraldomainpairwiseGrangercausalityPairwiseGC,presentedinthetimedomainintheprevioussection,canalsobeextendedtothespectraldomaininordertoobtainestimatesofthecausalrelationshipbetweenxi(t)andxj(t)atdifferentfrequencyvalues.DoingsorequiresrstrewritingtheVARmodelttoxi(t)andxj(t)inEq.( 2 ),intermsofthelagoperator,LkX(t)=X(t)]TJ /F5 11.955 Tf 11.95 0 Td[(k)whichgives(t)= A(0))]TJ /F9 7.97 Tf 18.27 15.21 Td[(pXk=1A(k)Lk!X(t), (2)whereA(0)=Iisthe2-dimensionalidentitymatrix.FouriertransformingbothsidesofEq.( 2 )thenleadstoA(f)X(f)=E(f). (2)RecallingthatH(f)=A)]TJ /F8 7.97 Tf 6.58 0 Td[(1(f)(cf.Eq.( 2 )),Eq.( 2 )canbewrittenasX(f)=H(f)E(f). (2)Eq.( 2 )canbeexpandedto0B@xi(f)xj(f)1CA=0B@Hii(f)Hij(f)Hji(f)Hjj(f)1CA0B@Ei(f)Ej(f)1CA. (2)ThespectralmatrixofX(t)fromEq.( 2 )canalsobeexpandedinthiswayas0B@Sii(f)Sij(f)Sji(f)Sjj(f)1CA=0B@Hii(f)Hij(f)Hji(f)Hjj(f)1CA0B@iiijjijj1CA0B@Hii(f)Hij(f)Hji(f)Hjj(f)1CA (2)wherethetermsSii(f)andSjj(f)representautospectraandSij(f)andSji(f)representcrossspectra.AccordingtoEq.( 2 ),theautospectrumofxi(t)is( Gardner 1988 )Sii(f)=Hii(f)iiHii(f)+2ijRe(Hii(f)Hij(f))+Hij(f)jjHij(f). (2) 27

PAGE 28

Whenxi(t)andxj(t)areindependent,ij=0,andtheautospectrumofxi(t)simpliestotwoterms.ThersttermisconsideredtheintrinsiccontributiontoSii(f)sinceitdependsonlyonHii(f)andthevarianceofi,2,whilethesecondtermisviewedasthecausalcontributionsinceitinvolvesHij(f)andthevarianceofj,2,whichcomesfromtheVARmodelequationforxj(t).Whenij6=0,itbecomesdifculttoseparatetheintrinsicandcausalcontributionstotheautospectrumofxi(t).Gewekeintroducedatransformation,however,thateliminatesthecrossterminEq.( 2 )( Geweke 1982 ).Thistransformation,callednormalization,consistsofleft-multiplyingbothsidesofEq.( 2 )bythematrixP=0B@10)]TJ /F8 7.97 Tf 10.49 6.11 Td[(ij ii11CA, (2)andtransformingtheresulttothefrequencydomainviaFouriertransform.Thisgives~A(f)X(f)=~E(f). (2)WritingEq.( 2 )intermsoftheexpandedtransferfunctionmatrixgives0B@xi(f)xj(f)1CA=0B@~Hii(f)Hij(f)~Hji(f)Hjj(f)1CA0B@Ei(f)~Ej(f)1CA (2)where~Ej(f)=Ej(f))]TJ /F8 7.97 Tf 13.17 6.11 Td[(ij iiEi(f),~Hii(f)=Hii(f)+ij iiHij(f),and~Hji(f)=Hji(f)+ij iiHii(f).Inthiscase,thecovariancematrixofthenoiseterms,Ei(f)and~Ej(f),is~=0B@ii00~jj1CA, (2)where~jj=jj)]TJ /F8 7.97 Tf 13.15 7.48 Td[(2ij ii.UsingEq.( 2 ),thenewautospectrumofxi(t)isSii(f)=~Hii(f)ii~Hii(f)+Hij(f)~jjHij(f), (2) 28

PAGE 29

withthersttermontherighthandsideinterpretedastheintrinsicpowerandthesecondtermasthecausalpowerduetoxj(t).AssuggestedbyGeweke,pairwiseGCinthespectraldomainisthendenedasthenaturallogarithmoftheratiooftotaltointrinsicpower( Geweke 1982 ):GCj!i(f)=lnjSii(f)j j~Hii(f)ii~Hii(f)j=lnjSii(f)j jSii(f))]TJ /F7 11.955 Tf 11.96 0 Td[((jj)]TJ /F8 7.97 Tf 13.15 7.55 Td[(2ij ii)jHij(f)j2j, (2)wherejjisabsolutevalue.Accordingtothisdenition,ifthecausalpowertermiszero,thenGCj!i(f)=0.Also,foraxedvalueoftotalpower,asthecausalpowertermincreases,GCj!i(f)increases.Thus,thisformulationaccuratelyreectstheintuitivemeaningofthetermcausality.InordertoobtainspectralGCfromxi(t)toxj(t),atransformationmatrixoftheformP=0B@1)]TJ /F8 7.97 Tf 10.59 6.11 Td[(ij jj011CA, (2)isagainleft-multipliedonbothsidesofEq.( 2 ).PerformingananalysissimilartotheoneaboveresultsinGCfromxi(t)toxj(t):GCi!j(f)=lnjSjj(f)j j~Hjj(f)jj~Hjj(f)j=lnjSjj(f)j jSjj(f))]TJ /F7 11.955 Tf 11.96 0 Td[((ii)]TJ /F8 7.97 Tf 13.24 7.54 Td[(2ij jj)jHji(f)j2j, (2)GewekedemonstratedthatthetimedomainGCmeasureistheoreticallyrelatedtothespectraldomainmeasureaccordingtogci!jZ1=2)]TJ /F8 7.97 Tf 6.59 0 Td[(1=2GCi!j(f)df, (2)where,inmanycases,theequalityholds( Geweke 1982 ). 29

PAGE 30

2.2.1.3ExampleapplicationsConsidertwotimeseries,xi(t)andxj(t)generatedaccordingtothefollowingstochasticARmodelxi(t)=(t) (2)xj(t)=xi(t)]TJ /F7 11.955 Tf 11.96 0 Td[(1)+(t),where(t)and(t)arewhitenoiseprocesseswithzeromeanandvariances,2,2,respectively.Clearly,xj(t)iscausallyinuencedbyxi(t),whichcanbewrittenasi!j.Setting=1and=0.2,Isimulatedthissystemforatotalof5,000timepoints.Theresultingtimeseries,xi(t)andxj(t),areshowninFig. 2-1 .Then,makingnoassumptionsabouttheprocessesgeneratingxi(t)andxj(t),IusedtheformulationpresentedabovetocalculatebothtimedomainandspectraldomainpairwiseGC.Forthisexampleandallofthefollowingexamplesinthischapter,IusedtheARtMatlabpackagetotARandVARmodelstothetimeseriesalongwithBICtoselecttheoptimalmodelorder( NeumaierandSchneider 2001 ; SchneiderandNeumaier 2001 ; Schwarz 1978 ).Also,inordertoguaranteezeromeantimeseries,Isubtractedthetemporalmeanofeachtimeseriesfromitself.TimedomainpairwiseGCforthissystemisgcj!i=ln1.026 1.026=0gci!j=ln1.067 0.039=3.309.ThisresultdemonstratesthattimedomainGCaccuratelycapturesthefactthatcausalityowsunidirectionallyfromxi(t)toxj(t).SpectraldomainpairwiseGCforthissystemispresentedgraphicallyinFig. 2-2 sinceitisafunctionoffrequency.Itisassumedthatthesamplingrateofbothtimeseriesis1Hz.Fromthisgure,themaximumGCacrossallfrequenciesis 30

PAGE 31

MAX(GCj!i(f))=0,MAX(GCi!j(f))=3.576.Thus,causalityfromxi(t)toxj(t)isaccuratelycapturedinthefrequencydomainaswell.Furthermore,Icanverifynumericallythatgcj!i=0Z1=2)]TJ /F8 7.97 Tf 6.59 0 Td[(1=2GCj!i(f)df=0gci!j=3.309Z1=2)]TJ /F8 7.97 Tf 6.59 0 Td[(1=2GCi!j(f)df=3.230.Asanexampleofanonlinearsystem,considertwoseriesxi(t)andxj(t)generatedbystochasticvanderPoloscillators( vanderPol 1922 )xi(t)=(1)]TJ /F5 11.955 Tf 11.95 0 Td[(x2i)_xi)]TJ /F11 11.955 Tf 11.96 0 Td[(!2ixi+ii(t)+"ij(xj)]TJ /F5 11.955 Tf 11.96 0 Td[(xi) (2)xj(t)=(1)]TJ /F5 11.955 Tf 11.96 0 Td[(x2j)_xj)]TJ /F11 11.955 Tf 11.95 0 Td[(!2jxj+jj(t)+"ji(xi)]TJ /F5 11.955 Tf 11.95 0 Td[(xj),wherei(t)andj(t)arebothGaussiandistributedwhitenoiseprocesses.Iset=5,!i=1.5,!j=1.48andi=j=1.5.Thecouplingstrengthfromxi(t)toxj(t)iscontrolledby"jiandcouplingfromxj(t)toxi(t),by"ij.Iset"ji=0.4and"ij=0sothatthereisunidirectionalcouplingfromxi(t)toxj(t).Isimulatedthissystemforatotalof30,000timepointsatarateof10Hz.Inthisexampleandforallfuturestochasticdifferentialequationsinmywork,IemployedtheEuler-Maruyamamethodtondanapproximatenumericalsolutiontothestochasticdifferentialequation( KloedenandPlaten 1999 ).Inthiscase,thestepsizewassettot=0.005.Theresultingtimeseries,xi(t)andxj(t),areshowninFig. 2-3 .Makingnoassumptionsabouttheprocessesgeneratingxi(t)andxj(t),IusedtheformulationpresentedabovetocalculatebothtimedomainandspectraldomainpairwiseGC.TimedomainpairwiseGCforthissystemisgcj!i=ln0.166 0.166=0gci!j=ln0.025 0.013=0.181. 31

PAGE 32

FromtheseresultsitispossibletoconcludethattimedomainpairwiseGCcorrectlyidentiestheunidirectionalinteractionfromxi(t)toxj(t).SpectraldomainpairwiseGCforthissystemispresentedgraphicallyinFig. 2-4 .Fromthisgure,themaximumGCacrossallfrequenciesisMAX(GCj!i(f))=0.037,MAX(GCi!j(f))=4.687.Inthiscase,thenonlinearlyofthesystemcannotbefullycapturedbythelinearARandVARmodelsttothetimeseries.ThisleadstosomeamountoferrorinthecalculationofGCj!i(f),whichisnolongeridenticallyzero.Nevertheless,itisstillquantitativelymuchlowerthanGCi!j(f)anditispossibletoconcludethatpairwise,spectraldomainGCcorrectlyidentiestheunidirectionalinteractionfromxi(t)toxj(t).Thisresulthighlightstheneedforsomeformofsignicancetesting,wherebyitcanbedeterminedwhetherthevaluesofGCj!i(f)andGCi!j(f)aresignicantlydifferentfromzero.SuchtestingisdescribedinSection 2.6 2.2.2ConditionalGrangerCausalityConditionalGCisatimeseriesinferencetechniquethatcanbeusedtomeasuretheeffectiveconnectivitybetweenthreestochasticprocesses,denotedasxi(t),xj(t),andxh(t).Apairwiseanalysiscanalsobeusedinthiscasebyconsideringonlythepairwiseinteractionsbetweenprocesses,however,thisapproachhassomeinherentlimitationswhichareillustratedinSection 2.2.2.4 .Mostimportantly,conditionalGChastheabilitytoresolvewhetheraninteractionbetweentwoprocessesisdirectorismediatedbyathirdprocess.However,likepairwiseGC,thistechnique,sinceitisbasedonlinearVARmodelsisitselflinearandparametric.Thetechnique,asdescribedinthefollowing,canalsobegeneralizedtothreesetsofstochasticprocesses. 2.2.2.1TimedomainconditionalGrangercausalityAssumingthatxi(t),xj(t),andxh(t)arezeromean,individuallyandjointlystationarystochasticprocesses,abivariateVARmodelcanbettoxi(t)andxh(t) 32

PAGE 33

sothatxi(t)=pXk=1Aii,2(k)xi(t)]TJ /F5 11.955 Tf 11.95 0 Td[(k)+pXk=1Aih,2(k)xh(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+ei,2(t) (2)xh(t)=pXk=1Ahi,2(k)xi(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+pXk=1Ahh,2(k)xh(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+eh,2(t),wherethecovariancematrixofthenoisetermsis2=0B@ii,2ih,2hi,2hh,21CA (2).Next,atrivariateVARmodelcanbettoxi(t),xj(t),andxh(t)suchthatxi(t)=pXk=1Aii,3(k)xi(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+pXk=1Aij,3(k)xj(t)]TJ /F5 11.955 Tf 11.95 0 Td[(k)+pXk=1Aih,3(k)xh(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+ei,3(t)xj(t)=pXk=1Aji,3(k)xi(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+pXk=1Ajj,3(k)xj(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+pXk=1Ajh,3(k)xh(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+ej,3(t)xh(t)=pXk=1Ahi,3(k)xi(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+pXk=1Ahj,3(k)xj(t)]TJ /F5 11.955 Tf 11.95 0 Td[(k)+pXk=1Ahh,3(k)xh(t)]TJ /F5 11.955 Tf 11.95 0 Td[(k)+eh,3(t), (2)wherethecovariancematrixofthenoisetermsis3=0BBBB@ii,3ij,3ih,3ji,3jj,3jh,3hi,3hj,3hh,31CCCCA (2).Then,timedomainGCfromxj(t)toxi(t)conditionedonxh(t)isdenedasgcj!ijh=lnii,2 ii,3, (2)withconditionalGCfortheotherinteractionssimilarlydened( Dingetal. 2006 ; Geweke 1984 ). 33

PAGE 34

Whencausalityfromxj(t)toxi(t)ismediatedentirelythoughxh(t),Aij,3(k)=0leavingii,3=ii,2.Thus,gcj!ijh=0,indicatingthatthereisnodirectcausalityfromxj(t)toxi(t).Ontheotherhand,whensomedirectcausalityfromxj(t)toxi(t)doesexist,addingxj(t)totheVARmodelofxi(t)reducesthevarianceoftheerrorterms,leadingtoii,30.Asthesecasesillustrate,thisdenitionofconditionalGCisinaccordwiththeintuitivemeaningofdirectcausality. 2.2.2.2SpectraldomainconditionalGrangercausalityFormulatingconditionalGCinthespectraldomaininvolvesrstnormalizingthebivariateVARmodelofEq.( 2 )byleft-multiplyingbothsidesoftheequationbythePmatrixinEq.( 2 ).Transformingtheresulttothefrequencydomaingives0B@xi(f)xh(f)1CA=0B@Gii(f)Gih(f)Ghi(f)Ghh(f)1CA0B@Ei,2(f)Eh,2(f)1CA, (2)where,duetonormalization,cov(Ei,2(f),Eh,2(f))=0.SincenormalizationdoesnotchangethevarianceofEi,2(f),itisstillthecasethatvar(Ei,2(f))=ii,2.Next,thetrivariateVARmodelofEq.( 2 )mustalsobenormalized.Thisisdonebyleft-multiplyingbothsidesoftheequationbythematrixP=P2P1( Dingetal. 2006 ; Geweke 1984 )whereP1=0BBBB@100)]TJ /F7 11.955 Tf 9.3 0 Td[(ji,3)]TJ /F8 7.97 Tf 6.58 0 Td[(1ii,310)]TJ /F7 11.955 Tf 9.3 0 Td[(hi,3)]TJ /F8 7.97 Tf 6.58 0 Td[(1ii,3011CCCCA,P2=0BBBB@1000100)]TJ /F8 7.97 Tf 10.5 7.21 Td[(hj,3)]TJ /F8 7.97 Tf 6.58 0 Td[(hi,3)]TJ /F16 5.978 Tf 5.75 0 Td[(1ii,3ij,3 jj,3)]TJ /F8 7.97 Tf 6.58 0 Td[(ji,3)]TJ /F16 5.978 Tf 5.76 0 Td[(1ii,3ij,311CCCCA. (2)Aftertransformingtheresulttothefrequencydomain,oneobtains0BBBB@xi(f)xj(f)xh(f)1CCCCA=0BBBB@Hii(f)Hij(f)Hih(f)Hji(f)Hjj(f)Hjh(f)Hhi(f)Hhj(f)Hhh(f)1CCCCA0BBBB@Ei,3(f)Ej,3(f)Eh,3(f)1CCCCA. (2) 34

PAGE 35

Thecovariancematrixofthenoiseterms,Ei,3(f),Ej,3(f),andEh,3(f),is^3=0BBBB@^ii,3000^jj,3000^hh,31CCCCA. (2)AkeyequalityprovenbyGewekeisthefollowing( Geweke 1984 ).:GCj!ijh(f)=GCjeh,2!ei,2(f), (2)whereeh,2andei,2representthetimeseries,eh,2(t)andei,2(t),thatresultfromtakingtheinverseFouriertransformofEh,2(f)andEi,2(f),respectively.ThisequalitymeansthatGCfromxj(t)toxi(t)conditionedonxh(t)isequivalenttounconditionalGCfromxj(t)andeh,2(t)toei,2(t).Thislatter,unconditionalGCcanbeformulatedbyrstcomputingtheautospectrumofei,2(t)andthendecomposingitintoanintrinsictermandacausalterm,similartoEqs.( 2 2 ).Theautospectrumofei,2(t)canbecomputedbyrstassumingthatxi(f)andxh(f)inEq.( 2 )areequivalenttoxi(f)andxh(f)inEq.( 2 ).Then,Eq.( 2 )andEq.( 2 )canbecombinedtoyield0BBBB@Ei,2(f)xj(f)Eh,2(f)1CCCCA=0BBBB@Gii(f)0Gih(f)010Ghi(f)0Ghh(f)1CCCCA)]TJ /F8 7.97 Tf 6.58 0 Td[(10BBBB@Hii(f)Hij(f)Hih(f)Hji(f)Hjj(f)Hjh(f)Hhi(f)Hhj(f)Hhh(f)1CCCCA0BBBB@Ei,3(f)Ej,3(f)Eh,3(f)1CCCCA (2)=0BBBB@Qii(f)Qij(f)Qih(f)Qji(f)Qjj(f)Qjh(f)Qhi(f)Qhj(f)Qhh(f)1CCCCA0BBBB@Ei,3(f)Ej,3(f)Eh,3(f)1CCCCA,whereQ(f)=G)]TJ /F8 7.97 Tf 6.59 0 Td[(1(f)H(f).WithQ(f)asthetransferfunctionand^3asthecovariancematrix,theautospectrumofei,2(t)canbecomputedas( Gardner 1988 )Sei,2ei,2(f)=Qii(f)^ii,3Qii(f)+Qij(f)^jj,3Qij(f)+Qih(f)^hh,3Qih(f). (2) 35

PAGE 36

Thersttermontherighthandsideofthisequationisconsideredtheintrinsicpowerandtheremainingtwotermsarethecombinedcausalityfromxj(t)andeh,2(t).WithGCdenedasthenaturallogarithmoftheratiooftotaltointrinsicpower,GCfromxj(t)andeh,2(t)toei,2(t)is( Dingetal. 2006 ; Geweke 1984 )GCjeh,2!ei,2(f)=lnjSei,2ei,2(f)j jQii(f)^ii,3Qii(f)j. (2)Becauseei,2(t)isawhitenoiseprocess,itsspectrum,Sei,2ei,2(f)isequaltoitsvariance,ii,2.Thus,thenalexpressionforconditionalGCfromxj(t)toxi(t)conditionedonxh(t)isGCj!ijh(f)=lnii,2 jQii(f)^ii,3Qii(f)j, (2)withconditionalGCfortheotherinteractionssimilarlydened.Onceagain,itwasshownbyGeweke( Geweke 1984 )thatgcj!ijhZ1=2)]TJ /F8 7.97 Tf 6.59 0 Td[(1=2GCj!ijh(f)df. (2) 2.2.2.3PartitionmatrixGrangercausalityAkeyassumptionintheabovederivationofconditionalGCisthatxi(f)andxh(f)inEq.( 2 )andEq.( 2 )areequivalent.Whiletheoreticallythisholdtrue,itmaybehardtosatisfynumericallywhenusingrealdataduetoerrorsinestimation( Chenetal. 2004 ).Chenetal.demonstratedthatsucherrorscanleadtospectralGCthatcontainsnegativevalues,whichhavenointerpretationintermsofcausality,and/orseveralartifactualpeaksacrossfrequency.Chenetal.alsointroducedamethodforresolvingthisissuewhichtheytermedthepartitionmatrixtechnique( Chenetal. 2004 ).Thepartitionmatrixtechniqueavoidsviolatingtheassumptionthatxi(f)andxh(f)inEq.( 2 )andEq.( 2 )areequivalentbyeliminatingthebivariateVARmodelinEq.( 2 )altogether.Initsplace,thistechniquepartitionsthetrivariateVARmodelinEq.( 2 )intoabivariateexpressionforxi(f)andxh(f).Inthisway,itisguaranteed 36

PAGE 37

thatxi(f)andxh(f)areequivalentinthebivariateandtrivariateexpressions.Thebivariateexpressionforxi(f)andxh(f),constructedbypartitioningEq.( 2 )is0B@xi(f)xh(f)1CA=0B@Hii(f)Hih(f)Hhi(f)Hhh(f)1CA0B@Ei,2(f)Eh,2(f)1CA, (2)whereitcanbeshownthat0B@Ei,2(f)Eh,2(f)1CA=0B@Ei,2(f)Eh,2(f)1CA+0B@Hii(f)Hih(f)Hhi(f)Hhh(f)1CA)]TJ /F8 7.97 Tf 6.59 0 Td[(10B@Hij(f)Hhj(f)1CAEj,2(f). (2)CondensingthetwoHmatricesintheaboveequationintoasingletermgives0B@Hij(f)Hhj(f)1CA=0B@Hii(f)Hih(f)Hhi(f)Hhh(f)1CA)]TJ /F8 7.97 Tf 6.59 0 Td[(10B@Hij(f)Hhj(f)1CA. (2)Then,thecovariancematrixofthenoiseterms,Ei,2(f)andEh,2(f),canbefoundusingEq.( 2 )andis(f)=0B@ii,3ih,3hi,3hh,31CA+0B@Hij(f)Hhj(f)1CAij,3hj,3+0B@ij,3hj,31CAHij(f)Hhj(f)+jj,30B@Hij(f)Hhj(f)1CAHij(f)Hhj(f), (2)withthevaluescomingfromEq.( 2 ).DuetotheH(f)terms,thecross-termsofthiscovariancematrixarenolongerrealbutsatisfytherelationshipih,3(f)=hi,3(f).NormalizingEq.( 2 )inordertomakethenoiseterms,Ei,2(f)andEh,2(f),independentnowrequiresthefollowingtransformationmatrix( Chenetal. 2004 )P=0B@10)]TJ /F8 7.97 Tf 11.48 7.55 Td[(ih,3(f) ii,311CA. (2) 37

PAGE 38

Afterapplyingthistransformationmatrix,thetransferfunctionofEq.( 2 )isG(f)=0B@Hii(f)Hih(f)Hhi(f)Hhh(f)1CA0B@10)]TJ /F8 7.97 Tf 11.48 7.54 Td[(ih,3(f) ii,311CA)]TJ /F8 7.97 Tf 6.58 0 Td[(1. (2)AsinEq.( 2 ),G(f)canbeexpandedtoobtainQ(f)=G)]TJ /F8 7.97 Tf 6.59 0 Td[(1(f)H(f).Now,ii,2=ii,3,yielding( Chenetal. 2004 ; Dingetal. 2006 )GCj!ijh(f)=lnii,3 jQii(f)^ii,3Qii(f)j. (2) 2.2.2.4ExampleapplicationsConditionalGCrequiresatleastthreetimeseriesbeavailableforanalysis.Therefore,inordertodemonstrateitsapplication,consideranextendedVARmodelofthestochasticARmodelpresentedinEq.( 2 )xi(t)=(t) (2)xj(t)=xi(t)]TJ /F7 11.955 Tf 11.95 0 Td[(1)+(t)xh(t)=xh(t)]TJ /F7 11.955 Tf 11.96 0 Td[(1)+xi(t)]TJ /F7 11.955 Tf 11.96 0 Td[(2)+"(t),where(t),(t),and"(t)arewhitenoiseprocesseswithzeromeanandvariances,2,2,2",respectively,andjj<1isaparameterofthesystem.Inthiscase,xi(t)iscausallyinuencingbothxj(t)andxh(t),butwithdifferentdelays.Setting=1,=0.2,"=0.3,and=0.5,Isimulatedthissystemforatotalof5,000timepoints.TheresultingtimeseriesweresimilartothoseinFig. 2-1 .Then,makingnoassumptionsabouttheprocessesgeneratingthetimeseries,IusedtheformulationpresentedabovetocalculatebothtimedomainandspectraldomainconditionalGC.Ibeginbyperformingapairwiseanalysisonthesystem,usingpairwisetimedomainGCtodeterminetheinteractionbetweeneachpairoftimeseries.Thisresultsinthe 38

PAGE 39

followinggci!j=ln1.028 0.041=3.222,gcj!i=ln0.995 0.995=0gch!j=ln1.028 1.028=0,gcj!h=ln1.091 0.1312=2.118gci!h=ln1.091 0.090=2.495,gch!i=ln0.995 0.995=0.Whilethecausalityfromxi(t)toxj(t)andfromxi(t)toxh(t)iscorrectlycaptured,erroneouscausalityfromxj(t)toxh(t)isalsocapturedduetothecommoninuenceofxi(t)onbothxj(t)andxh(t)withdifferenttimedelays.ThisexampleillustratesonelimitationofusingpairwiseGCanalysisonasystemcontainingmorethantwotimeseries.Next,IappliedconditionalGCtothesystem.ThetimedomainconditionalGCforthissystemisgci!jjh=ln1.029 0.041=3.222,gcj!ijh=ln0.996 0.996=0gch!jji=ln0.041 0.041=0,gcj!hji=ln0.090 0.090=0gci!hjj=ln0.131 0.090=0.377,gch!ijj=ln0.996 0.996=0ThisresultdemonstratesthattimedomainconditionalGCaccuratelycapturesthefactthatcausalityowsonlyfromxi(t)toxj(t)andfromxi(t)toxh(t).Theerroneousconnectionfromxj(t)toxh(t)thatwaspresentinthepairwiseanalysisiscorrectlyeliminatedbythetechnique.Inthespectraldomain,IappliedconditionalGCbutdidnotemploythepartitionmatrixmethod.Fig. 2-5 illustratestheresults.Inthiscase,GCi!jjh(f)appearscorrect,butbothGCh!ijj(f)andGCh!jji(f)takeonnegativevaluesforwhichthereisnocausalinterpretation.Uponfurtherinspectionofthesystem,theproblemliesinthefactthatxi(f)andxj(f),whichshouldbeequivalentaccordingtoEq.( 2 )andEq.( 2 ),arenotequal. 39

PAGE 40

Toresolvethisissue,IappliedthepartitionmatrixmethodandobtainedtheresultsinFig. 2-6 .Fromthisgure,themaximumconditionalGCacrossallfrequenciesfromxi(t)toxj(t)andfromxi(t)toxh(t)isMAX(GCi!jjh(f))=3.363andMAX(GCi!hjj(f))=3.912.ThemaximumconditionalGCacrossallremaininginteractionsandfrequencieswas0.056.Ignoringthissmallamountoferrorthatcreptintothecalculation,causalityonlyfromxi(t)toxj(t)andfromxi(t)toxh(t)isaccuratelycapturedinthefrequencydomainusingthepartitionmatrixapproach.Unfortunately,usingthepartitionmatrixmethodcausestheresultstoviolateGeweke'sinequalityfromEq.( 2 )( Geweke 1984 )whichcanbeseennumericallygci!jjh=3.222
PAGE 41

abouttheprocessesgeneratingxi(t),xj(t)andxh(t),IusedtheformulationpresentedabovetocalculatebothtimedomainandspectraldomainconditionalGC.TimedomainconditionalGCforthissystemisgci!jjh=ln0.254 0.227=0.112,gcj!ijh=ln0.167 0.167=0gch!jji=ln0.222 0.227=)]TJ /F7 11.955 Tf 9.3 0 Td[(0.022,gcj!hji=ln0.201 0.200=0gci!hjj=ln0.236 0.200=0.166,gch!ijj=ln0.167 0.167=0.005Again,thenonlinearlyofthesystemcannotbefullycapturedbythelinearARandVARmodelsttothetimeseries.Thisleadstosomeamountoferrorinthecalculationofgch!jjiandgch!ijjwhicharenolongeridenticallyzero.Nevertheless,thesevaluesarequantitativelymuchlowerthangci!jjhandgci!hjj,makingitpossibletoconcludethatGCcorrectlyidentiestheunidirectionalinteractionfromxi(t)toxj(t)andfromxi(t)toxh(t).SpectraldomainconditionalGCforthissystem,computedusingthepartitionmatrixmethod,ispresentedgraphicallyinFig. 2-7 .ThemaximumGCacrossallfrequenciesforthetwocausalinteractionsactuallypresentinthesystemisMAX(GCi!jjh(f))=6.72andMAX(GCi!hjj(f))=6.57.Incomparison,themaximumGCacrossallfrequenciesandacrossallinteractionsnotpresentisthesystemis2.88,whichcorrespondstoMAX(GCh!ijj(f)).Whilesignicancetestingisneededinordertoconrmthatthisvalueisnotsignicant,thesequantitativeresultssuggestthatspectraldomainconditionalGCaccuratelycapturesthecausalityinthissystem. 2.3PartialDirectedCoherencePartialdirectedcoherence(PDC)wassuggestedbyBaccalaandSameshimaasatimeseriesinferencetechniquefordeterminingthedirectionalityofinformationowbetweenmultivariatestochasticprocessesinthespectraldomain( BaccalaandSameshima 2001 ).Inageneralsense,thistechniquecanbeconsideredanothermeasureofGCsinceitreliesontheconceptoftemporalprecedenceandrestsonmanyofthesameassumptionsasthepairwiseandconditionalGCmeasurespreviously 41

PAGE 42

discussedinthischapter.LiketheseotherGCmeasures,applyingPDCtotimeseriesdatashouldenableonetodeterminetheeffectiveconnectivitybetweenthenodesfromwhichthetimeserieswererecorded.However,PDCisonlyformulatedinthespectraldomainandhasnotimedomainequivalent.Moreover,PDCisinherentlyamultivariatemeasurethatcanbeappliedtodatasetswithanynumberoftimeseries.ThisisincontrasttopairwiseGCwhichisdesignedforbivariatetimeseriesandforwhichadifferentformulation,namelyconditionalGC,mustbeusedwhenthenumberoftimeseriesisgreaterthantwo.SincethebasisforPDCisVARmodeling,itisconsideredalinearandparametrictechnique.ThePDCformulationthatfollowsisbasedon BaccalaandSameshima ( 2001 ).ConsideravectorofstochasticprocessesX(t)=[x1(t),x2(t),,xn(t)]Tthatarezeromeanandjointlystationary.AVARmodelttoX(t)hastheformX(t)=pXk=1A(k)X(t)]TJ /F5 11.955 Tf 11.96 0 Td[(k)+(t). (2)withthecovariancematrixofthenoiseterms,(t),denotedby.Thecoefcientmatrix,A(k),isannnmatrixwhereAji(k)denotestheVARmodelcoefcientrepresentingprocessxi(t)'scontributiontoprocessxj(t).TheformulationofPDCrestsontheobservationthatcausalityfromxi(t)toxj(t)shouldbereectedinthecoefcientmatrix,Aji(k),whichshouldbenonzero.ThisreasoningisextendedtotheFouriertransformofthecoefcientmatrix,A(k),(cf.Eq.( 2 )),whichisA(f).BaccalaandSameshimasuggestnormalizingeachcoefcientterminordertoarriveatthefollowingdenitionofPDCfromxi(t)toxj(t)( BaccalaandSameshima 2001 )PDCi!j=jAji(f)j p Ai(f)Ai(f) (2)whereAi(f)istheithcolumnoftheA(f)matrix.Directcausalityfromprocessxi(t)toprocessxj(t)isinferredbyanon-zerovalueofPDCi!j,whichisnormalizedbetweenzeroandone.Intuitively,thisdenitioncapturestherelativeinuenceofxi(t)onxj(t) 42

PAGE 43

comparedtotheinuenceofxi(t)onallotherprocessesinthesystem.PDCfromxj(t)toxi(t)isanalogouslydened.AnimportantdifferencebetweenPDCandthepairwiseandconditionalGCmeasurespresentedpreviouslyisthatPDCdoesnotincludetermsfromtheVARmodel'snoisecovariancematrixinitsformulation.Thefollowingexampledemonstratesalimitationthatarisesfromthisfact.ConsiderthreeindependentGaussianwhitenoiseprocesses,xi(t)=ii(t),xj(t)=jj(t),andxh(t)=hh(t)withvariancesi=1,j=h=500.TheresultsofaPDCanalysisofthissystemareshowninFig. 2.6 .Whilethereisnoactualcausalityinthesystem,PDCincorrectlyidentiescausalityfromxi(t)toxj(t)andfromxi(t)toxh(t).Thisfalsedetectionofcausalityisaresultofthedifferenceinvariancebetweenxi(t),xj(t),andxh(t).PDCincorrectlyidentiescausalityfromtheprocesswithlowvariancetotheprocesseswithmuchhighervariances.ThisiscausedbyerroneousVARmodelparameterestimateswhichareduetothevariancedifferencesinthesystem( Winterhalderetal. 2005 ).Forinstance,intheexampleabove,thecoefcientsjAij(f)j,jAhj(f)j,jAih(f)j,andjAij(f)jdonotexceedvaluesontheorderof10)]TJ /F8 7.97 Tf 6.59 0 Td[(3.Incontrast,MAX(jAji(f)j)=3.9andMAX(jAhi(f)j)=7.5.SincethevaluesofAji(f)andAhi(f)areseveralordersofmagnitudegreaterthanalloftheothercoefcientvaluesinthemodel,PDCerroneouslyidentiescausalityfromxi(t)toxj(t)andfromxi(t)toxh(t).GeneralizedPDC,whichisdiscussedinthenextsection,wasformulatedtoremedythislimitationofPDC. 2.3.1GeneralizedPartialDirectedCoherenceBaccalaetal.introducedgeneralizedPDC(GPDC)inordertoovercomethelimitationofPDCdiscussedabove( Baccalaetal. 2007 ).GPDCincludestermsfromtheVARmodel'snoisecovariancematrixinitsformulation,effectuallyaccountingforerrorsintheestimationofthecoefcientmatrixthatarisewhentherearelargedifferencesinvarianceinthesystem. 43

PAGE 44

GPDCfromxi(t)toxj(t)isdenedas( Baccalaetal. 2007 )GPDCi!j=1 iijAji(f)j s nPk=11 2kkAki(f)Aki(f), (2)whereiiandkkrepresentthevariancesofthenoisetermsintheVARmodel'sequationsforxi(t)andxk(t),respectively.Directcausalityfromprocessxi(t)toprocessxj(t)isinferredbyanon-zerovalueofGPDCi!j,whichisnormalizedbetweenzeroandone.TheintuitionbehindGPDCasameasureofcausalityisthesameasthatforPDC.GPDCfromxj(t)toprocessxi(t)isformulatedinananalogousfashion.AgainconsideringtheexampleofthreeindependentGaussianwhitenoiseprocesses,xi(t)=2ii(t),xj(t)=2jj(t),andxh(t)=2hh(t)withthesameparametersettingsasabove,itisshowninFig. 2-9 thatGPDCcorrectlyidentiesnocausalityinthesystem( Winterhalderetal. 2005 ). 2.3.2ExampleApplicationsConsiderthesystemdescribedinEq.( 2 ).TheresultswhenPDCandGPDCareappliedtothissystemareshowninFigs. 2-10 and 2-11 ,respectively.Fromtheseresults,itisclearthatbothmethodsareabletocorrectlyidentifycausalityinthesystem,specicallythatfromxi(t)toxj(t)andfromxi(t)toxh(t).EventhoughthevanderPolsystemisanonlinearone,linearmethodssuchasPDCandGPDCareabletocorrectlyidentifythecausalinteractionsinthesystem. 2.4PhaseDynamicsModelingPhasedynamicsmodeling(PDM)wasintroducedbyRosenblumandPikovskyasatimeseriesinferencetechniquedesignedspecicallyfordetectingthedirectionalityofinteractionsbetweenweaklycoupled,narrow-bandoscillators( RosenblumandPikovsky 2001 ).Whilethisclassofapplicablesystemsseemsrestrictive,narrow-bandoscillatorsarethetopicofawiderangeofstudiesastheyarerepresentativeofmanynaturalphenomena( Pikovskyetal. 2002 ).LikethepreviouslydescribedGC 44

PAGE 45

techniques,applyingPDMtotimeseriesofweaklycoupledoscillatorsshouldenableonetodeterminetheeffectiveconnectivitybetweentheoscillatorsfromwhichthetimeserieswererecorded.Whilerstformulatedasapairwisetechnique,PDMwaslaterextendedtoanalyzemultivariatetimeseriesaswell( SmirnovandBezruchko 2009 ).Inaddition,PDMisstrictlyatimedomainmeasureandhasnospectraldomainequivalent.AkeyfeatureofPDMisthefactthatitutilizestimeseriesofinstantaneousphaseratherthantheactualtimeseriessignalrecordedfromtheoscillatorsthemselves.Thesetimeseriesofinstantaneousphase,whichwedenoteas(t),canbeobtainedbytakingtheHilberttransformoftheoriginaltimeseriesoftheoscillator( Pikovskyetal. 2002 ).Forexample,givenatimeseriess(t),theHilberttransformofs(t)isdenedassH(t)=)]TJ /F8 7.97 Tf 6.59 0 Td[(1P.V.Z1s() t)]TJ /F11 11.955 Tf 11.96 0 Td[(d, (2)whereP.V.meansthattheintegralistakeninthesenseoftheCauchyprincipalvalue.Then,ananalyticsignal,(t),whichisacomplexfunctionoftime,isdenedby(t)=s(t)+isH(t)=A(t)ei(t). (2)Theinstantaneousamplitude,A(t),andtheinstantaneousphase,(t),ofs(t)aredenedaccordingtothisanalyticfunction.Whilethereareothertechniquesfordeningtheinstantaneousphaseofasignal( Pikovskyetal. 2002 ),IusetheHilberttransform.ThegeneralapproachofPDMissimilartothattakenbyGCandPDC.Amodelisrstttothetimeseriesofinstantaneousphasesandthecoefcientsofthismodelareusedtocalculatecausality.However,unlikeGCandPDCwhichttimeseriestolinearVARmodels,PDMtsthetimeseriestofunctionsofsineandcosine.Sincetheseresultingfunctionsarenonlinearbyvirtueofthefactthatsineandcosinearenonlinear,PDMcanbeconsideredanonlineartechnique.Thenonlinearitycapturedbyfunctions 45

PAGE 46

ofsineandcosine,however,islimitedandthusithasbeenshownthatPDMisonlycapableofcapturingsomespecicformsofnonlinearity( Smirnovetal. 2005 ).AnotherdistinguishingfeatureofPDMisthat,unlikethelinearmethodsGCandPDC,itisapplicabletodeterministictimeseries.Thisusecaseisgenerallyonlyconsideredintheory,however,sincetimeseriesrecordedfromphysicalprocesseswillnearlyalwayscontainsomenoise.ThePDMformulationthatfollowsisbasedonworkbyRosenblumetal.andSmirnovetal.( RosenblumandPikovsky 2001 ; SmirnovandBezruchko 2009 ). 2.4.1PairwisePhaseDynamicModelingLetxiandxjbetwosignalsrecordedfromtwoweaklycoupledoscillatorsandleti(t)andj(t)bethecorrespondinginstantaneousphasesasdenedabove.Itcanbeassumedthatthephaseincrementi(t+))]TJ /F11 11.955 Tf 12.63 0 Td[(i(t)isgeneratedbysomeunknowntwo-dimensionalnoisymapsuchthat i(t+))]TJ /F11 11.955 Tf 11.96 0 Td[(i(t)=Fi(i(t),j(t))+i(t).(2)Thevalueoftheconstanthasbeenshowntoberelativelyunimportantandistypicallysettoonebasicoscillationperiodofoscillatori.Therandomprocess,,isofzeromeanandthefunctionFi(i(t),j(t))isatrigonometricpolynomialoftheform( RosenblumandPikovsky 2001 )Fi(i(t),j(t))=Xm,`[am,`cos(mi(t)+`j(t))+bm,`sin(mi(t)+`j(t))]. (2)ThecoefcientsofEq.( 2 ),am,`andbm,`,areestimatedusingaleastsquaresttotheinstantaneousphaseincrementtimeseries,i(t+))]TJ /F11 11.955 Tf 12.23 0 Td[(i(t),andtheconstants,mand`determinetheorderofthefunctionFi(i(t),j(t)).Ahigher-orderfunctionisabletocapturemorenonlinearityinthesystembutcanalsoleadtomodelovertting.SincethirdorderfunctionshavebeenfoundtobesuitableforPDManalysis,Ichosem3and`3( RosenblumandPikovsky 2001 ). 46

PAGE 47

ThestrengthofcausalityfromoscillatorjtooscillatoriisdeterminedbythedegreeofdependenceofFionj(t),aquantitythatcanbemeasuredbythesensitivityofFitoj(t).Thus,PDMfromj!iisdenedas( RosenblumandPikovsky 2001 ) PDMj!i=s 1 22Z20Z20@Fi(i(t),j(t)) @j2didj,(2)where@Fi(i(t),j(t))=@j(t)isthesensitivityofFi(i(t),j(t))toj(t).Itcanbeshownthat( SmirnovandBezruchko 2003 )PDMj!i=s Xm,``2(a2m,`+b2m,`). (2)AvalueofPDMj!ithatissignicantlygreaterthanzerogenerallyindicatescausalityfromoscillatorjtooscillatori.Causalityfromoscillatoritooscillatorj,PDMi!j,isdenedanalogously. 2.4.2MultivariatePhaseDynamicModelingTheformulaforPDMpresentedaboveisapairwisemeasuresinceitisconstructedfortwoinstantaneousphasetimeseries,namelyi(t)andj(t).Extendingthismeasuretomultivariatetimeseriesingeneralrequiresrstcomputingtheinstantaneousphaseofalloscillatorsbeingconsidered.Assumingthattherearenoscillatorsinasystem,theseinstantaneousphasetimeseriescanbedenotedas1(t),,n(t).Asinthepairwisecase,thephaseincrementofi(t)canbemodeledbythefunction i(t+))]TJ /F11 11.955 Tf 11.96 0 Td[(i(t)=Fi(1(t),,n(t))+i(t),(2)whereFiisnowafunctionofallinstantaneousphasetimeseriesinthesystemandi(t)andaredenedastheywereinthepairwisecase.ThefunctionFiisnowatrigonometricpolynomialoftheform( SmirnovandBezruchko 2009 )Fi(1(t),,n(t))=i,0+nXj=1(j6=i)[i,jcos(j(t))]TJ /F11 11.955 Tf 11.96 0 Td[(i(t))+i,jsin(j(t))]TJ /F11 11.955 Tf 11.95 0 Td[(i(t))]. (2) 47

PAGE 48

Onceagain,thecoefcients,i,jandi,j,areestimatedusingaleastsquaresttotheinstantaneousphasetimeseries.MultivariatePDMisthendenedas( SmirnovandBezruchko 2009 )PDMj!i=1 2(2i,j+2i,j). (2)AvalueofPDMj!ithatissignicantlygreaterthanzerogenerallyindicatescausalityfromoscillatorjtooscillatori.Inthiscase,Eq.( 2 )isoforderonesotheformulaforPDMissimplied.However,higher-ordertermscanbeincorporatedintoEq.( 2 )ifnecessary( SmirnovandBezruchko 2009 ).InboththepairwiseandmultivariateformulationsofPDM,itisassumedthattheinstantaneousphasetimeseries,1(t),,n(t),areindependent.ThisisanecessaryassumptionforttingthephaseincrementstothefunctionsFi(i(t),j(t)))andFi(1(t),,n(t)))inEqs.( 2 2 )( RosenblumandPikovsky 2001 ). 2.4.3ExampleApplicationsInthissection,IwilldemonstratetheapplicationofmultivariatePDMtotwoexamplesystems.TherstistheVARmodeldescribedinEq.( 2 )andthesecondisthesystemofvanderPoloscillatorsdescribedinEq.( 2 ).First,IappliedPDMtothetrivariateVARmodelinEq.( 2 ).TheresultswereasfollowsPDMi!j=1.920,PDMi!h=1.341PDMj!i=1.116,PDMj!h=1.178PDMh!i=0.534,PDMh!j=0.586.ThevaluesofPDMj!iandPDMj!harenearlyashighasthevaluesofPDMi!jandPDMi!h,indicatingthatPDMmostlikelyidentiesthefollowingcausalinteractions:xi(t)toxj(t),xi(t)toxh(t),xj(t)toxi(t)andxj(t)toxh(t).Sincetheonlytruecausalinteractionsinthesystemarefromxi(t)toxj(t)andfromxi(t)toxh(t),PDMerroneously 48

PAGE 49

identiescausalityfromxj(t)toxi(t)andfromxj(t)toxh(t).Thisisnotunexpectedsince,fromthediscussionabove,PDMisdesignedforsystemsofweaklycoupledoscillatorsandthelinearVARmodelconsideredinthisexampledoesnotbehavelikeanoscillator.Next,IappliedPDMtothetrivariatevanderPolmodelinEq.( 2 ).UnlikethelinearVARmodelexample,vanderPoloscillatorsshouldbeamenabletoPDManalysis.TheresultswereasfollowsPDMi!j=0.902,PDMi!h=0.482PDMj!i=0.015,PDMj!h=0.012PDMh!i=0.015,PDMh!j=0.027.InthiscasePDMi!jandPDMi!hareanorderofmagnitudegreaterthanthePDMvaluesoftheotherinteractions.Thus,PDMlikelyidentiescausalityfromxi(t)toxj(t)andfromxi(t)toxh(t)anddoesnoterroneouslyidentifynonexistentinteractions. 2.5RecentAdvancesandApplicationsOfthetimeseriesinferencetechniquesdiscussedinthischapter,GCisthemostwidelyappliedtoneuraltimeseries.Despiteitsknownlimitations,pairwiseGCisoftenusedtoanalyzetimeseriesfrommulti-channelfMRIandEEGrecordings( Cadotteetal. 2010 ; Goebeletal. 2003 ; Roebroecketal. 2005 ).Morerecentstudies,however,haveappliedconditionalGCtosuchmultivariatedatasets( Liaoetal. 2010b ; Zhouetal. 2011 ).AnotherrecentlyproposedapproachutilizesprincipalcomponentanalysistorstreducethedimensionalityofmultivariatetimeseriespriortotheapplicationofconditionalGC( Zhouetal. 2009 ).Thisapproachgreatlydecreasesthecomputationalcostofperformingtheanalysis.PDChasalsobeenappliedtoanumberofneuraltimeseriesdatasetsincludingthosefromfMRIandEEGrecordings( Satoetal. 2009 ; Tropinietal. 2011 ).However,GPDCasproposedbyBaccalaetal.,( Baccalaetal. 2007 )hasyettobeappliedto 49

PAGE 50

recordedneuraltimeseries.Schelteretal.proposedanapproachsimilartoGPDC,whichtheytermedrenormalizedPDC,andappliedthistechniquetoEEGdatarecordedfrompatientssufferingfromParkinsoniantremors( Schelteretal. 2009 ).Becauseitisrestrictedtotheanalysisofnarrow-bandoscillatorydata,PDMhasnotyetgainedwidespreadpopularityasatimeseriesinferencetechniqueinneuroscience.ManyofthestudiesinvestigatingPDMusesimulateddatasets( Bezruchkoetal. 2003 ; Kralemannetal. 2011 ; RosenblumandPikovsky 2001 ).However,afewstudieshaveconsideredtheuseofPDMtodetermineeffectiveconnectivityintherespiratorysystem( Rosenblumetal. 2002 )andinParkinsoniantremors( Smirnovetal. 2008 ). 2.6SignicanceTestingAswasdemonstratedinSections 2.2.1.3 and 2.2.2.4 ,whenlineartimeseriesinferencetechniquesareappliedinthetimedomaintosystemsthatareinherentlylinearandautoregressive,theirresultsareusuallyexact.Forinstance,inEq.( 2 ),thereisnocausalityfromtimeseriesxj(t)toxi(t)andgcj!i=0.Ontheotherhand,thereiscausalityfromxi(t)toxj(t)andgci!j>0.Whensuchlineartechniquesareappliedtononlineartimeseries,however,theirresultstendtobeapproximatesincethelinearmodelstheyarebasedoncannotfullycapturethenonlinearnatureofthesystemunderstudy.Evenwhenthesetechniquesareappliedtothegeneralclassoflinearsystems,theymaynotbeexactduetotheapproximatenatureofVARmodeltting.UsingspectraldomaintechniquesfurtherconfoundsthisproblemsincethespectralmatrixobtainedviaaVARmodeltisalsoanestimationofthetruespectrum.ThissameargumentcanbeextendedtoPDMsinceitcannotfullycaptureallformsofnonlinearitypresentinagivensystemofoscillatorsandthettingofinstantaneousphasetimeseriestoapredenedmodelisapproximate.Therefore,itmustbedeterminedwhethercausalityestimatesobtainedbythesetechniquesaresignicantlygreaterthanzero,inwhichcasecausalityislikelytoexist. 50

PAGE 51

Onestrategyfordeterminingwhetheracausalityestimateobtainedthroughthesetechniquesissignicantlygreaterthanzeroistocompareittoanullcase.Herethenullcaseconsistsofasystemwithnocausalitybetweenitsnodes.Byapplyingeachtechniquetothisnullcase,itispossibletoobtainsignicancethresholdsabovewhichthecausalityestimatesareconsideredsignicantlygreaterthanzero( Fisher 1925 ).Inmywork,Iconstructnullcasesviasurrogateanalysis( PrichardandTheiler 1994 ; SchreiberandSchmitz 2000 ; Timmer 2000 ).Foragivensetoftimeseriesunderstudy,Igenerate1,000surrogatesetsoftimeseriesbywindowingthetimeseriesofeachnodeintodiscreteblocksandthenrandomlyshufingtheseblocks.Thismaintainstheenergycontentofeachnode'stimeseriesbutdestroysthetemporalrelationshipbetweentimeseriesintheset( Kaminskietal. 2001 ).Inthisway,Iobtain1,000setsoftimeserieseachofwhichrepresentsthenullcase.ThenIapplyeachtimeseriesinferencetechniquetothisnullcaseand,usingap-valueofp=0.05,determineasignicancethresholdforeachtechnique.InthecaseofspectraldomainGC,PDCandGPDC,Iconsideronlythemaximumcausalityestimateacrossallfrequencies( BlairandKarniski 1993 ; Chenetal. 2006 ).SurrogatetestingtodeterminesignicancelevelswasnecessaryinthecaseofspectralGCandGPDCsincenoanalyticalboundsonsignicancehavebeenderivedforthesetechniques.WhileanalyticalboundsdoexistforPDCandPDM( Schelteretal. 2006 ),IchosetousesurrogatetestingforthesemeasuresaswellinordertomaketheresultsmorecomparabletothoseobtainedbyGCandGPDC.Inaddition,thereissomeindicationthatnumericalmethodsperformbetteronnitedatasetsthananalyticalboundsderivedasymptotically( DavidsonandMacKinnon 2003 ; MacKinnon 2006 ). 51

PAGE 52

Figure2-1. Timeseriesxi(t),inblue(top),andxj(t),inred(bottom),generatedfromtwoAR,stochasticprocesses.Inthiscase,xi(t)iscausallyinuencingxj(t). 52

PAGE 53

Figure2-2. Theresultsobtainedfromapplyingspectraldomain,pairwiseGCtotwoexampletimeseriesgeneratedfromstochastic,ARmodels.GCfromxi(t)toxj(t)isinblueandGCfromxj(t)toxi(t)isinred.SinceGCi!j(f)isalwaysgreaterthanGCj!i(f),whichisidenticallyzero,GCcapturesthecorrectcausalityinthesystem,whichisfromxi(t)toxj(t). 53

PAGE 54

Figure2-3. Timeseriesxi(t),inblue(top),andxj(t),inred(bottom),generatedfromtwostochasticvanderPoloscillators.Inthiscase,xi(t)iscausallyinuencingxj(t). 54

PAGE 55

Figure2-4. Theresultsobtainedfromapplyingspectraldomain,pairwiseGCtotwoexampletimeseriesgeneratedfromstochastic,vanderPoloscillators.GCfromxi(t)toxj(t)isinblueandGCfromxj(t)toxi(t)isinred.SinceGCi!j(f)isalwaysgreaterthanGCj!i(f),whichisnearlyzero,GCcapturesthecorrectcausalityinthesystem,whichisfromxi(t)toxj(t). 55

PAGE 56

Figure2-5. Theresultsobtainedfromapplyingspectraldomain,conditionalGCwithoutthepartitionmatrixmethodtothreeexampletimeseriesgeneratedfromstochastic,ARmodels.GCfromxi(t)toxj(t)conditionedonxh(t)isinblue,fromxi(t)toxh(t)conditionedonxj(t)isinred,xh(t)toxi(t)conditionedonxj(t)isinyellow,andxh(t)toxj(t)conditionedonxi(t)isingreen.GCfromxh(t)toxi(t)conditionedonxj(t)andfromxh(t)toxj(t)conditionedonxi(t)takesonnegativevalueswhichhavenomeaningfulinterpretationintermsofcausality.Thus,thisillustratestheneedforthepartitionmatrixmethod. 56

PAGE 57

Figure2-6. Theresultsobtainedfromapplyingspectraldomain,conditionalGCwiththepartitionmatrixmethodtothreeexampletimeseriesgeneratedfromstochastic,ARmodels.GCfromxi(t)toxj(t)conditionedonxh(t)isinblue,fromxi(t)toxh(t)conditionedonxj(t)isinred,andtheremainingfourinteractions,whicharenearlyidenticaltooneanother,areinyellow.GCwiththepartitionmatrixmethodcapturesthecorrectcausalityinthesystem,whichisfromxi(t)toxj(t)andfromxi(t)toxh(t)anddoesnoterroneouslyidentifyanycausalinteractionsinthesystemthatarenotpresent. 57

PAGE 58

Figure2-7. Theresultsobtainedfromapplyingspectraldomain,conditionalGCwiththepartitionmatrixmethodtothreeexampletimeseriesgeneratedfromstochastic,vanderPoloscillators.ThemaximumvaluesofGCi!jjh(f)andGCi!hjj(f)aremuchgreaterthanthemaximumGC(f)valuesoftheotherinteractions.Thus,GCwiththepartitionmatrixmethodcapturesthecorrectcausalityinthesystem,whichisfromxi(t)toxj(t)andfromxi(t)toxh(t). 58

PAGE 59

Figure2-8. TheresultsofaPDCanalysisofthreewhitenoiseprocesses.Processxi(t)hasamuchlowervariancethaneitherxj(t)orxh(t)andbecauseofthis,PDCincorrectlyidentiescausalityfromxi(t)toxj(t)andxh(t),showninthetoptwoplots. 59

PAGE 60

Figure2-9. TheresultsofaGPDCanalysisofthreewhitenoiseprocesses.UnlikePDC,GPDCcorrectlyidentiesnocausalityinthesystem. 60

PAGE 61

Figure2-10. TheresultsofaPDCanalysisofthreeexampletimeseriesgeneratedfromstochastic,vanderPoloscillators.ThemaximumvaluesofPDCi!j(f)andPDCi!h(f)aremuchgreaterthanthemaximumPDC(f)valuesoftheotherinteractions.Thus,PDCcapturesthecorrectcausalityinthesystem,whichisfromxi(t)toxj(t)andfromxi(t)toxh(t). 61

PAGE 62

Figure2-11. TheresultsofaGPDCanalysisofthreeexampletimeseriesgeneratedfromstochastic,vanderPoloscillators.ThemaximumvaluesofGPDCi!j(f)andGPDCi!h(f)aremuchgreaterthanthemaximumGPDC(f)valuesoftheotherinteractions.Thus,GPDCcapturesthecorrectcausalityinthesystem,whichisfromxi(t)toxj(t)andfromxi(t)toxh(t). 62

PAGE 63

CHAPTER3CHARACTERIZATIONANDMITIGATIONOFTHEEFFECTSOFNOISEVARIANCEONTIMESERIESINFERENCETECHNIQUESNoiseisubiquitousinbiologicalsystems( Samoilovetal. 2006 ),withthenervoussystembeingnoexception( Faisaletal. 2008 ).Itisthereforeimportanttothoroughlyunderstandthateffectsofnoiseontimeseriesinferencestechniquesusedtodetermineeffectiveconnectivityinneuraltimeseriesrecordings.Noisecanbeendogenoustoaneuralsystembeinganalyzed,arising,forexample,fromthestochasticnatureofcellularmachineryorexogenous,suchasmeasurementnoisewhichmaybeaddedtosignalsbyrecordingdevices.Inthischapter,IinvestigatetheeffectsofnoiseonthetimeseriesinferencestechniquesdiscussedinChapter 2 .Specically,Iconsiderendogenousnoiseandexploretheimpactthatdifferencesinthevarianceofnoisebetweentimeserieshasonthesetechniques.Indthatpartialdirectedcoherence(PDC)andphasedynamicsmodeling(PDM)arebothhighlyaffectedbythesenoisevariancedifferences.WhilethisobservationhasbeenmadepreviouslyinregardstoPDC( Winterhalderetal. 2005 ),thisisthersttimeithasbeennotedasalimitationofPDM.Inlightofthisnding,IproposeamodiedPDMtechnique,whichIcallgeneralizedPDM(GPDM),thatisalongthesamelinesasapreviouslysuggestedmodicationofPDC,knownasgeneralizedPDC(GPDC)( Baccalaetal. 2007 ).UnlikePDM,GPDMisunaffectedbynoisevariancedifferences.IdemonstratetheapplicationofGPDMonanexamplesystemofcoupledoscillatorsandshowthatitisindeedimmunetonoisevariancedifferencesbetweentimeseries. 3.1NoiseVarianceandLinearTimeSeriesInferenceTechniques 3.1.1TheImpactofNoiseVarianceonLinearTimeSeriesInferenceTechniquesAsdemonstratedinSection 2.3 ,thelineartimeseriesinferencetechniquePDCishighlyaffectedbydifferencesinnoisevariancebetweentimeseries.Forinstance,giventwotimeseries,PDCislikelytoidentifycausalityfromthetimeserieswithalower 63

PAGE 64

noisevariancetothetimeserieswithahighernoisevarianceregardlessofthetrueeffectiveconnectivityofthesystem( Winterhalderetal. 2005 ).ThistendencyisduetothefactthatPDCdoesnotincludeinitsformulationthenoisevariancetermsofthelinearVARmodelrstttothetimeseriesdata,asexplainedinSection 2.3 .WhilethisledPDCtoerroneouslyidentifycausalityintheexamplegiveninSection 2.3 ,noisevariancedifferencesdonotinvariablyresultinPDCincorrectlyidentifyingcausality.Atrivialexampleofthisiswhencausalityinasystemisactuallydirectedfromalownoisevariancetimeseriestoahighernoisevariancetimeseries.Inthiscase,evenifPDCisaffectedthenoisevariancedifference,itsresultisnotincorrectsincecausalityfromthelowtohighnoisevariancetimeseriesisphysicallypresentinthesystem.Evenoutsideofthistrivialcase,however,theabovestatementholdstrue.AsIdemonstrateinthefollowing,theprobabilitythatPDCincorrectlyidentiescausalityfromalownoisevariancetimeseriestoahighernoisevariancetimeseriesdependsonthemagnitudeofthenoisevariancedifferenceaswellasthestrengthofthecausalconnectionfromthehighertothelowernoisevariancetimeseries.Considertwotimeseries,xi(t)andxj(t),suchthatxi(t)=xj(t)]TJ /F7 11.955 Tf 11.95 0 Td[(1)+ii(t)xj(t)=jj(t),wherei(t)andj(t)areGaussianwhitenoiseprocesseswithzeromeanandastandarddeviationofone.Inaddition,var(i(t))=2iandvar(j(t))=2j.When>0,thetruecausalityinthissystemisfromxj(t)toxi(t).Furthermore,giventhatij,thissystemdoesnotrepresentthetrivialcasewherecausalityowsfromthelowtothehighnoisevariancetimeseries.Withthissysteminmind,onemayask,underwhatconditionsdoesPDCincorrectlyidentifycausalityfromxi(t)toxj(t)?Inordertoanswerthisquestion,IsimulatedtheabovesystemforarangeofandjvaluesandcomputedtheprobabilitythatPDCincorrectlyidentiedcausality 64

PAGE 65

fromxi(t)toxj(t).Specically,Iseti=1whilevaryingintherange[0,1]andjintherange[1,500].Foreachparametersetting,Isimulatedthesystem1,000timesandappliedPDCtoeachinstance.Then,byaveragingoverthe1,000samplesofeachsystem,IestimatedtheprobabilitythatPDCcorrectlyidentiedcausalityfromxj(t)toxi(t)aswellastheprobabilitythatitincorrectlyidentiedcausalityfromxi(t)toxj(t).Aprobabilisticanalysiswasnecessaryduetothestochasticnatureofthetimeseries.TheresultsarepresentedinFig. 3-1 .TheprobabilitythatPDCincorrectlyidentiescausalityfromxi(t)toxj(t),whichIdenoteasP(PDCi!j),isafunctionofbothandj.Asincreases,P(PDCi!j)decreasesandasjincreases,P(PDCi!j)increases.Sinceiisxedatone,anincreaseinjisalsoanincreaseinthemagnitudeofthenoisevariancedifferencebetweenxi(t)andxj(t),namelyji)]TJ /F11 11.955 Tf 12.5 0 Td[(jj.Inallcases,aslongas>0,PDCalwayscorrectlyidentiescausalityfromxj(t)toxi(t).Thus,theseresultsshowthattheprobabilitythatPDCincorrectlyidentiescausalityfromalownoisevariancetimeseriestoahighnoisevariancetimeseriesdependsonboth,thestrengthofcausalinuencefromthehightothelownoisevariancetimeseries,aswellasonjj)]TJ /F11 11.955 Tf 10.11 0 Td[(ij,themagnitudeofthenoisevariancedifferencebetweenthetimeseries.Asthestrengthofcausalinuencefromthehightothelownoisevariancetimeseriesincreases,PDCislesslikelytoincorrectlyidentifycausalityandasthemagnitudeofthenoisevariancedifferencebetweenthetimeseriesincreases,PDCismorelikelytoincorrectlyidentifycausality. 3.1.2MitigatingtheImpactofNoiseVarianceonLinearTimeSeriesInferenceTechniquesOnetechniqueformitigatingtheimpactofnoisevarianceonPDCistheGPDCtechniquedescribedinSection 2.3.1 .GPDCincorporatesthevarianceofthettedVARmodel'snoisetermsintotheoriginalPDCformulation,effectivelyaccountingforanynoisevariancedifferencebetweenthetimeseries(cf.Eq.( 2 ))( Baccalaetal. 65

PAGE 66

2007 ; Winterhalderetal. 2005 ).WiththismodicationtoPDCinplace,GPDCisnotaffectedbydifferencesinthenoisevariancebetweentimeseries.Infact,repeatingthesameanalysisthatwasperformedaboveonGPDCdemonstratesthatGPDCdoesnotincorrectlyidentifycausalityfromalownoisevariancetimeseriestoahighnoisevariancetimeserieswhenthiscausalityisnotphysicallypresentinthesystem.TheresultsfromthisanalysisarepresentedinFig. 3-2 .GPDCalwaysidentiesthetruecausalityfromxj(t)toxi(t)when>0andhasaverylowprobabilityofincorrectlyidentifyingcausalityfromxi(t)toxj(t),regardlessofthevalueoforjj)]TJ /F11 11.955 Tf 11.95 0 Td[(ij.Grangercausality(GC)isalsoimmunetotheeffectsofnoisevariancedifference,aresultofthefactthatittooincorporatesthevarianceofttedVARmodel'snoiseterms( Dingetal. 2006 ).TheresultsoftheaboveanalysisperformedonGCarepresentedinFig. 3-3 andareidenticaltotheGPDCresults.Thus,theconclusionsmadeaboveinregardstoGPDCareequallyapplicabletoGC.Asthisexampleillustrates,theimpactofnoisevarianceonlineartimeseriesinferencetechniquescanbemitigatedbyincludingintheirformulationthenoisevariancetermsofthelinearVARmodelttothetimeseriesdata.Inthefollowingsections,IusethisobservationtoproposeanewtechniqueformitigatingtheeffectsofnoisevarianceonPDM,anonlineartimeseriesinferencetechnique. 3.2NoiseVarianceandPhaseDynamicsModeling 3.2.1TheImpactofNoiseVarianceonPhaseDynamicsModelingLikePDC,PDMdoesnotincorporateinitsformulationanyinformationregardingthenoisetermsofthemodelthatisrstttothedata.ThiscanbeseeninEq.( 2 )whichdescribesthemodelthatisusedinthePDMtechniquetotthetimeseriesofinstantaneousphaseincrements.Thenoisetermi(t)presentinthepreviousequation,Eq.( 2 ),isnotpresentinEq.( 2 ),noristhistermincludedinthenaldenitionofPDMgiveninEq.( 2 ).IshowinthefollowingthatthisfactresultsinPDMbeinghighlyaffectedbydifferencesinnoisevariancebetweentimeseriesmuchlikePDC. 66

PAGE 67

ConsiderthesystemoftwocoupledvanderPoloscillatorsfromEq.( 2 )with!i=0.5,!j=0.97,"ij=0and"ji=0.01sothatthereiscouplingonlyfromxi(t)toxj(t).Byxingj=0.01andvaryingi,IcanobservetheeffectsofnoisevariancedifferencesonthePDMtechnique.Thetrivialcasewherecausalityactuallyowsfromthelownoisevariancetothehighnoisevariancetimeseriesisavoidedbyensuringthatij.IperformedthisexperimentbysimulatingthecoupledvanderPolsystemfordifferentvaluesofiintherange[0.01,20]andapplyingthePDMtechniquetotheresultingtimeseries.TheresultsarepresentedinFig. 3-4 ,whereanyPDMvalueabovethesignicancethreshold(indicatedbytheblackdottedline)issignicant.Theseresultsshowthatwheni10,PDMincorrectlyidentiescausalityfromthelownoisevariancetimeseries,xj(t),tothehighnoisevariancetimeseries,xi(t).Inallcases,regardlessofthevalueofi,PDMcorrectlyidentiesthetruecausalityfromxi(t)toxj(t).TheseresultsdemonstratethatPDMishighlyaffectedbydifferencesinnoisevariancebetweentimeseries.Specically,ifthemagnitudeofthenoisevariancedifferencebetweentimeseriesislargeenough(inthiscaseji)]TJ /F11 11.955 Tf 12.58 0 Td[(jjj10)]TJ /F7 11.955 Tf 12.58 0 Td[(0.01j=9.99),PDMwillidentifycausalityfromthetimeserieswithalowernoisevariancetothetimeserieswithahighernoisevarianceregardlessofthetrueeffectiveconnectivityofthesystem. 3.2.2ANewTechniqueforMitigatingtheImpactofNoiseVarianceonPhaseDynamicsModelingInthecaseoflineartimeseriesinferencetechniques,theeffectsofnoisevariancedifferencesaremitigatedbyincludingintheformulationofthesetechniquesthenoisevariancetermsfromthelinearVARmodelttothetimeseriesdata.Inasimilarfashion,IproposeamodicationofthePDMtechniquewhichservestomitigatethenegativeimpactofnoisevariancedifferences,whichIcallgeneralizedPDM(GPDM).ConsidertheformulationofthepairwisePDMtechniquepresentedin 67

PAGE 68

Eqs.( 2 2 2 2 ).Denotetheestimatedfunction,Fi(i(t),j(t)),obtainedbydoingaleastsquaresttotheinstantaneousphaseincrementtimeseries,i(t+))]TJ /F11 11.955 Tf -458.7 -23.91 Td[(i(t),as^Fi(i(t),j(t)).Computetheerrorof^Fi(i(t),j(t))asi(t)=^Fi(i(t),j(t)))]TJ /F7 11.955 Tf 11.96 0 Td[((i(t+))]TJ /F11 11.955 Tf 11.95 0 Td[(i(t)). (3)Then,letibethesamplestandarddeviationofi(t),computedasi=vuut 1 N)]TJ /F7 11.955 Tf 11.96 0 Td[(1NXt=1(i(t))]TJ /F7 11.955 Tf 12.77 0 Td[(i)2 (3)wherei=1 NPNt=1i(t)isthesamplemeanandNisthetotalnumberofdatapointsini(t)andj(t).GPDMisthendenedasGPDMj!i=1 is Xm,``2(a2m,`+b2m,`). (3)FollowingthemultivariateformulationofPDMinEq.( 2 ),multivariateGPDMcanbedenedasGPDMj!i=1 2i(2i,j+2i,j). (3)IappliedGPDMtothesystemoftwocoupledvanderPoloscillatorsdescribedabove.TheresultsarepresentedinFig. 3-4 ,whereanyGPDMvalueabovethesignicancethresholdindicatedbytheblackdottedlineissignicant.Theseresultsdemonstratethat,inallcases,regardlessofthedifferencebetweeniandj,GPDMcorrectlyidentiescausalityfromxi(t)toxj(t)anddoesnotincorrectlyidentifycausalityfromxj(t)toxi(t).Thus,likeGPDC,GPDMdoesnotincorrectlyidentifycausalityfromalownoisevariancetimeseriestoahighnoisevariancetimeserieswhenthiscausalityisnotphysicallypresentinthesystem. 68

PAGE 69

Figure3-1. TheprobabilitythatPDCincorrectlyidentiescausalityfromtimeseriesxi(t)toxj(t)(topplot)andtheprobabilityPDCcorrectlyidentiescausalityfromxj(t)toxi(t)(bottomplot)arebothpresentedasafunctionof,thestrengthofthecausalconnectionfromxj(t)toxi(t),andj,thenoisevarianceofxj(t).Sincethenoisevarianceofxi(t)isxedati=1,jalsorepresentsthemagnitudeofthenoisevariancedifferencebetweenthetimeseries,ji)]TJ /F11 11.955 Tf 11.96 0 Td[(jj.PDChasahighprobabilityofincorrectlyidentifyingcausalityfromxi(t)toxj(t)whenislowandji)]TJ /F11 11.955 Tf 11.96 0 Td[(jjishigh. 69

PAGE 70

Figure3-2. TheprobabilitythatGPDCincorrectlyidentiescausalityfromtimeseriesxi(t)toxj(t)(topplot)andtheprobabilityGPDCcorrectlyidentiescausalityfromxj(t)toxi(t)(bottomplot)arebothpresentedasafunctionof,thestrengthofthecausalconnectionfromxj(t)toxi(t),andj,thenoisevarianceofxj(t).Sincethenoisevarianceofxi(t)isxedati=1,jalsorepresentsthemagnitudeofthenoisevariancedifferencebetweenthetimeseries,ji)]TJ /F11 11.955 Tf 11.96 0 Td[(jj.GPDChasalowprobabilityofincorrectlyidentifyingcausalityfromxi(t)toxj(t)regardlessoforji)]TJ /F11 11.955 Tf 11.95 0 Td[(jj. 70

PAGE 71

Figure3-3. TheprobabilitythatGCincorrectlyidentiescausalityfromtimeseriesxi(t)toxj(t)(topplot)andtheprobabilityGCcorrectlyidentiescausalityfromxj(t)toxi(t)(bottomplot)arebothpresentedasafunctionof,thestrengthofthecausalconnectionfromxj(t)toxi(t),andj,thenoisevarianceofxj(t).Sincethenoisevarianceofxi(t)isxedati=1,jalsorepresentsthemagnitudeofthenoisevariancedifferencebetweenthetimeseries,ji)]TJ /F11 11.955 Tf 11.96 0 Td[(jj.GChasalowprobabilityofincorrectlyidentifyingcausalityfromxi(t)toxj(t)regardlessoforji)]TJ /F11 11.955 Tf 11.96 0 Td[(jj. 71

PAGE 72

Figure3-4. TheresultsofaPDManalysisoftwocoupledvanderPoloscillatorsarepresentedwiththePDMvaluesfromoscillatoritoj(topplot)andfromoscillatorjtoi(bottomplot)asafunctionofthenoisevarianceofi,i.Sincethenoisevarianceofjisxedatj=0.01,ialsorepresentsthemagnitudeofthenoisevariancedifferencebetweentheoscillators,ji)]TJ /F11 11.955 Tf 11.95 0 Td[(jj.Theblackdottedlineindicatesthesignicancethresholdof0.024.Anyvalueabovethisthresholdindicatesasigniantcausalinteraction.PDMincorrectlyidentiescausalityfromoscillatorjtoiwhenji)]TJ /F11 11.955 Tf 11.96 0 Td[(jjj10)]TJ /F7 11.955 Tf 11.95 0 Td[(0.01j=9.99.PDMalwaysidentiesthetruecausalityfromitojregardlessofji)]TJ /F11 11.955 Tf 11.96 0 Td[(jj. 72

PAGE 73

Figure3-5. TheresultsofaGPDManalysisoftwocoupledvanderPoloscillatorsarepresentedwiththeGPDMvaluesfromoscillatoritoj(topplot)andfromoscillatorjtoi(bottomplot)asafunctionofthenoisevarianceofi,i.Sincethenoisevarianceofjisxedatj=0.01,ialsorepresentsthemagnitudeofthenoisevariancedifferencebetweentheoscillators,ji)]TJ /F11 11.955 Tf 11.95 0 Td[(jj.Theblackdottedlineindicatesthesignicancethresholdof0.15.Anyvalueabovethisthresholdindicatesasigniantcausalinteraction.GPDMneverincorrectlyidentiescausalityfromoscillatorjtoiandalwaysidentiesthetruecausalityfromitojregardlessofji)]TJ /F11 11.955 Tf 11.95 0 Td[(jj. 73

PAGE 74

CHAPTER4DETECTINGEFFECTIVECONNECTIVITYINNETWORKSOFCOUPLEDNEURONALOSCILLATORSTheimportanceofneuronaloscillatornetworksincoordinatingbrainfunctionhasbeenwell-establishedinnumerousstudies.Atthesingle-celllevel,specializedcircuitsofneuronaloscillatorshavebeenimplicatedinmemoryformation,visualprocessing,learning,andavarietyofothermotorandcognitivetasks( Buzsaki 2006 ; BuzsakiandDraguhn 2004 ; Deisteretal. 2009 ; Fries 2009 ).Ofparticularinterestistheabilityofneuronalnetworkstosynchronizetheirringactivity,sincethisbehaviorhasbeenlinkedtopathologicalbrainstatessuchasepilepsy( UhlhaasandSinger 2010 )andParkinson'sdisease( Hammondetal. 2007 ).Threekeycomponentsofneuronalnetworksthatplayacentralroleintheemergenceofpathologicalsynchronization,are:(a)thephysicalpresenceofcouplingbetweenneuronaloscillators,referredtoasstructuralconnectivity,(b)thestatisticaldependencebetweenneuronaloscillators,referredtoasfunctionalconnectivity,and(c)thecausaleffectsofoneneuronaloscillatoronanother,referredtoaseffectiveconnectivity( Spornsetal. 2004 ).TwopopulartimeseriesinferencetechniquesfordeterminingeffectiveconnectivityincoupledoscillatornetworksincludeGrangercausality(GC)( Geweke 1982 1984 ; Granger 1969 )andphasedynamicsmodeling(PDM)( RosenblumandPikovsky 2001 ).Anumberofrecentstudieshavefocusedoncomparingtheabilityofthesetechniquestocorrectlyidentifyeffectiveconnectivityinagivennetworkofcoupledoscillators( Lungarellaetal. 2007 ; Smirnovetal. 2007 ; Winterhalderetal. 2005 ).However,asystematicstudyoftheapplicabilitydomainsofthesetechniquestoneuronalnetworks,whichtakesintoaccountthecouplingstrength,thedegreeofnoiseinthenetwork,thelevelofneuronalnetworksynchronyandthetypeofsynapticinteraction(chemicalvs.electrical)isstilllacking.Inthischapter,IfocusontheseissueswithinthecontextofasimplenetworkofcoupledMorris-Lecar(ML)neuronaloscillators( MorrisandLecar 1981 ). 74

PAGE 75

IbeginbyconsideringthecaseoftwounidirectionallycoupledMLneuronsunderthreedifferentcouplingscenarios:1)lineardiffusivecouplingwhichmodelsanelectricalsynapseorgapjunction( Bennett 1997 ; ChowandKopell 2000 ),2)nonlinear,fastthresholdmodulationcouplingwhichmodelsafasttimescalechemicalsynapse( SomersandKopell 1993 ),and3)nonlinearthresholdcouplingthatmodelsaslowtimescalechemicalsynapse( Abarbaneletal. 2003 ).Foreachcouplingscheme,sampletimeseriesdatasetsaregeneratedbyvaryingthecouplingstrengths,noiselevels,andintrinsicringfrequenciesofthecoupledneuronaloscillators.Theresultisalargeensembleoftimeseriesdatasetsfromwhichtheregionsofapplicabilityofseveraltechniquesusedtoestimateeffectiveconnectivitycanbeinvestigated.Theapplicabilitydomainsofthesevarioustechniquesarethendeterminedusingdecisiontreeclassiersthataretrainedandtestedonthesetimeseriesdatasets.Theresultisametaheuristicapproachthatcanbeappliedbypractitionerswithoutanypriorknowledgeofnetworkstructureordynamicssinceallthetimeseriesfeaturesusedtodetermineapplicabilitycanbemeasureddirectlyfromthetimeseriesthemselves.Formystudy,Iconsiderthelinearspectralmethods:GC(Eqs.( 2 2 2 ))( Geweke 1982 1984 ; Granger 1969 ),partialdirectedcoherence(PDC)(Eq.( 2 ))( BaccalaandSameshima 2001 )andamodiedversionofPDCknownasthegeneralizedPDC(GPDC)(Eq.( 2 ))( Baccalaetal. 2007 ; Winterhalderetal. 2005 ).IalsoincludePDMinmyanalysissinceitisspecicallydesignedfornarrow-bandinteractingoscillatorsandcanhandleamoderateamountofnonlinearityinthedata(Eqs.( 2 2 ))( RosenblumandPikovsky 2001 ).Therestofthischapterisorganizedasfollows.First,Idescribetheneuronaloscillatormodelandnetworksusedtogeneratesampledatasets.Next,Ioutlinethemethodologyusedforgeneratingthetestdataandtheconstructionoftheclassiers.Theseclassiersarethenusedtoassesstheapplicabilityofthesetechniquesfor 75

PAGE 76

estimatingtheeffectiveconnectivityinthegivennetwork.Finally,Ipresenttheresults,adiscussionofmystudyandasummaryofthechapter. 4.1Neuron,SynapseandNetworkModels 4.1.1NeuronModelInthisstudy,IusetheMLspikingneuronmodelasageneralexampleofaminimal,conductance-basedneuronaloscillator( Izhikevich 2007 ; MorrisandLecar 1981 ).Thismodelisdescribedbythefollowingequations:Cdv dt=)]TJ /F5 11.955 Tf 9.3 0 Td[(gCam1(v)(v)]TJ /F5 11.955 Tf 11.96 0 Td[(VCa))]TJ /F5 11.955 Tf 11.95 0 Td[(gKw(v)]TJ /F5 11.955 Tf 11.96 0 Td[(VK))]TJ /F5 11.955 Tf 11.95 0 Td[(gL(v)]TJ /F5 11.955 Tf 11.95 0 Td[(VL)+Idc+Inoise+Isynapse (4)dw dt=w1(v))]TJ /F5 11.955 Tf 11.95 0 Td[(w 1(v)whereInoise=DIsynapse=kf(vpost,vpre)m1(v)=0.51+tanhv)]TJ /F5 11.955 Tf 11.95 0 Td[(V1 V2w1(v)=0.51+tanhv)]TJ /F5 11.955 Tf 11.96 0 Td[(V3 V41(v)=coshv)]TJ /F5 11.955 Tf 11.96 0 Td[(V3 2V4)]TJ /F8 7.97 Tf 6.59 0 Td[(1Herevrepresentsthetransmembranevoltageofaneuron,wrepresentstheactionofthepotassiumcurrent,Idcistheexternalinputcurrent,InoiseistheinputcurrentduetonoiseandIsynapseistheinputcurrentfromsynapticallyconnectedneurons.IntheequationforInoise,isGaussianwhitenoisewithzeromean,unitvarianceandDisthenoiseintensity.IntheequationforIsynapse,kisthestrengthofcouplingandthefunction,f(vpost,vpre),representsthecouplingtypewherevpreisthemembranevoltageofthepresynapticneuronandvpostisthemembranevoltageofthepostsynapticneuron.Thisfunctionischangedinordertomodeldifferentrealisticsynapses,specicallylineardiffusivecoupling,fastthresholdmodulationcoupling,andnonlinearsynapticthreshold 76

PAGE 77

coupling.Theremainingparametersaresetasfollows:V1=)]TJ /F7 11.955 Tf 9.29 0 Td[(1.2mV,V2=18mV,V3=12mV,V4=17.4mV,VL=)]TJ /F7 11.955 Tf 9.3 0 Td[(60mV,VK=)]TJ /F7 11.955 Tf 9.3 0 Td[(84mV,VCa=120mV,gL=2S/cm2,gK=8S/cm2,gCa=4S/cm2,=1 15s)]TJ /F8 7.97 Tf 6.59 0 Td[(1,andC=20F/cm2.Fortheseparameters,themodelexhibitsTypeIexcitability,whichmeansthatoscillationscanoccurwitharbitrarilylowfrequencyascurrentisinjectedintotheneuron( Ermentrout 1996 ). 4.1.2SynapseModelsThefollowingthreerealisticsynapticcouplingmodelsareconsidered: 1. Lineardiffusivecouplingwhichmodelsanelectricalsynapseorgapjunction( Bennett 1997 ; ChowandKopell 2000 ):f(vpost,vpre)=vpre)]TJ /F5 11.955 Tf 11.96 0 Td[(vpost (4) 2. Nonlinear,fastthresholdmodulationcouplingwhichmodelsasimpliedchemicalsynapse( SomersandKopell 1993 ):f(vpost,vpre)=0.51+tanhvpre)]TJ /F5 11.955 Tf 11.96 0 Td[(Vth Vslope(vpost)]TJ /F5 11.955 Tf 11.95 0 Td[(Vrev) (4)whereVth=0mV,Vslope=1mVandVrevrepresentsthesynapticpotential,whichissetto-84mVtomimictheeffectofinhibitorycoupling. 3. Nonlinear,thresholdcouplingwhichmodelsachemicalsynapseandaccountsforthedynamicsofthereleaseandabsorptionofneurotransmittersinthesynapticcleft( Abarbaneletal. 2003 ; Talathietal. 2010 ):f(vpost)=S(t)(vpost)]TJ /F5 11.955 Tf 11.95 0 Td[(Vrev)) (4)dS dt=S1(vpre(t)(t)))]TJ /F5 11.955 Tf 11.95 0 Td[(S(t) [S0)]TJ /F5 11.955 Tf 11.96 0 Td[(S1(vpre(t)(t))]whereVrevisdenedasinEq.( 4 ).(t)=Pi(t)]TJ /F5 11.955 Tf 12.15 0 Td[(ti)((ti+R))]TJ /F5 11.955 Tf 12.15 0 Td[(t)where()istheHeavisidestepfunctionandtiisthetimeoftheithpresynapticneuronalspike.Thetimescaleconstantgoverningreceptorbindingisrepresentedbyandissetto7.9ms,S0=1.013,andR=(S0)]TJ /F7 11.955 Tf 12.79 0 Td[(1).Finally,S1isthesigmoidalfunctiondenedasS1(v)=0.5(1+tanh(120(v)]TJ /F7 11.955 Tf 11.96 0 Td[(0.1))). 4.1.3NetworkModelIconsidersimplenetworkscomposedofeithertwoorthreecoupledMLneuronaloscillatorsinthiswork.Therstnetworkiscomprisedoftwounidirectionallycoupled 77

PAGE 78

neuronmodelswhosenetworkstructureisillustratedinFig. 4-1 A.Thesecondisanensembleof12different3-neuronnetworks,wherenetworksareconstructedbypermutingthepresenceandabsenceofthevariousconnectionsinFig. 4-1 B. 4.2ClassierConstructionMethodology 4.2.1DenitionofTimeSeriesFeaturesIbeginbydeningthetimeseriesfeaturesusedtocharacterizethedatasetsinvestigatedinthisstudy.Ithendescribehowthesefeaturesareusedtoobtainthedecisiontreesthatdeterminetheapplicabilitydomainsofthevarioustechniquesusedtoestimatingeffectiveconnectivity. 1. Phasecoherence(PC):isametricthatcanbeusedasameasureofsynchronybetweentwooscillators.Giventwotimeserieswithinstantaneousphase,i(t)andj(t),PCcanbedenedas( Mormannetal. 2000 ):PC(i,j)=p hcos(i(t))]TJ /F11 11.955 Tf 11.96 0 Td[(j(t))i2+hsin(i(t))]TJ /F11 11.955 Tf 11.95 0 Td[(j(t))i2 (4)wherehidenotestimeaveraging.Thismeasureissymmetricini(t)andj(t)andPC(i,j)=1whenthereiscompletephasesynchronybetweentheoscillators,i.e.,wheni(t))]TJ /F11 11.955 Tf 11.96 0 Td[(j(t)=constant. 2. Coefcientofvariationofinterspikeinterval(CV):isameasureofdynamicalnoiseintheneuronaloscillator.Givenneuronalmembranepotential,onecanestimatetheinter-spikeinterval(ISI)asthetimeintervalbetweenconsecutivecrossingsofthemembranepotentialwithsomethreshold.FromthedistributionofISIs,oneobtainsthemeanISIandthevariance2ISI.CVisthendenedas( Perkeletal. 1967 ):CV=ISI ISI (4)ACVofoneindicatesaPoissonprocesswhereallspikesoccurindependentlywhileaCVofzeroindicatesaseriesofperfectlyperiodicspikes.ThefollowingpairwisemeasureofCVisintroducedtoaccountforagivenpairofneurons:CVmean(i,j)=1 2(CVi+CVj) (4)CVdiff(i,j)=(CVi)]TJ /F1 11.955 Tf 11.96 0 Td[(CVj) 78

PAGE 79

3. Firingfrequency(FF):isobtainedasFF=1 ISI (4)Againintroducingthefollowingpair-wisemeasureforFFFFmean(i,j)=(FFi)]TJ /F1 11.955 Tf 11.95 0 Td[(FFj) max(FFi,FFj) (4)ThetimeseriesfeaturesgiveninEqs.( 4 4 4 )canbecalculatedfromthemembranevoltagetimeseriesofaneuronaloscillator.Thisisincontrasttomodelparameterssuchask,D,andIdcwhichareoftenhiddeninexperimentaldatasets.Previousstudieshavefocusedeitheronndingregionsofmodelparameterspacewherethetimeseriesinferencetechniquesforestimatingeffectiveconnectivityfail( Kayseretal. 2009 ; Lungarellaetal. 2007 ; Winterhalderetal. 2007 ),oronnon-dynamicalfeaturesofadatasetsuchassignal-to-noiseratioordatasetlengththatinuencetheaccuracyofagiventechniqueforestimatingeffectiveconnectivity( Astoletal. 2007 ; SmirnovandAndrzejak 2005 ).Unlikethesepastworks,mystudypresentsresultsinthespaceofthetimeseriesfeaturesoutlinedabove,makingitmoreapplicableforpractitionerswhowishtoapplythetimeseriesinferencetechniquesconsideredheretoestimateeffectiveconnectivity. 4.2.2DecisionTreeConstructionThefollowingisanoverviewoftheprocedureadoptedtoconstructdecisiontreesusingthetimeseriesfeaturesdescribedinsection 4.2.1 1. Generateanensembleoftimeseriesdatasetsbysystematicallyvaryingtheparameters,k,D,andIdcoftheMLneuronsinEq.( 4 )whicharecoupledaccordingtoFig. 4-1 A.RepeatthisforeachofthethreesynapsetypespresentedinSection 4.1.2 2. CalculatethetimeseriesfeaturesdenedinSection 4.2.1 foreachdataset. 3. EstimateeffectiveconnectivityusingGC,PDC,GPDC,andPDMforeachtimeseries. 79

PAGE 80

4. Trainaclassierforagiventimeseriesinferencetechniquebysettingthetimeseriesfeaturesaspredictorsandtheoutputgeneratedbythechosentechniqueastheresponse. 5. Testtheclassiersbygeneratinganewensembleof3-neurontimeseriesdatasets(Fig. 4-1 B)andevaluatingclassierperformanceonthisnewdata.Forstep1,atotalof1,000timeseriesdatasetsweregeneratedforeachsynapsetypebysystematicallyvaryingkofneuron2intherange[1e-4,0.5],Dofbothneuronsintherange[0,80],andIdcofbothneuronsintherange[40,85].Onlyunidirectionalcouplingswereconsidered,sok=0forneuron1.AtotalofN=51,000datapointsweregeneratedforeachtimeserieswiththerst1,000datapointsthrownouttoremovetransients.Eq.( 4 )wassolvednumericallyviathestochasticEuler-Maruyamamethodwithastepsizeoft=0.005( KloedenandPlaten 1999 ).Timeseriesfeatureswerecalculatedinstep2asdenedinSection 4.2.1 .Instantaneousphase,whichisneededtocalculatephasecoherence,wasobtainedviaHilberttransformofthevoltagetimeseries(cf.Eq.( 2 ))( Pikovskyetal. 2002 ).Thefourtechniquesfordeterminingeffectiveconnectivitywereappliedtothedatasetsinstep3.Theclassiersinstep4wereconstructedusingMatlab'sclassregtreefunctionwhichtrainsadecisiontreebasedonagivensetofpredictorsandresponses(alsoreferredtoasfeaturesandcategories,respectively)( Breimanetal. 1984 ; Hastieetal. 2001 ).Foreachmethodandcouplingtype,adecisiontreewasconstructed.Inthiswork,thepredictorsarethetimeseriesfeaturesdenedinSection 4.2.1 includingPC(i,j),CVmean(i,j),abs(CVdiff(i,j)),andabs(FFdiff(i,j))andtheresponsesaretheoutputsgeneratedbythefourtechniquesinstep3.AbsolutevaluesofCVdiff(i,j)andFFdiff(i,j)areusedfortheclassiersince,inanexperimentalsetting,thereisnoaprioriorderingofneurons.Asanexample,theGC/linearcouplingdecisiontreewasconstructedbysupplyingtheclassregtreefunctionwiththePC(i,j),CVmean(i,j),abs(CVdiff(i,j)),and 80

PAGE 81

abs(FFdiff(i,j))valuesofall1,000linearcouplingdatasetsaspredictors.Inaddition,thefunctionwasgiventhecorrespondingGCresultastheresponseforeachdataset.Responseswerelabeledaseither`1'ifGCdetectedthecorrectdirectionalityoftheinteractionor`0'ifGCfailedtodetectcorrectdirectionality.SimilardecisiontreeswereconstructedforPDC,GPDCandPDM.Decisiontreeswereusedinthisworkbecausetheyhadthelowestmisclassicationerrorcomparedtootherclassicationmethodssuchaslinearorquadraticdiscriminantanalysis,NaiveBayes,andsupportvectormachineswhenappliedtothesimulateddatasets(seeTable 4-1 andSection 4.2.3 ).Theperformanceofthedecisiontreeclassierswasveriedusinga10-foldcross-validationscheme( Schaffer 1993 ).Instep5,Itestedtheapplicabilityofthedecisiontreeclassierstothegeneralclassofamulti-nodeneuronalnetworkbyrstgeneratinganewensembleof3-neurontimeseriesdatasetsaccordingtothenetworkstructureinFig. 4-1 B.ValuesfortheMLmodelparametersk,D,andIdcwererandomlyselectedwithintheaforementionedrangesand1,000newtimeseriesdatasetsweregeneratedforeachcouplingtypeusingthepreviouslydescribedmethodology.Thepresenceanddirectionallyofthepossiblecouplingsinthenetworkwassetrandomly.Next,Icalculatedasetofpredictorsfromthisnewdatathatcouldbeclassiedbythepreviouslyconstructeddecisiontrees.Onceagain,thetimeseriesfeaturesdenedinSection 4.2.1 wereusedaspredictors.Inthiscase,timeseriesfeatureswerecalculatedforeachnewtesttimeseriesdatasetbyconsideringonlypairwiseinteractions.Thus,eachtimeseriesdatasetconsistedofthreevaluesforeachtimeseriesfeature,i.e.,PC(1,2)capturesthephasecoherencebetweenneuron1andneuron2,PC(1,3)thephasecoherencebetweenneuron1and3,andPC(2,3)thephasecoherencebetweenneuron2and3,withtheremainingtimeseriesfeaturesspeciedinasimilarfashion.Forinstance,givenasingledataset,threepredictorswereconstructed:oneconsistingofthePC(i,j),CVmean(i,j),abs(CVdiff(i,j)),and 81

PAGE 82

abs(FFdiff(i,j))valuesbetweenneurons1and2,asecondconsistingofthesetimeseriesfeaturesforneurons1and3,andathirdwiththesetimeseriesfeaturesforneurons2and3.Withanensemblesizeof1,000,atotalof3,000predictorswasconstructed.Thedecisiontreeclassierswerethenappliedtothisnewsetoftestpredictors,yieldingasetofclassierresponses.Theseclassierresponses,labeledeither`1'or`0',indicatedwhetheraparticulartimeseriesinferencetechniquewaspredictedtobecorrectorincorrectbyitscorrespondingdecisiontree.Finally,IcalculatedtheactualresponsesobtainedwhenIappliedeachofthefourtechniques,GC,PDC,GPDCandPDM,tothenew3-nodenetworkdatasets.PDCandGPDCwereappliedasdescribedinEqs.( 2 2 )sincetheseformulasimplicitlyhandlenetworkswithmorethantwonodes.MultivariateformulationsofGCandPDMwereusedinordertoproperlyhandlethismultivariatedata.Specically,conditionalspectraldomainGCwiththepartitionmatrixmethodgivenbyEq.( 2 )wasusedalongwithmultivariatePDM,givenbyEq.( 2 ).Theresultsofthefourtechniquesfordeterminingeffectiveconnectivitywerelabeledaseither`1'or`0,'denotingthatthetechniquewaseithercorrectorincorrect,respectively.Theclassierresponseswerethencomparedwiththeactualresponsesofagiventimeseriesinferencetechniqueinordertocalculatethetree'saccuracy.TheexperimentalproceduredescribedaboveisillustratedinFig. 4-2 4.2.3ResultingDecisionTreesDecisiontreeswereconstructedforeachofthefourtechniquesfollowingthemethodologydescribedinSection 4.2.2 .Inasimilarfashion,Ialsoconstructedclassiersbasedonlinearandquadraticdiscriminantanalysis,NaiveBayes,andsupportvectormachines.Iused10-foldcross-validationoneachoftheresultingclassiersinordertoestimatethetesterror,whichistheexpectedmisclassicationerrorofaclassierwhenitisappliedtodatathatisindependentoftheoriginaltraining 82

PAGE 83

data( Schaffer 1993 ).Theresults,presentedinTable 4-1 ,showthatthedecisiontreeclassierhadthelowesttesterroroftheveclassiersinvestigated.Therefore,Ichosedecisiontreeclassiersforthisstudy.Afterconstructingdecisiontreesforeachcouplingtype,itwasfoundthatthebasicstructureofthedecisiontrees,suchasthefeaturesoneachbranchandtheirordering,wasthesameacrossallthreetypesofcouplingthatwereconsidered.However,thespecicfeaturevaluesoneachdecisionbranchpointdidvaryamongthedifferentcouplingtypesforeachtecnique.AnexampleofthisisshowninFig. 4-3 forGC.Ratherthanpresentseparatedecisiontreesforeachcouplingtype,Ichosetoaveragethefeaturevaluesatthesebranchpointsinordertogetacompositedecisiontreeforeachofthefourtechniquesthatcouldbeusedregardlessofthetypeofcouplingbetweenneuronaloscillators.Sinceapractitionermaynotknowtheexactformofcouplingbetweenneuronsinagivenexperimentaldataset,thisapproachisalsothemostapplicableinpractice.ThecompositetreesareshowninFig. 4-4 andwereusedtoderivethefollowingresults.Eachdecisiontreecontainsonlyasubsetoftheoriginalfourtimeseriesfeaturesusedaspredictorsforgeneratingthetrees.Thisprovidesinsightintothepredictivefeaturesofagivendatasetthatarecriticalforeachofthefourtechniquesintermsoftheirabilitytocorrectlydetermineeffectiveconnectivityinthenetwork.ThePDCtreeisthemostextremeexampleasitonlyrequiresPC(i,j)topredictwhetherornotthemethodwillwork.TheresultthatPDCislikelytofailwhenPC(i,j)0.69correspondstotheresultsobtainedinSmirnovetal.'sstudywhichindicatedthatPC(i,j)valuesabove0.75or,insomecases,above0.5weredetrimentaltoPDC( Smirnovetal. 2007 ).ThetreesconstructedforGCandGPDCarealmostidenticalduetothefactthatthesetechniquesnearlyalwayspredictthesameeffectiveconnectivityfor2-nodenetworks.Forinstance,outofthe1,000,2-nodetrainingdatasetsconsidered,GCand 83

PAGE 84

GPDCagreedonthedirectionalityofinteraction97.0%,97.4%,and97.8%ofthetimeforlinear,fastthresholdmodulation,andsynapticcoupling,respectively.Thus,forbothtechniques,thedecisiontreesindicatethatPC(i,j)andCVmean(i,j)arecriticalfactorsfordeterminingcorrecteffectiveconnectivity.ThefactthatCVmean(i,j)mustbeequalorgreaterthan0.022inthecaseofbothGCandGPDC,ismostlikelyareectionofthefactthatbothtechniquesrequiresomeamountofdynamicalnoisebepresentinthesysteminorderforthemtowork.TherequirementthatPC(i,j)mustbelessthan0.92or0.87forGCandGPDC,respectively,restrictsthelevelofsynchronizationthatcanexistbetweentwosignalsandisanexpectedrequirementbasedonpreviousstudiesofthesetechniques( Smirnovetal. 2007 ).Incontrasttotheothertechniquesconsidered,PDMisable,tosomeextent,todetecteffectiveconnectivityacrossdifferentfrequenciesduetothemandntermsinEq.( 2 )( SmirnovandBezruchko 2009 ).Infact,variationintheintrinsicringfrequenciesbetweenneuronaloscillatorsisbenecialtothetechniquebecauseitresultsina1(t)versus2(t)trajectorythatllsupthephasespace( RosenblumandPikovsky 2001 ).OnlyinthiscasecanthefunctionF(i(t),j(t))inEq.( 2 )beproperlyestimatedwithi(t)andj(t)asindependentvariables.ThisisillustratedinFig. 4-5 wherethe1(t)versus2(t)phasespacetrajectoriesfortwodatasets,onewithahighabs(FFdiff(i,j))andonewithalowabs(FFdiff(i,j)),areshown.DespitepreviousobservationsthatPDMaccuracyishighlydependentonPC(i,j)( Smirnovetal. 2007 ),myresultsindicatethatPC(i,j)isnotascriticalafactorindeterminingwhetherthetechniquewillfailasothermetricssuchasFFdiff(i,j)andCVmean(i,j).Thesensitivityandspecicityofeachdecisiontreewascalculatedasfollows:sensitivity=#oftruepositives #oftruepositives+#offalsenegatives (4) 84

PAGE 85

specicity=#oftruenegatives #oftruenegatives+#offalsepositives (4)wheretruepositivesinthiscaseoccurwhenthedecisiontreelabelsadatasetas`1'andthetimeseriesinferencetechniqueproducesacorrectresult,truenegativesoccurwhenthedecisiontreelabelsadatasetas`0'andthetechniqueproducesanincorrectresult,falsepositivesoccurwhenthedecisiontreelabelsadatasetas`1'andthegiventechniqueproducesanincorrectresult,andfalsenegativesoccurwhenthedecisiontreelabelsadatasetas`0'andthegiventechniqueproducesacorrectresult.Asensitivityof100%meansthedecisiontreeidentiesallcaseswhenthetechniqueiscorrectandaspecicityof100%meansthedecisiontreeidentiesallcaseswhenthetechniqueproducesincorrectresults.Compositetreesforeachtechniquewereusedalongwith10-foldcross-validationtocalculatesensitivityandspecicityvalues.TheresultsarepresentedinTable 4-2 .Inlightoftheseresults,Icannowcommentonthechoiceofusingonlyunidirectionalcouplinginthetrainingdataensemble.Whentimeseriesdatasetswithbidirectionalcouplingwereincludedinthetrainingensemble,thesensitivityofthedecisiontreesdroppedsignicantly.However,whenIapplieddecisiontreesthatwerebasedonlyonunidirectionalcouplingtodatasetswithbidirectionalcoupling,thesensitivitywasnotaffected.Thisindicatesthatthespaceoftimeseriesfeatureswhereunidirectionalcouplingcanbedetectedismuchmorerestrictivethanthoseregionswherebidirectionalcouplingcanbedetected.Thereasonforthisisthatbidirectionalcouplingbiasesthedecisiontreestowardhighvaluesofphasecoherencesince,regardlessofactualcouplingdirectionality,ahighphasecoherencevaluealmostalwaysresultsinthetimeseriesinferencetechniquespredictingbidirectionalcouplingbetweenneurons.Thus,ifIincludedbidirectionalconnectionsinthetrainingdata,thedecisiontreeswouldfavordatasetswithahighphasecoherencevalueeventhoughthisis 85

PAGE 86

generallydetrimentalfordetectingunidirectionalcoupling.AsaresultIdecidedtouseonlyunidirectionallycoupledMLnetworkdatasetstoconstructtheclassiers.Sincethisanalysiswasperformedonnetworksof2nodes,thequestionremainsastotheutilityofthedecisiontreesderivedaboveforanalyzingtheapplicabilityoftimeseriesinferencetechniquetolargernetworks.IaddressthisquestioninthenextsectionwhereIdemonstratetheutilityofmydecisiontreeclassiersinthecaseof3-nodenetworks. 4.2.4ExtensiontoLargerNetworksIappliedtheclassierdecisiontreesofFig. 4-4 totimeseriesdatasetsgeneratedfromanensembleof3-neuronnetworks.Agivendatasetfromthisensemblewasclassiedatotalofthreetimesbythedecisiontree,onceforeachofitsthreepairwiseinteractions,withthefollowingresults:adatasetcouldeitherhaveallofitspairwiseinteractionsclassiedas`1',twopairwiseinteractionsclassiedas`1'andoneas`0',onepairwiseinteractionclassiedas`1'andtwoas`0',orallthreeclassiedas`0.'Multivariatetimeseriesinferencetechniqueswereappliedtoeachdataset.Ithencalculatedthelikelihoodthatagiventechniquecorrectlyidentiedthedirectionalityofallinteractionsinthe3-neuronnetworkgiventhenumberofpairwisenetworkinteractionsclassiedas`1'bythedecisiontree.Forexample,considera3-neuronnetworkwithcouplingfromneuron1to2andfromneuron3to1.Giventhatonlytwoofthethreepairwiseinteractionsinthisnetworkwereclassiedbythedecisionas`1',Icalculatedthelikelihoodthatatechniquewillcorrectlyidentifythecorrectnetworkcouplingfromneuron1to2and3to1.TheselikelihoodsarepresentedinTable 4-3 .Compositetreesforeachmethodwereusedforclassication.Inorderforallofthefourtechniquesconsideredheretobecorrectwithprobabilitygreaterthan50%,amajorityofthepairwiseinteractionsinthenetworkmustbeclassiedbythedecisiontreeas`1.'ThisresultisshowninthethirdcolumnofTable 4-3 whichgivestheaccuracyofeachtechniquegiventhattwoofthethreepairwise 86

PAGE 87

interactionsare`1.'TheseresultsarenotsignicantlydifferentfromthesensitivityvaluesinTable 4-2 .Ontheotherhand,accuracypercentagesforthetechniqueswhenallthreepairwiseinteractionsareclassiedas`0'ina3-nodenetworkarepresentedintherstcolumnofTable 4-3 .Inthiscase,theprobabilityofanygiventechniquesucceedingislessthan20%.Forinstance,theprobabilitythatGCgivescorrectresultsinthiscaseisonly166.6%.GiventheseaccuraciesintherstcolumnofTable 4-3 ,noneofthesetechniquesshouldbeappliedifallpairwiseinteractionsinadatasetarecategorizedas`0'bythedecisiontreesince,withhighprobability,theresultswillinfactbeincorrect. 4.3RecommendationsforDecisionTreeClassierUseinNetworksofCoupledNeuronalOscillatorsNumerousstudieshaveappliedlineartimeseriesinferencetechniquessuchasGCandPDCornonlinearmethodssuchasPDMtoneuraldatasetsinanattempttogaininsightintothedirectionalityofinteractionsinthebrain( Brovellietal. 2004 ; Cadotteetal. 2010 ; Chenetal. 2006 ; Havliceketal. 2010 ; Liaoetal. 2010a ; Satoetal. 2009 ).Datasetsaregenerallyconsideredamenabletosuchananalysisaslongastheyarestationary.Theresultsofthisstudydemonstratethenecessityforanalyzingtimeseriesdatasetsbeyondjustdeterminingtheirstationarity.Synchronyisoneexampleofatimeseriesfeaturethatcanresultinthefailureofmanytechniquesandpreviousworkhasindicatedapproximateboundsonthedegreeofphasecoherencebetweentwooscillatorsabovewhichseveralofthesetechniquestendtofail( Bezruchkoetal. 2003 ; Smirnovetal. 2007 ).Iextendedtheseresultsbyreningthelimitsonphasecoherenceaswellasconsideringlimitsonothertimeseriesfeatures,suchasnoiseandringfrequency,forthreeofthemostwidelyusedlinearmethodsfordeterminingeffectiveconnectivity,GC,PDCandGPDC,aswellasPDM,whichcanhandlesomenonlinearities.Thisistherstsystematiccomparativestudyofthesefourtechniquesastheyapplytonarrow-bandoscillatorydata. 87

PAGE 88

Inaddition,Iconsideredseveralrealisticmodelsforneuronalsynapses.Theobservedconsistencyinthederiveddecisiontreesregardlessofcouplingtypedemonstratesthatthesendingscanbeappliedtotimeseriesdatasetsevenwhenthecouplingtypeisunknown.Importantly,thefeaturesIconsideredareobservablefromthetimeseriesoftheoscillatorsthemselveswhichmeansthatpractitionerscanreadilyapplytheseresultstotheirexperimentaldatasets.Thisisincontrasttostudieswhichinvestigatetheapplicabilityoftimeseriesinferencetechniquesintermsofmodelparameters,suchascouplingstrength,whicharenotasevidentfromatimeseriesalone( Kayseretal. 2009 ; Lungarellaetal. 2007 ; Winterhalderetal. 2007 ).Thecentralcontributionofthischapteristheintroductionofamethodologytoassesswhetheragiventechniquefordeterminingeffectiveconnectivitywillproducecorrectresultswhenappliedtoagivendatasetbasedontheobservable,dynamicalcharacteristicsofthedataset.Theproposedmethodologyconsistsofdecisiontreeclassiersthatcanbeappliedwithoutanypriorknowledgeofnetworkstructure.Whenapplyingtechniquesfordeterminingeffectiveconnectivitytosmallnetworksof2-3nodes,Irecommendthatpractitionersrstusethedecisiontreesderivedinthisworktodetermineaclassicationforeachofthepairwiseinteractionsintheirnetwork.Ifamajorityoftheseinteractionsareclassiedas`0'bythedecisiontreeforagiventechnique,thetechniqueshouldnotbeappliedsincetheresultswillmostlikelybeincorrect.Ifamajorityareclassiedas`1',ontheotherhand,thetechniquecanbeappliedwithanexpectedprobabilityofsuccessgreaterthan66%.Extendingthisresulttoarbitrarilylargenetworksisbeyondthescopeofmywork.However,resultsfrom3-nodenetworksstronglysuggestthatforanynetwork,regardlessofsize,ifallpairwiseinteractionsareclassiedas`0'bythedecisiontreeofagiventechnique,thenthetechniqueshouldnotbeappliedtothedatasinceresultswillbeincorrectwithhighprobability.Itislikelythecasethatanevenlessrestrictivestatementcanbemade,namelythatthesetechniqueswillhaveverylowsuccessratesaslongasamajorityof 88

PAGE 89

thepairwiseinteractionsinanetworkareclassiedas`0.'Icautionagainsttheuseofagiventechniquefordeterminingeffectiveconnectivityinsuchacase.Insummary,thefollowingkeyconclusionscanbedrawnfromthischapter: Thedecisiontreespresentedprovideamethodologyfordeterminingtheapplicabilityofagiventimeseriesinferencetechniquetoagivensetoftimeseriesdata. Thisdecisiontreemethodologycanbeappliedwithoutanypriorknowledgeofnetworkstructureordynamicssinceallthetimeseriesfeaturesusedtodetermineapplicabilitycanbemeasureddirectlyfromthetimeseriesthemselves. Inthecaseofsmall,2-3nodenetworks,myresultscanbeapplieddirectly. Forlargernetworks,Irecommendthatagiventechniquenotbeappliedifamajorityofitspairwiseinteractionsareclassiedas`0'bythedecisiontreeforthattechnique. IncontrasttoGC,PDCandGPDC,differencesinintrinsicringfrequenciesbetweentwotimeseriesisbenecialforPDMsinceitresultsinamoredispersedinstantaneousphasetrajectory. 89

PAGE 90

Table4-1. Averagetesterrorofvariousclassierstrainedonsimulateddatasets.Testerrorisaveragedoverthefourtimeseriesinferencetechniquesandthreecouplingtypes. ClassierTypeTesterror Lineardiscriminantanalysis35.5%Quadraticdiscriminantanalysis35.2%NaiveBayeswithkerneldensityestimation26.0%Supportvectormachinewiththirdorderpolynomialkernel26.6%Decisiontree24.7% Table4-2. Decisiontreesensitivityandspecicityvalueswith95%condenceintervalsgiveninparentheses. sensitivityspecicity GCTree72.0(7.7)%81.7(3.1)%GPDCTree74.5(7.9)%81.5(3.0)%PDCTree66.4(4.5)%81.7(3.9)%PDMTree80.5(4.4)%77.0(3.3)% Table4-3. Likelihoodthatagiventimeseriesinferencetechniquecorrectlyidentiescausalityinanetworkgiventhenumberofpairwiseinteractionsclassiedas`1'bythedecisiontree.Ninetyvepercentcondenceintervalsaregiveninparentheses. 0outof3`1'1outof3`1'2outof3`1' GC12(3.3)%20(3.5)%67(7.0)%GPDC12(4.4)%43(4.6)%66(8.3)%PDC16(6.6)%46(4.0)%74(3.9)%PDM18(5.7)%25(2.4)%84(5.4)% 90

PAGE 91

Figure4-1. ThenetworkstructureofcoupledMLneuronaloscillatorsconsideredinthischapter.A)Neurons1and2areunidirectionallycoupledwithNeuron1drivingNeuron2.B)Neurons1,2and3arerandomlyconnectedwithdashedlinesindicatingpossiblecouplingbetweenneurons 91

PAGE 92

Figure4-2. Aowchartillustratingtheexperimentaldesignofmystudy.NumberscorrespondtothelistinSection 4.2.2 92

PAGE 93

Figure4-3. DecisiontreesconstructedforGCacrossthreedifferentcouplingtypesconsidered.Alltreessharethesamestructurebutvaryintheexactpredictorvaluesateachdecisionpoint. 93

PAGE 94

Figure4-4. Compositedecisiontreesforalltimeseriesinferencetechniques.Thesetreeswereconstructedbyaveragingthepredictorvaluesateachdecisionpointoverallcouplingtypesforeachtechnique. 94

PAGE 95

Figure4-5. Thetrajectoriesoftwoinstantaneousphases,1(t)versus2(t),fromdatasetswithlinearcouplingandA)abs(FF(i,j)diff)=2.4e-4,B)abs(FFdiff(i,j))=0.4.Alargerabs(FFdiff(i,j))valueresultsinthetrajectoryllingupmoreofthephasespace. 95

PAGE 96

CHAPTER5DETERMININGEFFECTIVECONNECTIVITYINNERUALMASSMODELNETWORKSInthepreviouschapter,Iconsideredtheproblemofeffectiveconnectivitydeterminationinsystemsofspikingoscillators.Suchoscillatorsarecommonlyusedtomodelbraindynamicsatthesingle-neuronscale( Izhikevich 2007 ; MorrisandLecar 1981 ).Inpractice,dataatthesingle-neuronscalecanberecordedbypatchclampingindividualcellsorthroughtheuseofmultielectrodearraysimplantedinvivo( Hamilletal. 1981 ; Potter 2001 ).Duetothespecializedandinvasivenatureofsuchrecordingtechniques,however,dataatthesingle-neuronscaleisnotaswidelyavailableaselectroencephalogram(EEG)orfunctionalmagneticresonanceimaging(fMRI)data.EEGsandfMRIsrecorddataattheneuronalpopulationscale,capturingthecombinedbehaviorofmillionsofindividualinteractingneuronsthatmakeupvariousregionsofthebrain( Kandeletal. 2000 ).Thenoninvasivenatureofthesetoolshasmadethempopularforstudyingthebrain'selectricalactivityandasaresult,thereisnoshortageofEEGandfMRIdataavailableforanalysis.NumerousstudieshavemadeuseofthisEEGandfMRIdatatoanalyzeeffectiveconnectivitybetweenbrainregions( Davidetal. 2008 ; Kiebeletal. 2009 ; LiandOuyang 2010 ; Ludtkeetal. 2010 ).Forinstance,timeseriesinferencetechniquesincludingGrangercausality(GC),partialdirectedcoherence(PDC),thedirectedtransferfunction(DTF),aswellasmanyothershaveallbeenappliedtoEEGandfMRIdatainanattempttouncoverthedirectedowofinformationbetweenregionsofinterestinthebrain( Brovellietal. 2004 ; Cadotteetal. 2010 ; Chenetal. 2006 ; Havliceketal. 2010 ; Liaoetal. 2010a ; Satoetal. 2009 ).However,fewstudieshaveexploredthevalidityofsuchtechniqueswhenappliedtononlinearandpotentiallybroadbandtimeseriesdatafromEEGandfMRIrecordings.Infact,somecriticismhasbeenleveledagainsttheapplicationofthesetechniquestofMRIdatasinceanfMRImeasuresthehemodynamic(bloodow)responsetoneuralactivityinthebrainratherthantheelectricalsignalsof 96

PAGE 97

theneuronsthemselves( Davidetal. 2008 ; Friston 2009b ).Becauseofthis,observedfMRIdatamayviolatetheassumptionthatthereareuniformresponseandrecordingdelaysinsystemvariables,onwhichthesetimeseriesinferencetechniquesarebased(cf.Chapter 1 ).Methodsforreducingtheeffectsofnon-uniformdelayscanbeemployedtomitigatethisproblembutarebeyondthescopeofmywork( Roebroecketal. 2005 ).Instead,inthischapter,IwillfocusontheapplicationoftimeseriesinferencetechniquestoEEGtimeseries.WhiletimeseriesinferencetechniqueshavebeenwidelyappliedtoexperimentallyrecordedEEGtimeseries,itisdifculttovalidatetheresultsobtainedsince1)thetruestructuralconnectivityofthebrainisrarelyknownwithabsolutecertaintyand2)thepossibleconnectivitycongurationsthatcanbeexploredarelimitedtothosethatareexperimentallyobserved.AmorerigorousapproachtovalidatingthesetechniquesinvolvessimulatingamodelofinterconnectedneuronalpopulationsthatiscapableofreproducingobservedEEGsignals.Becausethestrengthanddirectionalityofconnectionsinthismodelcanbesetbyhand,thetrueconnectivityofthemodelwouldbeknownpriortotheapplicationoftimeseriesinferencetechniques.Inthisway,theresultsofthesetechniquescouldbecomparedagainstknownconnectivity.Furthermore,thismodelwouldallowfortheexplorationofawiderangeofconnectivitycongurationssothatthetimeseriesinferencetechniquescouldbeevaluatedunderanumberoffeasiblescenarios.AssumingthatthemodelchosentoreplicatetheEEGsignalsisasatisfactoryrepresentationoftheneuronalpopulationsgeneratingtheEEGsignals,suchanapproachwouldservetodemonstratethevalidityofapplyingtimeseriesinferencetechniquestoEEGdata.Infact,severalsuchmodelsofneuronalpopulationsarealreadyinuse.Inthemid-1970s,LopesdaSilvaetal.andFreemanproposedmassactionmodelsofcorticalcolumnsinthebrain,whichwereshowntoreplicatesomeaspectsofEEGsignals( Freeman 1978 ; daSilvaetal. 1976 ).Thesemodelsweresubsequentlyimproved 97

PAGE 98

uponinworkbyJansenandRitandWendlingetal.andbecameknownasneuralmassmodels(NMMs)( JansenandRit 1995 ; Wendlingetal. 2000 ).NMMshavebeenusedtosimulateawiderangeofEEGfeatures,includingalpharhythms( JansenandRit 1995 ),synchronizationinthegammaband( SchillenandKonig 1994 ),andepilepticpatterns( Wendlingetal. 2002 ).Morerecently,NMMshavebeenshowncapableofreproducingtheentirespectrumofEEGsignals( DavidandFriston 2003 ).ThefactthatNMMsnotonlyproduceoutputthatisrepresentativeofrecordedEEGsignalsbutarealsobasedonthephysicalstructureofneuronalpopulationsinthebrain,makesthemexcellentcandidatesforinvestigatingtheapplicationoftimeseriesinferencetechniquestoEEGsforthepurposeofdeterminingeffectiveconnectivity.However,onlyafewstudieshaveconsideredNMMsforthispurpose.Forinstance,Astoletal.appliedseveraltimeseriesinferencetechniquesinthespectraldomainincludingthedirectedtransferfunction(DTF)anditsmodicationknownasthedirectDTF(dDTF)alongwithpartialdirectedcoherence(PDC)toNMMs( Astoletal. 2007 ).Thegoaloftheirworkwastodeterminetheeffectofsignal-to-noiseratio(SNR)anddatalength(DL)ontheeffectiveconnectivityestimatesproducedbythesetimeseriesinferencetechniques.However,sinceAstoletal.wereconcernedsolelywiththeeffectsofSNRandDL,theyonlyconsideredthemagnitudeoftheeffectiveconnectivityestimates.Forexample,toinvestigatetheimpactofSNR,theylookedatthemagnitudeoftheoutputproducedbyeachtechniqueasafunctionofSNRanddeterminedthatmagnitudeandSNRwereindeedcorrelated.Thisapproachdidnotconsider,however,theaccuracyofeachtechniquesincenothresholdsofsignicanceweretakenintoaccount.Thus,thisstudydidnotinvestigatethevalidityofthesetechniqueswhenappliedtoNMMs.Similarly,LiandOuyangcomparedtheapplicationofGrangercausality(GC)andtheirproposedmethodofpermutationconditionalmutualinformationtoNMMsbutonlyconsideredthemagnitudeoftheoutputproducedbythesetechniques( LiandOuyang 98

PAGE 99

2010 ).Thus,whiletheywerealsoabletodemonstratetherelativeeffectsofSNRandDLonthesetechniques,theycouldnotdrawconclusionsabouttheaccuracyofGCingeneralwhenappliedtoNMMs.AninvestigationintotheaccuracyoftimeseriesinferencetechniquesastheyapplytoeffectiveconnectivitydeterminationinNMMshasyettobeexploredintheliteraturebutisofcriticalimportancesincethesetechniquesareoftenusedtoanalyzeexperimentalEEGdata( Brovellietal. 2004 ; Cadotteetal. 2010 ; Chenetal. 2006 ).Furthermore,thefactthatmanyofthesetechniquesarelinear,includingGCandPDC,callsintoquestiontheirapplicabilitytononlinearEEGsignals.Evenanonlineartechniquesuchasphasedynamicsmodeling(PDM),maynotbeappropriatetoapplytoEEGdatasincePDMrequiresnarrow-bandinputandmanyEEGsignalsarespreadacrossarangeoffrequencies( RosenblumandPikovsky 2001 ).Inthefollowingstudy,IinvestigatetheaccuracyofseveraltimeseriesinferencetechniquesastheyapplytoeffectiveconnectivitydeterminationinNMMnetworks.Specically,IconsidertheaccuracyofGC,generalizedPDC(GPDC)andPDMwhenappliedtoanNMMnetworkintroducedbyZavagliaetal.( Zavaglia 2008 ),whichinturnisbasedonanNMMproposedbyWendlingetal.( Wendlingetal. 2000 ).IbeginwithasimplecasewhereeachneuronalpopulationintheNMMnetworkexhibitsrhythmicactivityatasingledominantfrequency.UsingthisNMMnetwork,IshowthatGPDCisabletoaccuratelyrecoverthetrueconnectivityofthemodelwithhighprobability.Next,Iconsiderthecaseofinterconnectedneuronalregionsofinterest(ROIs),eachofwhichcontainsseveralNMMswithdifferentdominantrhythmicfrequencies( Ursinoetal. 2007 ; Zavagliaetal. 2006 ).ThismodelismorecomplexsinceeachROIexhibitsrhythmicactivityoverarangeoffrequencyvalues.Inthiscase,Ishowthatnoneoftheaforementionedtimeseriesinferencetechniquesaccuratelyrecoversthetrueconnectivityofthemodel. 99

PAGE 100

InordertoaddresstheobservedinadequaciesofthetimeseriesinferencetechniquesthatIconsidered,IproposeanewtechniquebasedonPDMandtheempiricalmodedecomposition(EMD)thatiscapableofaccuratelyrecoveringeffectiveconnectivityinanROI-basedNMM.Inthefollowing,IdescribetheapplicationofmyproposedtechniqueanddemonstrateitseffectivenessonanROI-basedNMM. 5.1NeuralMassModelsInthefollowing,Iconsidertheneuralmassmodel(NMM)introducedbyZavagliaetal.( Zavaglia 2008 ),whichisamodiedversionoftheoneoriginallyproposedbyWendlingetal.( Wendlingetal. 2000 ).Thismodelconsistsoffourinterconnectedneuronalgroupsnamelypyramidalneurons,excitatoryinterneurons,inhibitoryinterneuronswithslowsynaptickinetics,andinhibitoryinterneuronswithfastsynaptickinetics.Neuronsarelumpedtogetherineachgroupanditisassumedthattheysharethesamemembranepotential.Therearenodendritesorintrinsicconductancesinthemodel,insteadlumpedrepresentationsofneuronsineachgroupcommunicateviatheaverageringrateoftheentiregroupofneurons.ThesinglepopulationNMMdescribedaboveisillustratedinFig. 5-1 .Eachneuronalgroupisrepresentedbyemboldenedrectangles.Thesignals,yi(i=0,1,2,3),owingintoeachoftheseneuronalgroupsrepresenttheaveragepostsynapticmembranepotentialofanotherneuronalgroup.Withineachgroup,thesemembranepotentialsarerstconvertedintoanaveragepresynapticspikedensityviaastaticsigmoidalfunction.Theresultingaveragepresynapticspikedensityisthenconvertedbackintopostsynapticmembranepotentialviaasecondorderlineartransferfunction.Thesetransferfunctionshaveimpulseresponseshe,hi,andhg,representingthesynapticeffectofexcitatory(pyramidalandexcitatoryinterneurons),slowinhibitoryinterneuronsandfastinhibitoryinterneurons,respectively.Thesystemofdifferentialequationsdescribingthismodelareasfollows: 100

PAGE 101

PyramidalNeuronsdy0(t) dt=y5(t)dy5(t) dt=Aa1z0(t))]TJ /F7 11.955 Tf 11.95 0 Td[(2a1y5(t))]TJ /F5 11.955 Tf 11.96 0 Td[(a21y0(t)z0(t)=2e0 1+er(s0)]TJ /F9 7.97 Tf 6.58 0 Td[(v0)v0(t)=C2y1(t))]TJ /F5 11.955 Tf 11.95 0 Td[(C4y2(t))]TJ /F5 11.955 Tf 11.96 0 Td[(C7y3(t)ExcitatoryInterneuronsdy1(t) dt=y6(t)dy6(t) dt=Aa1z1(t)+p(t) C2)]TJ /F7 11.955 Tf 11.95 0 Td[(2a1y6(t))]TJ /F5 11.955 Tf 11.96 0 Td[(a21y1(t)z1(t)=2e0 1+er(s0)]TJ /F9 7.97 Tf 6.58 0 Td[(v1)v1(t)=C1y0(t)SlowInhibitoryInterneuronsdy2(t) dt=y7(t)dy7(t) dt=Bb1z2(t))]TJ /F7 11.955 Tf 11.96 0 Td[(2b1y7(t))]TJ /F5 11.955 Tf 11.95 0 Td[(b21y2(t)z2(t)=2e0 1+er(s0)]TJ /F9 7.97 Tf 6.58 0 Td[(v2)v2(t)=C3y0(t)FastInhibitoryInterneuronsdy3(t) dt=y8(t)dy8(t) dt=Gg1z3(t))]TJ /F7 11.955 Tf 11.95 0 Td[(2g1y8(t))]TJ /F5 11.955 Tf 11.96 0 Td[(g21y3(t)z3(t)=2e0 1+er(s0)]TJ /F9 7.97 Tf 6.58 0 Td[(v3)v3(t)=C5y0(t))]TJ /F5 11.955 Tf 11.95 0 Td[(C6y2(t) (5) 101

PAGE 102

Intheaboveequations,vi(i=0,1,2,3),representstheaveragemembranepotentialsofthefourdifferentgroups.Thesesignalsaresuppliedasinputstothesigmoidfunctionswhichconvertthemintoaveragespikedensities,representedbyzi(i=0,1,2,3).Theaveragespikedensities,inturn,entersynapses,eachofwhichisdescribedbyanaveragegainfactorandatimeconstant.TheaveragegainsareA,B,andGandthecorrespondingtimeconstantsarea1,b1,andg1forexcitatory,slowinhibitoryandfastinhibitorysynapses,respectively.Thesynapseoutputsareagainpostsynapticmembranepotentialsrepresentedbyyi(i=0,1,2,3).InteractionsamongneuronsarecapturedbythesevenconnectivityconstantsCi(i=1,...7)whichrepresenttheaveragenumberofsynapticcontacts.Finally,p(t),aGaussianwhitenoisesignalwithmean,,andvariance2,representsallexogenousinputstothemodelincludingtheaveragedensityofafferentactionpotentialsandexcitationfromexternalsources.Byvaryingthesynaptickineticsofthevariousgroupsofneuronsinthismodel,onecanchangetheintrinsicringrhythmofthepopulationasawhole.Forexample,theparametersA,B,G,a1,b1,andg1ofthismodelcanbetunedinordertomaketheNMMasawholeexhibitanintrinsicringrhythmatalow,mediumorhighfrequency.Morespecically,Zavagliaetal.identiedvaluesfortheseparameterssuchthatthepopulationexhibitedeithertheta/alphabandrhythms(4-12Hz),betabandrhythms(12-30Hz),orgammabandrhythms(>30Hz)( Zavaglia 2008 ).TheseparametervaluesaregiveninTables 5-1 and 5-2 .ThepopulationscorrespondingtothesevariousrhythmsaredenotedasLFforlowfrequency,MFformediumfrequency,andHFforhighfrequency,respectively.Thepowerspectraldensity(PSD)ofanLF,MF,andHFpopulationisillustratedinFig. 5-2 .Zavagliaetal.usedtheseNMMstoinvestigatetheeffectsofdifferentconnectivitypatternsbetweenLF,MF,andHFpopulations( Zavaglia 2008 ).Specically,theyconsideredanetworkcomposedofthreeNMMpopulations,oneLFpopulation,oneMFpopulation,andoneHFpopulationwiththeconnectivitypatternsillustratedin 102

PAGE 103

Fig. 5-3 ,anddemonstratedthatthePSDofeachpopulationandthecoherencebetweenpopulationswerebothafunctionoftheunderlyingconnectivityofthenetwork.Next,theyusedthenetworkinFig. 5-3 Aandadjustedthestrengthsofcouplingbetweenthepopulationsinordertotthemodel'sPSDtothePSDofanEEGrecordedfromthecingulatedcortexoftestsubjects.Zavagliaetal.alsoextendedtheirmodeltoconsiderentireROIswithinthebrain( Zavagliaetal. 2006 ).Inthiscase,eachROIiscomposedofthreeNMMpopulations(oneLF,oneMFandoneHF)andcanbeconnectedtoanynumberofotherROIstoformanetwork.AsimplenetworkoftwocoupledROIsisillustratedinFig. 5-4 .BecauseeachROIiscomposedofthreedifferentpopulations,eachwithadifferentintrinsicringrhythm,theROIasawholecanexhibitrhythmicringacrossarangeoffrequencies.ThisisillustratedinFig. 5-5 whichshowsthePSDofasingleROI.Muchliketheirpreviousstudy,Ursinoetal.rstinvestigatedtheimpactthatdifferentROInetworkconnectivitypatternshadonthePSDofeachROI( Ursinoetal. 2007 ).Finally,usinganetworkofthreeROIs,theyadjustedthestrengthofvariousconnectionsinthenetworkinordertotthePSDofeachROItothePSDofEEGsrecordedfromseveraltestsubjectsperformingngermovingtasks.Ursinoetal.putthismethodforwardasatechniqueforestimatingeffectiveconnectivityofthebrainviaphysiologicalmodels( Ursinoetal. 2007 ).Determiningconnectivityinthiswayisamodel-drivenapproachsinceitreliesonanassumednetworkofROIs,witheachROIcomposedofseveralNMMpopulationspossessingparticularfrequencycharacteristics.Generaltimeseriesinferencetechniquesaremuchmoreversatilesincetheyarenotpredicatedonaparticularmodeloftheunderlyingsystem.Asofyet,however,theirefcacywhenappliedtononlinear,EEG-likedatahasnotbeenveried. 103

PAGE 104

5.2DeterminingEffectiveConnectivityinNetworksofNeuralMassModelPopulationsIbeganbyconsideringnetworksofNMMpopulationssimilartothatillustratedinFig. 5-3 ( Zavaglia 2008 ).Specically,thenetworksIconsideredwereeachcomposedofthreeNMMpopulations,oneLF,oneMFandoneHFpopulation,alongwithallpossiblecongurationsofeffectiveconnectivitybetweenthepopulations.Inordertoevaluatetheaccuracyofvarioustimeseriesinferencetechniqueswhenappliedtosuchnetworks,Irstsimulated1,000ofthesenetworkswiththepresenceanddirectionalityofcouplingbetweenpopulationschosenatrandom.1Bidirectionalconnectionswereallowedbutself-connectionswerenot.Foreachnetwork,IappliedthetimeseriesinferencetechniquesconditionalGC,GPDCandmultivariatePDMinordertoestimatetheeffectiveconnectivitybetweenNMMpopulations(refertoChapter 2 forimplementationdetailsofeachtechniquealongwithcalculationofsignicancethresholds).AnexampleofoneofthesenetworksalongwithoutputfromGC,GPDC,andPDMwhenappliedtothenetworkisshowninFig. 5-6 .Sincethetrueeffectiveconnectivityofeachnetworkwasknowninadvance,Iwasabletocalculatetheaccuracyofeachtechniqueacrossthe1,000simulatednetworks.Anaccuratetechniqueisonethata)identiesalleffectiveconnectionsthatarepresentinthenetworkandb)doesnotidentifyanyeffectiveconnectionsthatarenotinthenetwork.Thesefeaturescanbequantiedusingthefollowingstatisticaltermsastheyrelatetoidentifyingeffectiveconnectivityusingtimeseriesinferencetechniques: Truepositive-thetechniqueidentiesaneffectiveconnectionthatisactuallypresentinthenetwork 1ThestochasticdifferentialequationsgiveninEq. 5.1 weresolvedusingthestochasticEuler-Maruyamamethodwithastepsizeoft=0.001( KloedenandPlaten 1999 ). 104

PAGE 105

Truenegative-thetechniquedoesnotidentifyaneffectiveconnectionthatisnotactuallypresentinthenetwork Falsepositive-thetechniqueidentiesaneffectiveconnectionthatisnotactuallypresentinthenetwork Falsenegative-thetechniquedoesnotidentifyaneffectiveconnectionthatisactuallypresentinthenetworkUsingthesestatistics,sensitivityandspecicityvaluesforeachofthetechniquescanbecalculatedusingthefollowingformulas:sensitivity=#oftruepositives #oftruepositives+#offalsenegatives (5)specicity=#oftruenegatives #oftruenegatives+#offalsepositives (5)Asensitivityof100%meansthatthetechniqueidentiesalleffectiveconnectionsthatarepresentinthenetworkwhileaspecicityof100%meansthatthetechniquedoesnotidentifyanyeffectiveconnectionsthatarenotinthenetwork.SensitivityandspecicityvaluesforeachofthethreetechniquesarepresentedinTable 5-3 .ThehighspecicityandlowsensitivityofGCandPDMindicatesthatthesetechniquestendtorejectmanyconnectionsthatareactuallypresentinthesystem.Clearly,GPDCoutperformsbothofthesetechniquesintermsofbothsensitivityandspecicity,withvalues92%and90%,respectively.Furthermore,thesehighsensitivityandspecicityvaluesareagoodindicationthatGPDCisaviabletechniquefordeterminingeffectiveconnectivityinnetworksofNMMpopulationswheneachpopulationexhibitsasingledominantringfrequency.Severalreal-worldexamplesofsystemsexhibitingsingledominantfrequenciesincludebetabandoscillationsobservedintheprimarymotorcortexduringmotormaintenancebehaviorandthalamocorticalalphabandoscillationsobservedattheonsetofsleep( Brovellietal. 2004 ; Steriadeetal. 1993 ).SuchsystemsaregoodcandidatesforanalysisviatheGPDCtechnique. 105

PAGE 106

5.3DeterminingEffectiveConnectivityinNetworksofNeuralMassModelRegionsofInterest 5.3.1DeterminingEffectiveConnectivityViaGrangerCausality,PartialDirectedCoherenceandPhaseDynamicsModelingNext,IconsiderednetworkscomposedoftwointerconnectedNMMROIssuchasthatillustratedinFig. 5-4 ( Ursinoetal. 2007 ; Zavagliaetal. 2006 ).Similartomypreviousanalysis,Isimulated1,000ofthesenetworkswiththepresenceanddirectionalityofconnectionschosenatrandom.2Bidirectionalconnectionswereallowedbutself-connectionswerenot.IthenappliedconditionalGC,GPDCandmultivariatePDMtoeachofthesenetworksinordertoobtainestimatesofthenetworks'effectiveconnectivityandonceagaincalculatedeachtechnique'ssensitivityandspecicitystatistics.TheresultsarepresentedinTable 5-4 .Inthiscase,PDMistheleastaccurateofthethreetechniques.Itslowspecicityvaluemeansthatitoftenconcludesthataneffectiveconnectionispresentevenwhenitnotactuallyinthenetwork.PDM'sinaccuracyisanexpectedresultconsideringthatPDMisdesignedforuseonnarrow-bandsignalsandNMMROIsproducesignalsspreadacrossmultiplefrequencies.However,theseresultsalsoindicatethatGCandGPDCarenotreliablewhenappliedtosuchnetworks.Thelowspecicityvaluesofbothtechniquesindicatethattheyoftenincorrectlyidentifyaneffectiveconnectionaspresentwhenitisnotactuallyinthenetwork.ThiscallsintoquestionthevalidityofresultsobtainedwhenapplyingthesetechniquestogeneralEEGsignalscontainingoscillatoryactivityacrossarangeoffrequencies.Inthefollowing,Ipresentanovelalternativetothesetechniquesforuseonsuchsignals. 2ThestochasticdifferentialequationsgiveninEq. 5.1 weresolvedusingthestochasticEuler-Maruyamamethodwithastepsizeoft=0.001( KloedenandPlaten 1999 ). 106

PAGE 107

5.3.2EMD-GPDM:ANovelTechniqueforDeterminingEffectiveConnectivityinNetworksofNeuralMassModelRegionsofInterestThetimeseriesinferencetechniqueIintroduceherefordeterminingeffectiveconnectivityisbasedonPDMandempiricalmodedecomposition(EMD)( Huangetal. 1998 ).Byitself,PDMisnotapplicabletotimeseriesderivedfromNMMROIssince,asIdemonstratedabove,thespectralcharacteristicsofthesetimeseriesarenotnarrow-band.Inordertoovercomethislimitation,IuseEMDtorstseparateanNMMROItimeseriessignalintoseveralofitsconstituentsignalcomponentsknownasintrinsicmodefunctions(IMFs).DetailsontheextractionofIMFsusingEMDisgiveninSection 5.4 .AsaconsequenceofEMD,eachIMFisnarrow-bandandhasawell-denedinstantaneousphase.Thus,IMFsarewell-suitedtoPDManalysis.AsIdemonstratebelow,byapplyingPDMtotheseIMFsindividually,oneisabletodeterminetheeffectiveconnectivityofanetworkofNMMROIs.ItshouldbenotedthatinorderforPDMtoworkproperlyinthisapplication,amodicationtothetechniqueisneeded.ThismodiedPDMtechniqueisreferredtoasgeneralizedPDM(GPDM)andisdiscussedinmoredetailinChapter 3 .GPDMisessentialformyproposedapproachsinceIMFsvarywidelyintheamountofnoisetheycontain.TherstIMFsthatEMDextractsfromasignalcontainnearlyallofthenoisethatwasintheoriginalsignalwhilethelastfewIMFscontainalmostnonoise.Forinstance,letusconsiderIMFiandIMFjwhenivar(IMFj),wherevar()denotesvariance.ThisisillustratedinFig. 5-7 whichshowstwoIMFsextractedfromthetimeseriesofanNMMROI.ThevarianceofIMF4is2.08whilethevarianceofIMF9is0.162.Asaconsequence,whenPDMisappliedtoIMF4andIMF9,effectiveconnectivityisnearlyalwaysidentiedasgoingfromIMF9toIMF4,regardlessofthetrueeffectiveconnectivityinthesystem.Inthisexamplecase,thereisnoactualeffectiveconnectivitybetweenthesignalssincebothoriginatefromthesameNMMROI 107

PAGE 108

whichdoesnothaveaselfconnection.Nevertheless,PDMgivesavalueof0.77intheIMF9toIMF4directioncomparedtoavalueof0.005intheIMF4toIMF9direction.ThemechanismbehindPDM'sbiasofmisidentifyingeffectiveconnectivityascomingfromalessnoisysignalandgoingtoamorenoisysignalisgiveninChapter 3 .Inaddition,IintroducedinthischapteratechniquetermedGPDMwhichremovesthisbiasandisthereforeabletoidentifyeffectiveconnectivityacrossdifferentIMFsregardlessoftheirnoisedifferenceor,inotherwords,regardlessoftheorderinwhichtheywereextractedfromthesignal.BecausemynewtechniqueutilizesbothEMDandGPDMtoextracteffectiveconnectivityfromNMMROItimeseries,IrefertoitasEMD-GPDM.TheapplicationofEMD-GPDMtotwotimeseries,x(t)andy(t),involvesthefollowingsteps: 1. ApplyEMDtox(t)andy(t)inordertoobtainIMFsforeachtimeseries.AssumethetotalnumberofIMFsofx(t)isnandthetotalnumberofIMFsofy(t)ism.DenotetheseIMFsasIMFixandIMFjywhereiandjrangefrom1tonandm,respectively. 2. CalculatethePSDofeachIMF,denotedasPSDix(f)andPSDjy(f)fortimeseriesx(t)andy(t),respectively.ThenndthemaximumpowerofeachIMFasPSDmaxix=max(PSDix(f))orPSDmaxjy=max(PSDjy(f)). 3. FindtheIMFforeachtimeseriesthatsatises:^i=argmaxi(PSDmaxix)^j=argmaxj(PSDmaxjy)Inotherwords,foreachtimeseries,ndtheIMFwiththegreatestoverallpower. 4. UseGPDMtoestimatetheeffectiveconnectivitybetweenIMF^ixandIMF^jy. 5. IfGPDMfromIMF^ixtoIMF^jyissignicant,theneffectiveconnectivityfromx(t)toy(t)existsinthenetwork.Similarly,ifGPDMfromIMF^jytoIMF^ixissignicant,theneffectiveconnectivityfromy(t)tox(t)existsinthenetwork. 108

PAGE 109

AnexampleoftheapplicationofEMD-GPDMtotwoNMMROItimeseriesispresentedinFig. 5-8 .IappliedmyEMD-GPDMtechniquetothe1,000previouslysimulatedNMMROInetworksinordertodemonstrateitsaccuracy.TheresultingsensitivityandspecicitystatisticsoftheEMD-GPDMtechniquearepresentedinTable 5-5 .Withsensitivityandspecicityvaluesofatleast90%,itisclearthatEMD-GPDMcorrectlydetermineseffectiveconnectivityinanNMMROInetworkwithhighprobability.TheseresultsprovideastrongindicationthatEMD-GPDMisaviabletechniquefordeterminingeffectiveconnectivityinsystemswithtimeseriessimilartothosederivedfromnetworksofNMMROIs.SuchsystemsarecommonlyencounteredandincludealargefractionofrecordedEEGsignals.Thus,thistechniqueiswidelyapplicabletoalargenumberofrecordedEEGtimeseries.AsfurtherproofoftheneedforGPDMinstep4oftheEMD-GPDMtechnique,IrepeatedtheaboveanalysiswithPDMinplaceofGPDM.TheresultsarepresentedinthesecondrowofTable 5-5 .Fromthespecicityvalue,itcanbeseenthatPDMperformspoorlywhenusedinplaceofGPDMforthereasonsdiscussedabove.ThisprovidesaclearmotivationfortheuseoftheGPDMtechniquewhenconsideringnoisytimeseries. 5.4EmpiricalModeDecompositionEMDisatechniqueproposedbyHuangetal.foranalyzingnonlinear,non-stationaryand/orbroadbandsignals( Huangetal. 1998 ).Suchsignalshaveoftenposedaproblemfortraditionalharmonicandempiricalorthogonalfunction(EOF)analyses.Forinstance,theFouriertransformofanon-stationarysignalisill-denedsinceatraditionalFourieranalysisassumesthatthefrequencyofasignalisdenedbyaconstantsine/cosinefunctionthatspanstheentiredatalength.Fornon-stationarydata,however,thefrequencyofthesignalvariesovertime,andsothissine/cosinefunctionwillnotbeconstantovertheentiredatalength( Priestley 1967 ).Another 109

PAGE 110

exampleistheHilberttransformcomplexEOF.AccordingtoMerrieldandGuza( MerrieldandGuza 1990 ),thiscomplexEOFonlymakessenseifthefrequencydistributionoftheoriginalsignalisnarrowband(thoughthereisstillsomecontroversysurroundingthisissue).Thus,itisarguedthattheHilberttransformofabroadbandsignalisuninterpretable.EMDresolvesbothissuesdescribedabovefornon-stationaryandbroadbandsignalsbydecomposingasignalintoitsconstituentintrinsicmodefunctions(IMFs)( Huangetal. 1998 ).EachIMFmustsatisfytwoconditions:1)fortheentiredatalength,thenumberofextremaandthenumberofzerocrossingsmustbeeitherequalordifferbyatmostoneand2)atanypoint,themeanvalueoftheenvelopesdenedbythelocalmaximaandlocalminimamustbezero.Therstconditionissimilartoastandarddenitionofanarrowbandsignal,namelythattheexpectednumberofextremaandzerocrossingsmustbeequal.Thesecondconditionmeansthatlocally,themeanofeachIMFiszeroandthussymmetricwithrespecttothelocalzeromean.Sincethisconceptoflocalitycanbedifculttodene,especiallyfornon-stationarysignals,EMDusesthelocalmeanoftheenvelopesdenedbythelocalmaximaandlocalminimatoforcelocalsymmetry.ThetwoconditionsoutlinedaboveleadtotheEMDprocessforextractinganIMFfromasignal,x(t): 1. Identifyallthelocalextremaofx(t) 2. Connectallthelocalmaximbyacubicsplinelineanddenethisastheupperenvelope 3. ConnectallthelocalminimabyacubicsplinelineanddenethisasthelowerenvelopeDenotethemeanoftheupperandlowerenvelopesasm1.Thenthedifferencebetweenx(t)andm1istherstcomponent,h1.Inorderwords,x(t))]TJ /F5 11.955 Tf 12.64 0 Td[(m1=h1.Atthispoint,h1probablydoesnotqualifyasanIMFandsoasiftingprocesstakesplacewhich 110

PAGE 111

eliminatesridingwavesandmakesthewave-prolesmoresymmetric.Inthissiftingprocess,h1istreatedasthedatax(t)instep1aboveandtheremainingstepsarerepeatedresultinginh1)]TJ /F5 11.955 Tf 11.38 0 Td[(m11=h11.Siftingcontinuesktimesuntilh1kisanIMFatwhichpointitisdesignatedasc1=h1k.Becausethesiftingprocesssmoothsoutunevenamplitudesinanattempttomakethewave-prolessymmetric,carryingtheprocesstotheextremecouldeliminateallphysicallymeaningfulamplitudeuctuationsintheoriginalsignal.Thus,astoppingcriteriaisneededtoguaranteethatIMFsretaintheirphysicalrelevance.ThestoppingcriterionproposedbyHuangetal.( Huangetal. 1998 )issimilartotheCauchyconvergencetestandisdenedbyasumofthedifference(SD)asSDk=PTt=0jhk)]TJ /F8 7.97 Tf 6.59 0 Td[(1(t))]TJ /F5 11.955 Tf 11.96 0 Td[(hk(t)j2 PTt=0h2k)]TJ /F8 7.97 Tf 6.59 0 Td[(1(t) (5)Typically,whenSDkreachesavaluebetween0.2and0.3thesiftingprocessstops.Oncethesiftingprocessiscomplete,therstIMFobtainedisc1anditshouldcontaintheshortestperiodcomponentoftheoriginalsignal.Theprocesscontinuesbyremovingc1fromx(t)andndingthenextIMFinthedata.Specically,x(t))]TJ /F5 11.955 Tf 12.29 0 Td[(c1=r1,wheretheresiduer1isnowusedasinputtotheEMDprocedure.TheEMDprocedurenallystopswhentheresidue,rn,becomesamonotonicfunctionfromwhichnomoreIMFscanbeextracted.Atthispoint,nIMFshavebeenextractedandtheoriginalsignalx(t)canbewrittenasx(t)=nXj=1cj+rn (5)Thus,EMDformsacompletebasisfortheoriginalsignalx(t).Inaddition,Huangetal.observedthatlocally,anytwoIMFcomponentsshouldalsobeorthogonalforallpracticalpurposes( Huangetal. 1998 ).TheotherbenetsofEMDarethatisintuitiveandadaptivetoavarietyofdifferenttypesofsignals.However,beinganempiricalapproach,itdoesnotallowfortheoreticalanalysisandperformanceevaluation. 111

PAGE 112

Withinthecontextofeffectiveconnectivitydetermination,EMDisaperfectcomplementtothePDMtechniquesincePDMrequiresinputsignalsthatarenarrow-band( RosenblumandPikovsky 2001 ).DecomposingasignalviaEMDintoitsIMFsandthenapplyingPDMacrosstheseIMFsisanaturalextensionofthePDMtechnique,enablingittobeappliedtobroadbandsignals.TheEMD-GPDMtechniquethatIproposeinthischapterformalizesthisapproachandprovestobeanaccuratetechniquefordeterminingeffectiveconnectivityinbroadbandEEGsignals. 112

PAGE 113

Table5-1. Neuralmassmodelbasalparametervaluesthatareconsistentacrossallpopulationtypes ParameterValue C1135C2108C333.75C433.75C540.5C613.5C7108s06e02.5r0.56 Table5-2. Neuralmassmodelbasalparametervaluesthatdifferamongpopulationtypes ParameterLFValueMFValueHFValue A2.75.25.6B3.24.53.8G394375a1(s)]TJ /F8 7.97 Tf 6.59 0 Td[(1)4085110b1(s)]TJ /F8 7.97 Tf 6.58 0 Td[(1)203040g1(s)]TJ /F8 7.97 Tf 6.58 0 Td[(1)3003507902202020000 Table5-3. SensitivityandspecicityvaluesofGC,GPDC,andPDMwhenappliedtonetworksofthreecoupledNMMpopulations SensitivitySpecicity GC34%97%GPDC92%90%PDM20%79% Table5-4. SensitivityandspecicityvaluesofGC,GPDC,andPDMwhenappliedtonetworksoftwocoupledNMMROIs SensitivitySpecicity GC100%33%GPDC99%39%PDM76%15% 113

PAGE 114

Table5-5. SensitivityandspecicityvaluesofEMD-GPDMandEMD-PDMwhenappliedtonetworksoftwocoupledNMMROIs SensitivitySpecicity EMD-GPDM90%92%EMD-PDM79%0% Figure5-1. Asystemdiagramofasinglepopulationneuralmassmodel 114

PAGE 115

Figure5-2. Powerspectraldensitiesofneuralmassmodelpopulationswithlowintrinsicringfrequency(LF)between4-12Hz,mediumintrinsicringfrequency(MF)between12-30Hz,andhighintrinsicringfrequency(HF)greaterthan30Hz. 115

PAGE 116

Figure5-3. Twoneuralmassmodelpopulationnetworks,eachcomposedofonelowfrequency(LF)population,onemediumfrequency(MF)populationandonehighfrequency(HF)population 116

PAGE 117

Figure5-4. Twocoupledneuralmassmodelregionsofinterest.ThestrengthofconnectionfromROI1toROI2atlowfrequencyisdenotedbyWL12,atmediumfrequencybyWM12andathighfrequencybyWH12.Theinput,p(t)toeachpopulationisaGaussianwhitenoisesignal. Figure5-5. Powerspectraldensityofasingleneuralmassmodelregionofinterest 117

PAGE 118

Figure5-6. Anexampleneuralmassmodelpopulationnetworkandtheresultsobtainedfromtimeseriesinferencetechniqueswhenappliedtothenetwork.GCandGPDCoutputisafunctionoffrequency,soitispresentedinplotswithnormalizedfrequencyonthex-axis.PDMresultsarescalarvalueswhicharegiveninthetableatbottom.SignicancethresholdsareshownintheplotsforGCandGPDCasdashedlines.Signicanteffectiveconnectivityasdeterminedbyeachtechniqueishighlightedbygreyrectangles.GPDCproperlyidentieseffectiveconnectivityfromtheLFtotheMFpopulationandfromtheMFtotheHFpopulation.GCdoesnotidentifyanyoftheconnectionsandPDMidentiesthetwotrueconnectionsaswellasoneincorrectconnectionfromtheMFtotheLFpopulation. 118

PAGE 119

Figure5-7. AnexampleNMMROIsignalandtwoofitsconstituentIMFsextractedbyEMD.IMF4isextractedviaEMDbeforeIMF9andIMF4hasamuchlargervariancethanIMF9. 119

PAGE 120

Figure5-8. TheEMD-GPDMtechniqueappliedtoanexampleNMMROInetwork 120

PAGE 121

CHAPTER6CONCLUSIONSNumerousstudieshaveappliedlineartimeseriesinferencetechniquessuchasGrangercausality(GC)andpartialdirectedcoherence(PDC)ornonlineartechniquessuchasphasedynamicsmodeling(PDM)toneuraldatasetsinanattempttogaininsightintothedirectionalityofinteractionsinthebrain( Brovellietal. 2004 ; Cadotteetal. 2010 ; Chenetal. 2006 ; Havliceketal. 2010 ; Liaoetal. 2010a ; Satoetal. 2009 ).Datasetsaregenerallyconsideredamenabletosuchananalysisaslongastheyarestationary.Theresultsofmystudydemonstratethenecessityforanalyzingtimeseriesdatasetsbeyondjustdeterminingtheirstationarity.Forinstance,IdemonstratedthatnoisevariancedifferencesbetweentimeseriescanhaveasignicantimpactontheresultsofPDCandPDM.AmodiedPDCtechnique,termedgeneralizedPDC(GPDC)waspreviouslydevelopedasanoise-variance-immunealternativetoPDC.Inasimilarfashion,IproposedanddemonstratedtheuseofanovelPDMtechnique,whichIrefertoasgeneralizedPDM(GPDM),whichisalsoimmunetonoisevariancedifferences.Synchronyisanotherexampleofatimeseriesfeaturethatcanresultinthefailureofmanytechniquesandpreviousworkhasindicatedapproximateboundsonthedegreeofphasecoherencebetweentwooscillatorsabovewhichseveralofthesetechniquestendtofail( Bezruchkoetal. 2003 ; Smirnovetal. 2007 ).Iextendedtheseresultsbyreningthelimitsonphasecoherenceaswellasconsideringlimitsonothertimeseriesfeatures,suchasnoiseandringfrequency,forthreeofthemostwidelyusedlinearmethodsfordeterminingeffectiveconnectivity,GC,PDCandGPDC,aswellasPDM,whichcanhandlesomenonlinearities.Animportantcontributionofmyworkistheintroductionofamethodologytoassesswhetheragiventechniquefordeterminingeffectiveconnectivitywillproducecorrectresultswhenappliedtoagivendatasetbasedontheobservable,dynamical 121

PAGE 122

characteristicsofthedataset.Theproposedmethodologyconsistsofdecisiontreeclassiersthatcanbeappliedwithoutanypriorknowledgeofnetworkstructure.Idevelopedthismethodologyanddemonstrateditsusewithinthecontextofnarrow-bandneuronaloscillators,whicharerepresentativeofdataonthesingle-neuronscale.Furthermore,Iperformedamuch-neededinvestigationintotheapplicabilityoftimeseriesinferencetechniquestoneuralmassmodels(NMMs),whicharerepresentativeofelectroencephalographic(EEG)datacollectedatthescaleoflargeneuronalpopulations.WhilethesetechniquesareoftenusedtoanalyzeexperimentalEEGdata( Brovellietal. 2004 ; Cadotteetal. 2010 ; Chenetal. 2006 ),fewstudieshaveexploredthevalidityofthesetechniqueswhenappliedtosuchdata.IshowedthatwhileGPDCisaccuratewithhighprobabilitywhenappliedtoNMMsthatexhibitrhythmicactivityatasingledominantfrequency,noneofthetechniquesIconsideredwereaccuratewhenappliedtoNMMsthatexhibitrhythmicactivityoverarangeoffrequencyvalues.IintroducedanewtechniquewhichIcallEMD-GPDM,todetermineeffectiveconnectivityinsuchNMMs.Theempiricalmodedecomposition(EMD)aspectofthetechnique,whichdecomposesasignalintoconstituent,narrow-bandsignals,perfectlycomplementstheapplicationofGPDMsinceGPDMrequiresnarrow-bandsignals.MyproposedEMD-GPDMtechniqueprovestobeanaccuratemethodfordeterminingeffectiveconnectivityinNMMswithbroadbandcharacteristics,whicharerepresentativeofalargeclassofexperimentalEEGrecordings. 122

PAGE 123

REFERENCES AbarbanelH,GibbL,HuertaR,RabinovichM(2003)Biophysicalmodelofsynapticplasticitydynamics.BiolCybern89:214 AkaikeH(1969)Fittingautoregressivemodelsforprediction.AnnIStatMath21:243 AkaikeH(1974)Anewlookatthestatisticalmodelidentication.IEEETAutomatContrAC-19:716 AlboZ,PriscoGD,ChenY,RangarajanG,TruccoloW,FengJ,VertesR,DingM(2004)Ispartialcoherenceaviabletechniqueforidentifyinggeneratorsofneuraloscillations?BiolCybern90(5):318 AllumJ,Hepp-ReymondM,GysinR(1982)Cross-correlationanalysisofinterneuronalconnectivityinthemotorcortexofthemonkey.BrainRes231:325 AstolL,CincottiF,MattiaD,MarcianiM,BaccalaL,FallaniF,SalinariS,UrsinoM,ZavagliaM,DingL,EdgarJ,MillerG,HeB,BabiloniF(2007)Comparisonofdifferentcorticalconnectivityestimatorsforhigh-resolutioneegrecordings.HumBrainMapp28:143 BaccalaL,SameshimaK(2001)Partialdirectedcoherence:anewconceptinneuralstructuredetermination.BiolCybern84(6):463 BaccalaL,SameshimaK,BallesterG,ValleAD,TimoIariaC(1998)Studyingtheinteractionbetweenbrainstructuresviadirectedcoherenceandgrangercausality.AppSignalProcess5(1):40 BaccalaL,SameshimaK,TakahashiD(2007)Generalizedpartialdirectedcoherence.In:Proceedingsofthe15thInternationalConferenceonDigitalSignalProcessing,Cardiff,Wales,UK,pp163 BennettM(1997)Gapjunctionsaselectricalsynapses.JNeurocytol26:349 BezruchkoB,PonomarenkoV,RosenblumM,PikovskyA(2003)Characterizingdirectionofcouplingfromexperimentalobservations.Chaos13:179 BlairR,KarniskiW(1993)Analternativemethodforsignicancetestingofwaveformdifferencepotentials.Psychophysiology30:518 BoxG,JenkinsG,ReinselG(2008)Timeseriesanalysis:ForecastingandControl,4thedn.Wiley,Hoboken,NJ BreimanL,FriedmanJ,StoneC,OlshenR(1984)Classicationandregressiontrees.ChapmanandHall,NewYork,NY 123

PAGE 124

BrovelliA,DingM,LedbergA,ChenY,NakamuraR,BresslerS(2004)Betaoscillationsinalarge-scalesensorimotorcorticalnetwork:directionalinuencesrevealedbygrangercausality.PNatlAcadSciUSA101(26):9849 BucknerR,SepulcreJ,TalukdarT,KrienenF,LiuH,HeddenT,Andrews-HannaJ,SperlingR,JohnsonK(2009)Corticalhubsrevealedbyintrinsicfunctionalconnectivity:mapping,assessmentofstability,andrelationtoalzheimer'sdisease.JNeurosci29:1860 BuzsakiG(2006)Rhythmsofthebrain.OxfordUniversityPress,Oxford,NY BuzsakiG,DraguhnA(2004)Neuronaloscillationsincorticalnetworks.Science304(5679):1926 CadotteA,DeMarseT,MareciT,ParekhM,TalathiS,HwangDU,DittoW,DingM,CarneyP(2010)Grangercausalityrelationshipsbetweenlocaleldpotentialsinananimalmodeloftemporallobeepilepsy.JNeurosciMeth189:121 ChenY,RangarajanG,FengJ,DingM(2004)Analyzingmultiplenonlineartimeserieswithextendedgrangercausality.PhysLettA324(1):26 ChenY,BresslerS,DingM(2006)Frequencydecompositionofconditionalgrangercausalityandapplicationtomultivariateneuraleldpotentialdata.JNeurosciMeth150(2):228 ChowC,KopellN(2000)Dynamicsofspikingneuronswithelectricalcoupling.NeuralComput12(7):1643 DavidO,FristonK(2003)Aneuralmassmodelformeg/eegcouplingandneuronaldynamics.NeuroImage20:1743 DavidO,GuillemainI,SailletS,ReytS,DeransartC,SegebarthC,DepaulisA(2008)Identifyingneuraldriverswithfunctionalmri:Anelectrophysiologicalvalidation.PLoSBiol6:e315 DavidsonR,MacKinnonJ(2003)Econometrictheoryandmethods.OxfordUniversityPress,Oxford,NY DeisterC,TeagardenM,WilsonC,PaladiniC(2009)Anintrinsicneuronaloscillatorunderliesdopaminergicneuronbursting.JNeurosci29:15,888,897 DingM,ChenY,BresslerS(2006)Grangercausality:basictheoryandapplicationtoneuroscience.In:SchelterB,WinterhalderM,TimmerJ(eds)HandbookofTimeSeriesAnalysis,Wiley-VCHVerlage,Wienheim,Germany,pp451 ErmentroutB(1996)Type1membranes,phaseresettingcurves,andsynchrony.NeuralComput8(5):979 124

PAGE 125

FaisalA,SelenL,WolpertD(2008)Noiseinthenervoussystem.NatRevNeurosci9:292 FisherR(1925)StatisticalMethodsforResearchWorkers.OliverandBoyd,Edinburgh,UK FlorinE,GrossJ,PfeiferJ,FinkG,TimmermannL(2010)Theeffectoflteringongrangercausalitybasedmultivariatecausalitymeasures.NeuroImage50(2):577 FreemanW(1978)Modelsofthedynamicsofneuralpopulations.ElectroencephalogrClinNeurophysiol34:9 FriesP(2009)Neuronalgamma-bandsynchronizationasafundamentalprocessincorticalcomputation.AnnuRevNeurosci32:209 FristonK(1994)Functionalandeffectiveconnectivityinneuroimaging:asynthesis.HumBrainMapp2:56 FristonK(2009a)Causalmodellingandbrainconnectivityinfunctionalmagneticresonanceimaging.PLoSBiol7(2):221 FristonK(2009b)Causalmodellingandbrainconnectivityinfunctionalmagneticresonanceimaging.PLoSBiol7:0220 FristonK,HarrisonL,PennyW(2003)Dynamiccausalmodelling.NeuroImage19:1273 FritschG,HitzigE(1870)Ontheelectricalexcitabilityofthecerebrum.In:vonBoninG(ed)Somepapersonthecerebralcortex,BlackwellScienticPublishing,Springeld,IL,pp73 GardnerW(1988)Statisticalspectralanalysis.Prentice-Hall,EnglewoodCliffs,NJ GewekeJ(1982)Measurementoflineardependenceandfeedbackbetweenmultipletimeseries.JAmStatAssoc77(378):304 GewekeJ(1984)Measuresofconditionallineardependenceandfeedbackbetweentimeseries.JAmStatAssoc79(388):907 GoebelR,RoebroeckA,KimD,FormisanoE(2003)Investigatingdirectedcorticalinteractionsintime-resolvedfmridatausingvectorautoregressivemodelingandgrangercausalitymapping.MagnResonImaging21:1251 GrangerC(1969)Investigatingcausalrelationsbyeconometricmodelsandcross-spectralmethods.Econometrica37(3):424 GrangerC(1980)Testingforcausality:Apersonalviewpoint.JEconDynControl2:329 125

PAGE 126

GuoS,SethA,KendrickK,ZhouC,FengJ(2008)Partialgrangercausality:Eliminatingexogenousinputsandlatentvariables.JNeurosciMeth172(1):79 HamillO,MartyA,NeherE,SakmannB,SigworthF(1981)Improvedpatch-clamptechniquesforhigh-resolutioncurrentrecordingfromcellsandcell-freemembranepatches.ArchivEuroJPhysiol391:85 HamiltonJ(1994)TimeSeriesAnalysis.PrincetonUniversityPress,Princeton,NJ HammondC,BergmanH,BrownP(2007)Pathologicalsynchronizationinparkinson'sdisease:networks,modelsandtreatments.TrendsNeurosci30(7):357 HastieT,TibshiraniR,FriedmanJ(2001)TheElementsofStatisticalLearning:DataMining,InferenceandPrediction.SpringerVerlag,NewYork,NY HavlicekM,JanJ,BrazdilM,CalhounV(2010)Dynamicgrangercausalitybasedonkalmanlterforevaluationoffunctionalnetworkconnectivityinfmridata.NeuroImage53:65 HuangN,ShenZ,LongS,WuM,ShihH,ZhengQ,YenN,TungC,LiuH(1998)Theempiricalmodedecompositionandthehilbertspectrumfornonlinearandnon-stationarytimeseriesanalysis.PRoySocLondAMat454(1971):903 IzhikevichE(2007)Dynamicalsystemsinneuroscience:Thegeometryofexcitabilityandbursting.TheMITPress,Cambridge,MA JansenB,RitV(1995)Electroencephalogramandvisualevokedpotentialgenerationinamathematicalmodelofcoupledcorticalcolumns.BiolCybern73:357 KaminskiM,BlinowskaK(1991)Anewmethodofthedescriptionoftheinformationowinthebrainstructures.BiolCybern65:203 KaminskiM,DingM,TruccoloW,BresslerS(2001)Evaluatingcausalrelationsinneuralsystems:Grangercausality,directedtransferfunctionandstatisticalassessmentofsignicance.BiolCybern85(2):145 KandelE,SchwartzJ,JessellT(2000)Principlesofneuralscience.McGraw-Hill,NewYork,NY KayserA,SunF,D'EspositoM(2009)Acomparisonofgrangercausalityandcoherencyinfmri-basedanalysisofthemotorsystem.HumBrainMapp30(11):3475 KiebelS,GarridoM,MoranR,ChenC,FristonK(2009)Dynamiccausalmodelingforeegandmegdata.HumBrainMapp30:1866 KlineR(2010)Principlesandpracticeofstructuralequationmodeling.TheGuilfordPress,NewYork,NY KloedenP,PlatenE(1999)NumericalSolutionofStochasticDifferentialEquations.Springer,Berlin,Germany 126

PAGE 127

KralemannB,PikovskyA,RosenblumM(2011)Reconstructingphasedynamicsofoscillatornetworks.Chaos22:1 LiX,OuyangG(2010)Estimatingcouplingdirectionbetweenneuronalpopulationswithpermutationconditionalmutualinformation.NeuroImage52:497 LiaoW,MantiniD,ZhangZ,PanZ,DingJ,GongQ,YangY,ChenH(2010a)Evaluatingtheeffectiveconnectivityofrestingstatenetworksusingconditionalgrangercausality.BiolCybern102(1):57 LiaoW,MantiniD,ZhangZ,PanZ,DingJ,GongQ,YangY,ChenH(2010b)Evaluatingtheeffectiveconnectivityofrestingstatenetworksusingconditionalgrangercausality.BiolCybern102:57 LudtkeN,LogothetisN,PanzeriS(2010)Testingmethodologiesforthenonlinearanalysisofcausalrelationshipsinneurovascularcoupling.MagnResonImaging28:1113 LungarellaM,IshiguroK,KuniyoshiY,OtsuN(2007)Methodsforquantifyingthecausalstructureofbivariatetimeseries.IntJBifurcatChaos17:903 LutkepohlH(1985)Comparisonofcriteriaforestimatingtheorderofavectorautoregressiveprocess.JTimeSerAnal6:35 LutkepohlH(2005)NewIntroductiontoMultipleTimeSeriesAnalysis.Springer,NewYork,NY MacKinnonJ(2006)Bootstrapmethodsineconometrics.EconRec82:S2S18 MalsburgCVD(1994)Thecorrelationtheoryofbrainfunction.In:vanHemmenJ,SchultenK,DomanyE(eds)ModelsofneuralnetworksII:Temporalaspectsofcodingandinformationprocessinginbiologicalsystems,Springer-Verlag,NewYork,NY,pp95 MarinazzoD,PellicoroM,StramagliaS(2008)Kernelmethodfornonlineargrangercausality.PhysRevLett100(14):144,103 MassiminiM,FerrarelliF,HuberR,EsserS,SinghH,TononiG(2005)Breakdownofcorticaleffectiveconnectivityduringsleep.Science309:2228 McCullochW(1945)Aheterarchyofvaluesdeterminedbythetopologyofnervousnets.BMathBiol7:89 MerrieldM,GuzaR(1990)Detectingpropagatingsignalswithcomplexempiricalorthogonalfunctions:Acautionarynote.JPhysOceanogr20:1628 127

PAGE 128

MorecraftR,UgoliniG,LanciegoJ,WouterloodF,PandyaD(2009)Classicandcontemporaryneuraltracttracingtechniques.In:Johansen-BergH,BehrensT(eds)DiffusionMRI:Fromquantitativemeasurementtoin-vivoneuroanatomy,Elsevier,Burlington,MA,pp274 MorettinP(1984)Thelevinsonalgorithmanditsapplicationsintimeseriesanalysis.IntStatRev52(1):83 MormannF,LehnertzK,DavidP,ElgerC(2000)Meanphasecoherenceasameasureforphasesynchronizationanditsapplicationtotheeegofepilepsypatients.PhysicaD144:358 MorrisC,LecarH(1981)Voltageoscillationsinthebarnaclegiantmuscleber.BiophysJ35(1):193 NalatoreH,DingM,RangarajanG(2007)Mitigatingtheeffectsofmeasurementnoiseongrangercausality.PhysRevE75(3):031,123 NeumaierA,SchneiderT(2001)Estimationofparametersandeigenmodesofmultivariateautoregressivemodels.ACMTMathSoftware27:27 PeneldW,JasperH(1954)EpilepsyandtheFunctionalAnatomyoftheHumanBrain.Little,BrownandCo.,NewYork,NY PerkelD,GersteinG,MooreG(1967)Neuronalspiketrainsandstochasticpointprocesses:I.thesinglespiketrain.BiophysJ7(4):391 PikovskyA,RosenblumM,KurthsJ(2002)Synchronization:Auniversalconceptinnonlinearsciences.CambridgeUniversityPress,Cambridge,UK vanderPolB(1922)Onoscillation-hysteresisinasimpletriodegenerator.PhilMag43:700 PotterS(2001)Distributedprocessinginculturedneuronalnetworks.ProgBrainRes130:49 PrichardD,TheilerJ(1994)Generatingsurrogatedatafortimeserieswithseveralsimultaneouslymeasuredvariables.PhysRevLett73(7):951 PriestleyM(1967)Powerspectralanalysisofnon-stationaryrandomprocesses.JSoundandVibration6:86 RoebroeckA,FormisanoE,GoebelR(2005)Mappingdirectedinuenceoverthebrainusinggrangercausalityandfmri.NeuroImage25:230 RogersB,MorganV,NewtonA,GoreJ(2007)Assessingfunctionalconnectivityinthehumanbrainbyfmri.MagnResonImaging25:1347 RosenblumM,PikovskyA(2001)Detectingdirectionofcouplingininteractingoscillators.PhysRevE64(4):045,202 128

PAGE 129

RosenblumM,CimponeriuL,BezerianosA,PatzakA,MrowkaR(2002)Identicationofcouplingdirection:Applicationtocardiorespiratoryinteraction.PhysRevE65(4):041,909 SamoilovM,PriceG,ArkinA(2006)Fromuctuationstophenotypes:Thephysiologyofnoise.SciSTKE2006:re17re25 SatoJ,TakahashiD,ArcuriS,SameshimaK,MorettinP,BaccalaL(2009)Frequencydomainconnectivityidentication:anapplicationofpartialdirectedcoherenceinfmri.HumBrainMapp30(2):452 SchafferC(1993)Selectingaclassicationmethodbycross-validation.MachLearn13:135 SchelterB,WinterhalderM,EichlerM,PeiferM,HellwigB,GuschlbauerB,LuckingC,DahlhausR,TimmerJ(2006)Testingfordirectedinuencesamongneuralsignalsusingpartialdirectedcoherence.JNeurosciMeth152(1-2):210 SchelterB,TimmerJ,EichlerM(2009)Assessingthestrengthofdirectedinuencesamongneuralsignalsusingrenormalizedpartialdirectedcoherence.JNeurosciMeth179:121 SchillenT,KonigP(1994)Bindingbytemporalstructureinmultiplefeaturedomainsofanoscillatoryneuronalnetwork.BiolCybern70:397 SchneiderT,NeumaierA(2001)Algorithm808:Artamatlabpackageforestimationandspectraldecompositionofmultivariateautoregressivemodels.ACMTMathSoftware27:58 SchreiberT(2000)Measuringinformationtransfer.PhysRevLett85(2):461 SchreiberT,SchmitzA(2000)Surrogatetimeseries.PhysD142(3-4):346 SchwarzG(1978)Estimatingthedimensionofamodel.AnnStat6:461 SethA(2005)Causalconnectivityofevolvedneuralnetworksduringbehavior.etwComputNeuralSyst16:1 SethS,PrincipeJ(2010)Atestofgrangernon-causalitybasedonnonparametricconditionalindependence.In:Proceedingsofthe20thInternationalAssociationforPatternRecognition,Istanbul,Turkey,pp2620 daSilvaFL,canRotterdamA,BartsP,vanHeusdenE,BurrW(1976)Modelsofneuronalpopulations:thebasicmechanismsofrhythmicity.ProgBrainRes45:281 SmirnovD,AndrzejakR(2005)Detectionofweakdirectionalcoupling:Phase-dynamicsapproachversusstate-spaceapproach.PhysRevE71(3):036,207 129

PAGE 130

SmirnovD,BezruchkoB(2003)Estimationofinteractionstrengthanddirectionfromshortandnoisytimeseries.PhysRevE68(4):46,204 SmirnovD,BezruchkoB(2009)Detectionofcouplingsinensemblesofstochasticoscillators.PhysRevE79(4):046,204 SmirnovD,BodrovM,VelazquezJ,WennbergR,BezruchkoB(2005)Estimationofcouplingbetweenoscillatorsfromshorttimeseriesviaphasedynamicsmodeling:Limitationsandapplicationtoeegdata.Chaos15:024,102 SmirnovD,SchelterB,WinterhalderM,TimmerJ(2007)Revealingdirectionofcouplingbetweenneuronaloscillatorsfromtimeseries:Phasedynamicsmodelingversuspartialdirectedcoherence.Chaos17:013,111 SmirnovD,BarnikolU,BarnikolT,BezruchkoB,HauptmannC,BuhrleC,MaaroufM,SturmV,FreundH,TassP(2008)Thegenerationofparkinsoniantremorasrevealedbydirectionalcouplinganalysis.EurophysLett83:295 SomersD,KopellN(1993)Rapidsynchronizationthroughfastthresholdmodulation.BiolCybern68:393 SpornsO(2010)Networksofthebrain.MITPress,Cambridge,MA SpornsO,ChialvobDR,KaisercM,HilgetagCC(2004)Organization,developmentandfunctionofcomplexbrainnetworks.TrendsCognSci8:418 SpornsO,TononiG,KotterR(2005)Thehumanconnectome:astructuraldescriptionofthehumanbrain.PLoSComputBiol1:245 SteriadeM,McCormickD,SejnowskiT(1993)Thalamocorticaloscillationsinthesleepingandarousedbrain.Science262:679 TalathiS,HwangDU,CarneyP,DittoW(2010)Synchronywithshuntinginhibitioninafeedforwardinhibitorynetwork.JComputNeurosci28:305 TimmerJ(2000)Whatcanbeinferredfromsurrogatedatatesting?PhysRevLett85(12):2647 TropiniG,ChiangJ,WangZ,TyE,McKeownM(2011)Altereddirectionalconnectivityinparkinson'sdiseaseduringperformanceofavisuallyguidedtask.Neuroimage32:1 UddinL,KellyAC,BiswalB,CastellanosF,MilhamM(2009)Functionalconnectivityofdefaultmodenetworkcomponents:correlation,anticorrelation,andcausality.HumBrainMapp30:625 UhlhaasP,SingerW(2010)Abnormalneuraloscillationsandsynchronyinschizophrenia.NatRevNeurosci11:100 130

PAGE 131

UrsinoM,ZavagliaM,AstolL,BabiloniF(2007)Useofaneuralmassmodelfortheanalysisofeffectiveconnectivityamongcorticalregionsbasedonhighresolutioneegrecordings.BiolCybernpp351 Valdes-SosaP,Sanchez-BornotJ,Lage-CastellanosA,Vega-HernandexM,Bosch-BayardJ,Melie-GarciaL,Canales-RodriguezE(2005)Estimatingbrainfunctionalconnectivitywithsparsemultivariateautoregression.PhilosTRSocB360:969 WendlingF,BellangerJ,BartolomeiF,ChauvelP(2000)Relevanceofnonlinearlumped-parametermodelsintheanalysisofdepth-eegepilepticsignals.BiolCybern83:367 WendlingF,BellangerJ,BartolomeiF,ChauvelP(2002)Epilepticfastactivitycanbeexplainedbyamodelofimpairedgabaergicdendriticinhibition.EurJNeurosci15:1499 WhittleP(1963)Onthettingofmultivariateautoregressions,andtheapproximatecanonicalfactorizationofaspectraldensitymatrix.Biometrika50:129 WienerN(1956)Thetheoryofprediction.In:BeckenbachE(ed)ModernMathematicsforEngineers,McGraw-Hill,York,PA,pp165 WinterhalderM,SchelterB,HesseW,SchwabK,LeistritzL,KlanD,BauerR,TimmerJ,WitteH(2005)Comparisonoflinearsignalprocesstechniquestoinferdirectedinteractionsinmultivariateneuralsystems.SignalProcess85(11):2137 WinterhalderM,SchelterB,TimmerJ(2007)Detectingcouplingdirectionsinmultivariateoscillatorysystems.IntJBifurcatChaos17:3725 Zavaglia(2008)Theeffectofconnectivityoneegrhythms,powerspectraldensityandcoherenceamongcoupledneuralpopulations:Analysiswithaneuralmassmodel.IEEETransBiomedEng55(1):6977 ZavagliaM,AstolL,BabiloniF,UrsinoM(2006)Aneuralmassmodelforthesimulationofcorticalactivityestimatedfromhighresolutioneegduringcognitiveormotortasks.JNeurosciMeth157(2):317 ZellnerA(1979)Causalityandeconometrics.In:BrunnerK,MeltzerA(eds)ThreeAspectsofPolicyMaking:Knowledge,DataandInstitutions,Carnegie-RochesterConferenceSeriesonPublicPolicy,vol.10,North-Holland,Amsterdam,Netherlands,pp9 ZhouZ,ChenY,DingM,WrightP,LuZ,LiuY(2009)Analyzingbrainnetworkswithpcaandconditionalgrangercausality.HumBrainMapp30:2197 ZhouZ,WangX,KlahrN,LiuW,AriasD,LiuH,vonDeneenK,WenY,LuZ,XuD,LiuY(2011)Aconditionalgrangercausalitymodelapproachforgroupanalysisinfunctionalmagneticresonanceimaging.MagnResonImaging29:418 131

PAGE 132

BIOGRAPHICALSKETCH ErinBoykin,bornErinTaylor,wasraisedinOrlando,FloridaandgraduatedfromLakeHowellHighSchoolin2001.ShereceivedherB.S.incomputerengineeringfromtheUniversityofFloridain2005,graduatingcumlaude,andcontinuedontoearnherM.S.degreeinelectricalandcomputerengineeringatUFin2007.ShemarriedPatrickOscarBoykinin2009. 132