Assessing hippocampal functional reserve in temporal lobe epilepsy: A multi-voxel pattern analysis of fMRI data

jo u r n al ho me p ag e :w w w . e l s e v i e r . c o m / l o c a t e / e p i l e p s y r e s

Assessing

hippocampal

functional

reserve

in

temporal

lobe

epilepsy:

A

multi-voxel

pattern

analysis

of

fMRI

data

Heidi

M.

Bonnici

,

Meneka

Sidhu

,

Martin

J.

Chadwick

,

John

S.

Duncan

,

Eleanor

A.

Maguire

a_{WellcomeTrustCentreforNeuroimaging,UCLInstituteofNeurology,12QueenSquare,LondonWC1N3BG,UnitedKingdom} b_{DepartmentofClinicalandExperimentalEpilepsy,UCLInstituteofNeurology,QueenSquare,LondonWC1N3BG,}

UnitedKingdom

c_{EpilepsySociety,ChalfontStPeter,Buckinghamshire,UnitedKingdom}

Received19September2012;receivedinrevisedform29December2012;accepted3January2013 Availableonline23January2013

KEYWORDS Epilepsy; Episodicmemory; MVPA; Hippocampus; Hippocampal sclerosis; FunctionalMRI

Summary Assessingthefunctionalreserveofkeymemorystructuresinthemedialtemporal lobes(MTL)ofpre-surgicalpatientswithintractabletemporallobeepilepsy(TLE)remainsa challenge.Conventional functionalMRI(fMRI)memoryparadigmshaveyettofully convince oftheirabilitytoconfidently assesstheriskofapost-surgical amnesia.An alternativefMRI analysismethod,multi-voxelpatternanalysis(MVPA),focusesonthepatternsofactivityacross voxelsinspecificbrainregionsthatareassociatedwithindividualmemorytraces.Thismethod makesitpossibletoinvestigatewhetherthehippocampusandrelatedstructurescontralateral toanyproposedsurgeryarecapableoflayingdownandrepresentingspecificmemories.Here weusedMVPA-fMRItoassessthefunctionalintegrityofthehippocampiandMTLinpatients withlong-standingmedicallyrefractoryTLEassociatedwithunilateralhippocampalsclerosis (HS).Patientswereexposedtomovieclipsofeverydayeventspriortoscanning,whichthey subsequentlyrecalledduringhigh-resolutionfMRI.MTLstructuresweredelineatedandpattern classifierswere trainedtolearn thepatternsofbrain activityacrossvoxelsassociated with eachmemory.Predictablepatternsofactivityacrossvoxelsassociatedwithspecificmemories couldbedetectedinMTLstructures,includingthehippocampus,onthesidecontralateralto theHS,indicatingtheirfunctionalviability.Bycontrast,nodiscerniblememoryrepresentations wereapparentinthesclerotichippocampus,butadjacentMTLregionscontaineddetectable informationaboutthememories.ThesefindingssuggestthatMVPAinfMRImemorystudiesof TLEcanindicatehippocampalfunctionalreserveandmaybeusefultopredicttheeffectsof hippocampalresectioninindividualpatients.

©2013ElsevierB.V.

∗_{Correspondingauthor.Tel.:+442034484347;fax:+442078131445.}

E-mailaddress:[email protected](E.A.Maguire). 1_{Theseauthorsmadeequalcontributions.}

0920-1211 ©2013ElsevierB.V.

http://dx.doi.org/10.1016/j.eplepsyres.2013.01.004

Open access under CC BY license.

Introduction

Hippocampalsclerosis(HS)isacommonsubstratefor refrac-toryfocalepilepsy(SalmenperaandDuncan,2005).Surgical resection carries a 60—70% chance of an individual being seizure free and having improved quality of life (Wiebe etal.,2001;deTisietal.,2011)butbringscognitiverisks, particularly tolanguage and memory, if structuresin the hemisphere contralateral to surgery are not fit for pur-pose(ScovilleandMilner,1957;Baxendale,1998).Formany years the Wada test (or intracarotid amobarbital proce-dure;Wada,1949)wasemployedtohelppredicttheriskof languageimpairmentandamnesiafollowingtemporallobe resection.However, the invasive natureof the Wada test anditsmethodologicalshort-comingshaveprompteda sig-nificant decline in its use in recent years (Loring et al., 1992; Baxendale,2000; Baxendale etal.,2005).In paral-lel,non-invasivefunctionalMRI(fMRI)hasbeenincreasingly deployedtoaidpre-surgicalassessment oflanguage domi-nanceandmemory(Baxendale,2002;Rausch,2002;Binder etal.,2008;Powelletal.,2008;Duncan,2009).

fMRI is a reliable means of lateralising expressive lan-guage dominance in pre-surgical temporal lobe epilepsy patients(Baxendale,2002;Salmenpera andDuncan,2005; Duncan,2009).However,assessingthefunctionalintegrity andreserveof criticalmemorystructuressuchasthe hip-pocampusismoredifficult(Baxendale,2002;Loring,2005; Jones-Gotman,2005;Jones-Gotmanetal.,2008;Baxendale and Thompson, 2010; Bonnelli et al., 2010). This partly stemsfromthe mass-univariateapproach thatis typically employed in fMRI studies. This involves creatinga model oftheexperimentaldesign thatisfittedtothefMRIBOLD response at each voxel independently, the aim being to findactivityinvoxelsthatconsistentlyshowsarelationship withtheexperimentaldesign,averagedovernumerous tri-als.This methodishighlyeffective for investigatingmany typesofbrainfunction,butisgenerallyinsensitiveto fine-grainedlevelsofrepresentationsuchasindividualmemory traces.Thus,inthecase ofmemory,onecannotknowfor certainthatactivationtrulysignalsabrainregion’scapacity torepresentindividualmemories.

In recent years an alternative analysis approach has emergedwhichexploitstheintrinsicallymultivariatenature offMRIdata.Thisismotivatedbytheviewthattheremay beinformationpresentinthedistributedpatternsof activa-tionacrossvoxelsthatismissedwhenlookingateachvoxel independentlyasinthemass-univariatemethod(Haynesand Rees,2006;Normanetal.,2006).Thismultivariateapproach iscommonlyknownasmulti-voxelpatternanalysis(MVPA), or decoding.Itinvolves traininga supportvector machine (SVM)classifiertorecognisethepatternsofactivityacross voxelsassociatedwithparticularstimuli(e.g.specific mem-ories).Theclassifieristhenappliedtoapreviouslyunseen portionof the data thatwas notused for training.If the classifierissuccessfulatpredicting(ordecoding)which par-ticular memory was being recalled in this test data set, this indicates that there is information about that mem-ory represented in thebrain region where thepattern of voxels wassuccessfully identified.MVPA therefore affords theopportunitytoexaminetheneuralsubstratesofspecific memorytracesinindividualparticipants,localisedtobrain regionsofinterest.Recentstudiesinhealthysubjectshave

documented the power of this method for investigating neuralinformationat thelevel of individualmemory rep-resentations (reviewed in Rissman and Wagner, 2012). A number of these studies specifically focused on the hip-pocampusandsurroundingmedialtemporallobes(MTL)and successfullyinterrogatedtheircapacitytorepresentvarious types of memory(Hassabis etal., 2009; Chadwick etal., 2010,2011,2012;Bonnicietal.,2012).

The ability to establish the functional capacity of the contralesional medial temporal lobe to sustain specific memories on an individual patient level has clear impli-cations for surgical planning. Here we describe a novel approach of decoding specific neural signatures of mem-oryrepresentationswithinthemedialtemporallobesusing MVPA-fMRIinpatientswithTLE.Initially,wesoughtto ascer-tainwhetherthisapproachwouldprovideresultsthatwere concordantwithestablishedclinicalfeaturesinindividuals withwell-defined unilateral HS and memoryprofiles con-cordantwiththelateralityoftheirpathology. IfMVPAhas potentialutilityinthecontextofTLE,wewouldexpectthe classifieroperatingonvoxels inthesclerotichippocampus toperformatchance,asviablememorytracesareunlikely tobelaiddownthere.Bycontrast,thecontralesional hip-pocampus should contain memory representations that a classifiercandetect.

We employed a paradigm devised by Chadwick et al. (2010)inwhich participantsviewedthreeshortmoviesof everydayeventspriortoscanning,andthenrecalledthem during high-resolution fMRI. This task had a number of advantages for ourpurpose; the memories involved were realisticandclosetotheexperiencesofday-to-daylife,all participantswere exposed to thesame events permitting greaterexperimentalcontrol,andthetaskwassimpleand suitableforarangeofabilities.Mostimportantly,thistask revealed that decodable informationabout thememories was present bilaterally within the hippocampi, entorhi-nal/perirhinal and parahippocampal cortices in healthy subjects,makingitsuitableforuseinindividualswithboth leftandrightHS.Wetherefore testedthehypothesisthat the contralesional hippocampusand MTL would show the samepatternoffindingsasinthehealthycontrols(Chadwick et al., 2010) while results for the sclerotic hippocampus wouldbeatchance.

Methods

Participants

ThisstudywasapprovedbytheNational Hospitalfor Neu-rologyandNeurosurgeryandtheInstituteofNeurologyJoint ResearchEthicsCommittee,andwritteninformedconsent wasobtainedfromallparticipants,inlinewiththe Declara-tionofHelsinki.Tenpatients(8female;9righthanded)with amedian ageof 41years,interquartilerange16.75, took part (Table 1). Allhad long-standing medically refractory TLEassociatedwithunilateralHS,whichhadbeenidentified on3TstructuralMRIscansfollowingqualitativeassessment byexpertneuroradiologistsandquantificationof hippocam-palvolumesandT2relaxationtimesshowingunilateralHS andapparently normalcontralateral MTL structures.Nine

Table1 Summaryofpatientdetails.

ID Gender Age(yrs) H’ess Cognitivesummarya _Ageatonset

(yrs)

Duration (yrs)

MRI Seizuretypeand frequency 1 F 41 R IQlowaverage;verbalmem

imp;visualmemaverage

32 9 LHS CPS:2—3/month; SGS:2/month 2 F 51 R IQaverage;verbalmem

average;visualmemabove average

20 31 LHS CPS:1/year;SGS:2in thepast

3 M 20 L IQaverage;verbalmem average;visualmem average

13 7 LHS CPS:4—8/month; SGS:2/year 4 F 55 R IQaverage;verbalmem

borderlineimp;visualmem lowaverage

14 41 LHS CPS:2/week;SGS: 2/year

5 F 42 R IQaverage;verbalmem imp;visualmemlow average

11 31 LHS CPS:3—4/month

6 M 29 R IQaverage;verbalmem imp;visualmemaverage

6 23 LHS CPS:1/month;SPS: 1/month;SGS:1/year 7 F 44 R IQlowaverage;verbalmem

imp;visualmemabove average

18 26 LHS PreviouslyCPS: 2—3/month; currentlycontrolled onAED

8 F 52 R Chosenottoattendfor cognitivetesting

12 40 LHS CPS:1/month;SGS: 1/year

9 F 50 R IQaverage;verbalmem imp;visualmemabove average

35 15 LHS CPS:16/month;SGS: 2/year

10 F 37 R IQaverage;verbalmemlow average;visualmemimp

2 35 RHS CPS:4—6/month;SGS S:2/year

yrs=years; H’ess=handedness; R=right; L=left; imp=impaired; mem=memory; LHS=left hippocampal sclerosis; RHS=right hip-pocampal sclerosis; CPS=complex partial seizure; SGS=secondarily generalised tonic—clonic seizure; SPS=simple partial seizure; AED=anti-epilepticdrugs.

a_{Impaired=<2nd percentile; borderline impaired=2nd—9th percentile; low average=10th—24th percentile; average=25th—75th} percentile;aboveaverage=76th—90thpercentile;superior=91st+percentile.See‘‘Methods’’sectionfordetailsofspecific neuropsy-chologicaltests.

individualshadleftHS,andonehadrightHS,andallwere takinganti-epilepticmedicationatthetimeofthestudy.

Neuropsychological assessment (Table 1) showed that theiroverall intellectuallevelwaslowaverage oraverage (WASI-III;Wechsler,1997).ThreepatientswithleftHShad some mild word-finding difficulties (Graded Naming Test; McKennaand Warrington, 2007), but all could read profi-ciently andreadily comprehended and complied withthe taskinstructions.Executivefunction(TheWeiglColour-Form Sorting Test; Weigl,1927) and visual perception (subtests fromtheVisualObjectSpacePerceptionBattery;Warrington andJames,1991)werewithintheaveragerange.Memory (recallandrecognition;visualandverbal)wasassessedusing several British-normed tests: the BMIPB (Coughlan etal., 2007),the Camden Memory Tests(Warrington, 1996),and theRecognition Memory Tests (Warrington, 1984). In gen-eralmemoryscoresechoedthelateralityofpathology,i.e. impaired or weakerverbalcompared tovisual memoryin patientswithleftHS,andtheoppositepatterninthepatient withrightHS(Table1).

Our main interest was in comparing the affected MTL withthecontralesionalMTLwithinsubjects,withpatients

effectivelyactingastheirowncontrols.Inthiswayitwas possibletocontrolcompletelyfor factorssuchasage,IQ, seizurehistoryandmedicationregimen.However,inorder toestablishhowthepatientsperformedmoregenerally,we alsocomparedtheirfMRIdatatoagroupofhealthycontrol participants(alson=10;sixfemale;meanage21.1years, SD 1.8) who had performed the same task, in the same scannerusingthesamescanningparameters,anddata anal-ysismethod.Thesecontrolparticipantsalsogaveinformed written consent to participation in accordance with the localresearchethicscommittee,andtheirdatahavebeen reportedpreviously(seeChadwicketal.,2010).

Pre-scantraining

Theexperimentalprotocolwasthesameasthatemployed byChadwicketal.(2010),andthekeyfeaturesarereprised herefor convenience.Duringapre-scan trainingperiod,a patientviewedthreefilmclipsofeverydayevents.Eachclip was7slongandfeaturedawoman(adifferentoneineach clip)carryingoutashortseriesofactions.The filmswere

Figure1 Theexperimentalprotocol.(A)Stillphotographstakenfromoneofthefilmclipsviewedduringpre-scantraining.The clipdepictedawomantakingadrinkfromadisposablecoffeecupandthenputtingitinabin(trashcan).(B)Timelineofanexample trialduringfMRIscanning.

shotoutdoorsinthreedifferenturbansettings.Thesestimuli ensured that memories would be episodic-like in nature, and that allparticipants recalled the same set of memo-ries.Oneclipfeaturedawomantakingaletteroutofher handbag,postingitinapostbox(mailbox),andthen walk-ingoff.Anotherclipfeaturedawomantakingadrinkfrom a disposablecoffee cup, putting the cup in a rubbish bin (trashcan), and then walking off (Fig. 1A). The final clip featured a woman picking up a bicycle that was leaning againstsomerailings,adjustingherhelmetandwalkingoff withthe bicycle. Patients viewedeach clip 15 times and practiced recalling them asvividly aspossible and within the 7s timeframe. Patientswere alsotrained ona rating procedurewheretheyevaluatedeachrecallperiodforthe levelof vividnessaswell ashowconsistenttherecall had been relativetoprevious recollectionsof the same mem-ory.Thispracticeperiodthereforeensuredthatthememory tracesforthethreemovieswerestablebeforegoingintothe scanner,aprerequisiteforthesubsequentMVPAanalysis.

Scanningtask

Duringscanningpatientsrecalledeachmovietwentytimes inapseudo-randomorder,ensuringthatthesamememory wasnotrepeatedtwoormoretimesinarow.Oneachtrial (seeFig.1B)thepatientviewedacueonscreenindicating whichofthethreefilmeventstheyshouldrecall.Patients were instructed not tobegin the recall process until the cuetoclosetheireyesandrecallappeared.Toensurethe patients were concentrating, and that the recall approx-imated the original 7s length of a clip, the patient had topressa button(usingascanner-compatiblebutton-box) whentheyhadfinishedrecallingtheclip.Ifthebuttonwas

pushedtoosoon(<5.5s)ortheyfailedtopushitwithin10s thenatonewouldsound,andamessagewouldappearfor 1.5s indicating that their recall had been too fastor too slow. Any suchtrials were excluded fromthe subsequent analysis.Ifthepatientpressedthebuttonbetween5.5and 10s,a fixationcrossappeared onscreenfor 1.5s. Patients weretrainedtoopentheireyesassoonastheyhadpressed thebutton or if theyheard the tone. Following this, the patientwasrequired toprovide ratingsabout the preced-ingrecalltrialusingthebutton-box,justastheyhadbeen trainedtododuringthepre-scantrainingsession.First,they ratedhowvividtheprecedingrecalltrialwas(scale:1—5, where1wasnotvividatall,and5wasveryvivid).Second, theyrated howaccurately therecalled memoryreflected theactualfilmclip(scale:1—5,where1wasnotaccurateat all,and5wasveryaccurate).Anytrialswhereaparticipant recordedaratingoflessthan3wereexcludedfromthe sub-sequentanalysis.Followingtheratings,participantsrested for4sbeforestarting thenexttrial. Once excludedtrials werediscounted(andsimilartothecontroldatainChadwick etal.,2010),thisresultedinanaverageof12.3(SD1.34) tri-alsformemory1,12.5(SD2.42)trialsformemory2and12.9 (SD1.97)trialsformemory3makinganaverageof36trials intotalperpatientbeingenteredintotheMVPAanalysis.

Dataacquisition

Usinghigh-resolution fMRI(Carretal.,2010),weacquired datain a limited volume focussing onthe medial tempo-rallobe,particularlythehippocampusentorhinal/perirhinal and posterior parahippocampal cortices. A 3T Magnetom AllegraheadonlyMRIscanner(SiemensHealthcare, Erlan-gen,Germany)operatedwiththestandardtransmit-receive

headcoilwasusedtoacquirethefunctionaldatawitha T2*-weighted single-shot echo-planar imaging (EPI) sequence in a single session (in-plane resolution=1.5mm×1.5mm; matrix=128×128; field of view=192mm×192mm; 35 slicesacquiredininterleavedorder;slicethickness=1.5mm withno gap between slices; echo time TE=30ms; asym-metric echo shifted forward by 26 phase-encoding (PE) lines;echospacing=560␮s; repetition timeTR=3.5s; flip angle ˛=90◦). All data were acquired at 0◦ angle in the anterior-posterior axis. An isotropic voxel size of 1.5mm×1.5mm×1.5mmwaschosenforanoptimal trade-offbetweenBOLDsensitivityandspatialresolution.Further, theisotropicvoxeldimensionreducedre-samplingartefacts whenapplying motion correction.To ensureoptimal data quality,images werereconstructed online and underwent online quality assurance (Weiskopf et al., 2007). For dis-tortioncorrection (Hutton et al., 2002), field maps were acquiredwiththestandardmanufacturer’sdoubleecho gra-dient echo field map sequence (TE=10.0 and 12.46ms, TR 1020ms; matrix size, 64×64), using 64 slices cov-ering the whole head (voxel size 3mm×3mm×3mm). In addition to the functional scans, a whole brain T1-weighted3DFLASHsequencewasacquiredwitharesolution of1mm×1mm×1mm. AT1-weighted high-resolution 3D modifieddrivenequilibriumFouriertransformwhole-brain structural MRI scan was acquired for each patient after the main scanning session with 1mm isotropic resolution (Deichmannetal.,2004).

Regionofinterest(ROI)segmentation

Givenourspecificinterestinthehippocampusandmedial temporal lobes, manual segmentation of the hippocam-pus(HC),entorhinal/perirhinalcortex(EPC,combined)and parahippocampal cortex (PHC) was performed using ITK-SNAP (Yushkevich et al., 2006 — www.itksnap.org) using the1mm×1mm×1mmT1structuralimages(seeexample in Fig.2). Hippocampal anatomy wasidentified usingthe Duvernoyhippocampusatlas(Duvernoy,2005).The entorhi-nal/perirhinal cortex and parahippocampal cortex were segmentedaccordingtotheprotocol describedbyInsausti etal.(1998).AuthorHMBperformedthesegmentations,and inter-raterreliabilitywascalculatedonarandomselection of four patient scans which had also been segmented by authorMS. The Diceoverlapmetrics (Dice,1945)were as follows: HC 0.81, EPC 0.63, PHC 0.60. Mean volumes (in cubicmm)andstandarddeviations(SD)fortheROIsinthe left(L)andright(R)MTLinthegroupofnineLHSpatients were as follows: HCL 2435.11 (SD 467.94), HCR 3415.89 (453.01), EPCL 3085.33 (453.01), EPCR 3159.33 (326.71), PHCL1591.22(386),andPHCR1543.56(272.99).Arepeated measuresANOVArevealedasignificanthemispherebyregion interaction(F=24.37, p<0.0001).Furtherinvestigation of thiseffectusingpairedt-testsconfirmedthatavolume dif-ferencebetweentheleftandrighthippocampusdrovethis result(t=−6.06,p<0.0001)withtherighthippocampal vol-ume greater than the left. There were no differences in volumebetweentheleftandrightsidesfortheotherMTL structures(EPC:t=−1.07,p=0.32;PHC:t=0.42;p=0.69). ForthepatientwithrightHS,thehippocampalvolume dif-ferencewasalsoapparent,but withtheleftgreaterthan

theright:HCL3503,HCR2059,EPCL2569,EPCR2956,PHCL 1297,andPHCR1433.

Imagepre-processing

Pre-processingofthefMRIdatawasperformed usingSPM8 (www.fil.ion.ucl.ac.uk/SPM).ThefirstsixEPIvolumeswere discardedtoallowforT1equilibrationeffects(Frackowiak etal.,2004).TheremainingEPIimageswerethenrealigned tocorrectformotioneffects,andminimallysmoothedwith a 3mm FWHM Gaussian kernel. A linear detrendwas run on the images to remove any noise due to scanner drift (LaConteetal.,2005).Nextthedatawereconvolvedwith the canonical haemodynamic response function (HRF) to increasethesignal-to-noiseratio(Frackowiaketal.,2004). ThisHRFconvolutioneffectivelydoubledthenaturalBOLD signal delay,givingatotaldelay ofapproximately12s.To compensate for this delay, all onset times were shifted forwardintimebythreevolumes,yieldingthebest approx-imationtothe12sdelaygivenaTRof3.5sandroundingto thenearestvolume(Dudaetal.,2001).Functionalvolumes were extracted fromthevivid recall periodof eachtrial, leadingtoatotaloffourfunctionalvolumespertrial.

Multi-voxelpatternanalysis

A support vector machine classifier wascreated for each regionofinterest.Eachclassifierwastrainedonaportion of the fMRI data relating tothe three episodic memories andthentestedonanindependentsetofinstancesofthese memories. We used a standard MVPA procedure that has been describedelsewhere (Chadwicketal., 2010;Bonnici et al., 2012; for an in-depth review of MVPA in the hip-pocampus see Chadwick etal., 2012). Briefly, theoverall classificationprocedureinvolvedsplittingthefMRIdatainto two segments: a ‘‘training’’ set usedto train a classifier withfixedregularisationhyperparameterC=1,inorderto identify responsepatterns relatedto the memoriesbeing discriminated,anda‘‘test’’setusedtoindependentlytest theclassificationperformance(Dudaetal.,2001),usinga cross-validationprocedure.Weusedastandardk-fold cross-validationtestingregimewhereinkequalledthenumberof experimentaltrials,withthedatafromeachtrialsetaside inturnasthetestdata,andtheremainingdatausedasthe trainingset.ThisthereforegeneratedksetsofSVMtraining andtestsetswhichproducedanoverallclassification accu-racyfromtheproportionofcorrectclassification‘‘guesses’’ acrossallkfoldsof thecross-validation.Theclassification wasperformedusingtheLIBSVMimplementation(Changand Lin,2011).

Prior to multivariate classification, feature selection (GuyonandElisseeff,2003)wasperformedonthedatafrom the trainingset (therebyensuringthatthis stepwasfully independent fromfinal classification, which is criticalfor avoiding ‘‘double-dipping’’ — Kriegeskorte et al., 2009). The purpose of feature selection is to reduce the set of features (in this case, voxels) in a dataset tothose most likelytocarryrelevantinformation. Thisiseffectivelythe sameasremovingvoxelsmostlikelytocarrynoise,andisa way of increasing thesignal-to-noise ratio. This was con-ducted using a standard multivariate searchlight strategy

Figure2 Brainregionsofinterest.Thetoppanelshowssagittalviewsthroughtheleftandrighthemispheres,andthelowerpanel acoronalsection,fromthestructuralMRIscanofoneofthepatientswithleftHS;L=leftsideofthebrain,R=rightsideofthe brain.Thelefthippocampus(sclerosed)iscolouredinred,therighthippocampusingreen,theleftEPCinblue,therightEPCin yellow,theleftPHCinturquoise,andtherightPHCinmagenta.(Forinterpretationofthereferencestocolourinthisfigurelegend, thereaderisreferredtothewebversionofthearticle.)

withinthegivenROI.Foragivenvoxel,wefirstdefineda smallspherewitha radiusof threevoxels centredonthe givenvoxel(Kriegeskorteetal.,2006;Hassabisetal.,2009; Chadwicketal.,2010,2012;Bonnicietal.,2012).Notethat thesphereswererestrictedsothatonlyvoxelsfallingwithin thegivenregionofinterestwereincluded.Therefore,the shapeofthesphereandthenumberofvoxelswithinit var-ieddependingontheproximity totheregionofinterest’s borders. Thisprocedure thenallowedthe selectionof the searchlightvoxelsetthatcontainedthegreatestdegreeof decodinginformationwithinthetrainingdataset.Usingthis voxelsubset,theSVMclassifierwastrainedtodiscriminate betweenthethreeepisodicmemoriesusingthe‘‘training’’ dataset,andtestedontheindependent‘‘test’’dataset.

StandardSVMsarebinaryclassifiersthatoperateon two-classdiscriminationproblems,whereasourdatainvolveda three-class problem (i.e. three memories). The SVM can, however,bearbitrarilyextendedtoworkincasesinwhich there are more than two classes. Typically this is done by reducing the single multiclass problem into multiple binaryclassificationproblemsthatcanbesolvedseparately andthen recombined toprovide thefinal classprediction (Allwein etal.,2000). Weusedthe well-establishedError Correcting Output Codesapproach (Dietterich and Bakiri, 1994)andcomputingof theHammingdistance (Hamming, 1950;Hassabisetal.,2009;Chadwicketal.,2010,2011).

Dataanalysis

The classifier accuracy valuesfor each brain region were comparedtochanceusingone-tailedt-tests.Becausethere werethreememories,chancewas33%inthisstudy.Other

within-subjects comparisons were made using repeated measuresANOVAandpaired t-tests.Wereportformal sta-tistical analyses for the group (n=9) of left HS patients, andpresenttheindividualdataoftherightHSpatient.We alsocomparedtheperformanceofthepatients(n=9with lefthippocampalsclerosis)withthe10healthyparticipants reportedbyChadwicketal.(2010)usingANOVA.Athreshold ofp<0.05wasemployedthroughout.

Results

Foreachbrainregionofinterest (see‘‘Methods’’section) aclassifierwasfirsttrainedonaportionof thefMRIdata relatingtothethreeepisodic-likememoriesandthentested onanindependentsetoftrialsofthesememories.A clas-sificationresultsignificantlyabovechance (33%)wouldbe expected if information was present in the patterns of fMRIactivitythatenableddiscriminationbetweenthethree memories.WeperformedthisanalysisforeachROIinboth hemispheres.

Inordertocontextualisethecurrentfindings,itisworth notingthatwhen usingthis paradigmin healthysubjects, Chadwick et al. (2010) found that all three MTL regions producedclassificationresultsthatweresignificantlyabove chance, showing that it was possible to detect informa-tionabout individualepisodic-like memoriesfromactivity patternswithinthehippocampusandsurroundingMTL(see Fig.3A). Theresults werehighlysimilarfor leftand right MTL. They also found that classification accuracy for the hippocampus was significantly better than for the other MTLregions,suggestingthatepisodic-likememoriesmaybe

Figure3 Results ofthe MVPAanalyses. (A)Control data (n=10, collapsed across hemispheres)from Chadwick etal. (2010)

usingthesametaskasthatemployed inthecurrentstudy.Meanclassifieraccuracyperformances(±onestandarderrorofthe mean)areshown;chancewas33%(representedbytheblackdashedline).HC=hippocampus;EPC=entorhinal/periphinalcortex; PHC=parahippocampalcortex.(B)DataforthecontralesionalrightMTL(fromthe9patientswithleftHS).(C)Resultsfortheaffected leftMTL,whereclassifierperformanceforthesclerosedhippocampus(highlightedinredonthex-axis)wasnotsignificantlydifferent tochance.ClassifierperformancesintheotherleftMTLregionsandallareasinthecontralesionalMTLweresignificantlyabove chance.(Forinterpretationofthereferencestocolourinthisfigurelegend,thereaderisreferredtothewebversionofthearticle.)

moredistinctwithinthehippocampusthanthesurrounding cortex.

In a similar vein, we found that in the patients, clas-sifiers operating onthe patterns of activity across voxels in each regionof the contralesional MTL predicted which memory was being recalled significantly above chance (HCR:t=7.299,p=0.0001;EPCR:t=2.206,p=0.029;PHCR: t=2.450;p=0.020)(Fig.3B).Classificationaccuracywas sig-nificantlyhigherin thehippocampuscomparedtotheEPC (t=2.551, p=0.034; not significant compared to the PHC t=1.788,p=0.112).Thisshowsthattheepisodic-like mem-oriesof themovieswererepresented inthe hippocampus inparticular,andalsoinadjacentMTLstructures, suggest-ingoperationalmemorycapacityofthesecontralateralMTL regions.

Bycontrast,theclassifieraccuracy withinthe sclerotic left hippocampus for the three memories was not sig-nificantly greater than chance (HCL: t=1.225, p=0.128; Fig. 3C), while classifier performances in the surround-ing MTL regions were significantly above chance (EPCL: t=3.067,p=0.008;PHCL:t=3.444,p=0.004).Classification accuracy was significantly higher in the parahippocam-pal cortex than the sclerotic hippocampus (t=−3.853, p=0.005). This suggests that discernable memory repre-sentationswere absent in thesclerotic hippocampus, but presentinadjacentMTLstructures.

Having examined MTL structures in each hemisphere separately, we then made direct comparisons using a repeated-measures ANOVA. There was a significant inter-actionbetween thetwofactors ofhemisphereand region (F=4.561; p=0.026) that was driven by a significant dif-ference in classifier performance between the sclerotic andcontralesionalhippocampi(t=−3.989,p=0.004),with theclassifieroperatingonvoxels inthe scleroticleft hip-pocampusunabletodistinguishbetweenthethreeepisodic memories.

Eight of the nineleft HS patientsshowed a difference inperformance between thehippocampi inthe predicted direction; the performance of the classifier operating on voxelsinthecontralesionalrighthippocampuswasgreater

than that of the sclerotic left hippocampus (mean clas-sifier performance HCL=34.94%, SD=7.16; HCR=46.10%, SD=5.51).DatafortheonepatientwithrightHS wasalso intheexpecteddirection,withdeceasedclassifieraccuracy withinthe sclerosedright hippocampus(HCL 44.44%,HCR 36.67%).Wealsoexaminedwhethertheat-chanceresultin thesclerotichippocampimighthavebeenassociatedwithits smallervolume.Nocorrelationbetweenclassification accu-racyandhippocampalvolumewasfound(Pearsonr=0.511, p=0.16).

The ninth left HS patient’s results were equivocal (patient 9: HCL=44.95%, HCR=43.12%; volumes for this patient: HCL 2970, HCR 3478). To investigate if this was relatedtoclinicalfactors,correlationsofclassifieraccuracy withvariablessuchaspatients’currentage,ageofepilepsy onset, epilepsyduration andseizure frequency were per-formed.Nosignificantcorrelationswereobserved,although thismaybeduetotherelativelysmallsamplesize.

Finally, we directly compared the fMRI data from the nine left hippocampal sclerosis patients with that of the ten healthy control participants reported by Chadwick etal.(2010).Despitethecontrols beingyoungerthanthe patients, there were no significant differences in classi-fier accuracies between the patient and controls for any MTL region except the left hippocampus, where classi-fier accuracy was significantly less in the patients (HCL F=6.52, p=0.02; HCR F=0.001, p=0.97; EPCL F=1.57, p=0.23;EPCRF=0.94,p=0.35;PHCLF=0.7,p=0.41;PHCR F=1.3, p=0.27). Ofparticular note, therewas no differ-encebetweenthegroupsinclassifieraccuracyfortheright hippocampus.Thesefindingsthereforeaffirmtheviewthat thescleroticlefthippocampiinthepatientswere dysfunc-tional,whiletherighthippocampiretainedanormallevel ofmnesticfunctioning.

Discussion

Wefound thatpredictable patternsof activityacross vox-els associated with specific memories could be detected

inMTL structures,includingthehippocampus, onthe side contralateral to focal unilateral hippocampal sclerosis in a group of patients with TLE.By contrast,no discernible memoryrepresentationswereapparentinthesclerotic hip-pocampus,althoughtheadjacentMTLregionsonthatside containeddetectableinformationabout thememories.To our knowledge, this is the first MVPA-fMRI study to be reportedinTLE,andthereforeshowsthat theMVPA tech-nique,designedtodetectspecificmemoryrepresentations inpatternsoffMRIdata,permitsinterrogationofMTL func-tionalityandinparticularhippocampalfunctionalreserve, complementingexistinginvestigativeprotocolsinTLE.

Theparadigmweemployedinvolvedexposingallpatients tothesamesetof moviesthat depictedeverydayevents. Memoriesofthesenaturalisticstimulihavebeenfoundtobe encodedandrepresentedintheleftandrighthippocampus ofhealthycontrols,suchthatrecallofthesememoriescould bepredictedfrompatternsoffMRIbrainactivityacross hip-pocampalvoxels(Chadwicketal.,2010).Bycapitalisingon thebilateralnatureofthememorytracesassociatedwith theseepisodic-likememories,wecouldexamineand com-parethememorycapacityofthehippocampiinpatientsin whomonehippocampuswassclerotic.Inthisstudywe there-foredeliberatelyfocussedonpatientswherethestructural imagingand neuropsychologicaltests pointed toselective involvementofonesclerotichippocampus.Thatthis dam-agecompromisedthemnesticcapacityofthathippocampus wasclearintheat-chanceperformanceofthepattern classi-fier,whichcouldnotdiscernanyreliablepatternsofactivity associatedwiththememories.Thiswasindirectcontrastto thecontralateralhippocampusinwhichthememorytraces were readily detected. Moreover, the contralateral hip-pocampuscontainedsimilaramountsof informationabout thememorieswhenthepatientswerecomparedtohealthy controlparticipants, while the sclerotichippocampuswas severelycompromised.Overall,therefore, thewithin-and between-subjects effects we observed suggest the MVPA approachisreliable,andoffersnewinsightsinto hippocam-palfunctionalreserveinTLE.

This methodwasnotonlyabletointerrogatethe func-tionalstatusofthehippocampus,butalsothesurrounding medialtemporallobetissue.Theentorhinal/perirhinaland parahippocampalcorticesinthecontralesionalhemisphere ofthepatientscontaineddecodableinformationaboutthe memories,andlesssothanthehippocampus,confirmingthe functionalintegrity,andfunctionalhierarchy,ofthewider MTL memory system onthat side. Of particular note was the status of the EPC and PHC on the affected side, as information about the memories was detectable in these structures.ThisshowsthatinthepresenceofHS,adjacent MTLregionscanretainacapacitytocontributetomemory recall.TheclassifieroperatingonvoxelsinthePHConthe affectedsideseemedtoperformbetterthanitscounterpart ontheunaffectedside (Fig.3BandC),perhapsindicating acompensatory mechanism.However,whendirectly com-pared,therewasnosignificantdifferencebetweenthetwo (t=0.97;p=0.36),andnodifferencewhencomparedtothe controlparticipants.

HowdoourMVPAfindingsrelatetoresultsfrom conven-tional fMRI memory studies in TLE? After all, asymmetric MTLactivationshavealsobeenobservedusingstandardfMRI paradigms,andmuchinformationcanbegleanedfromsuch

studies. For instance, in tasks that were associated with bilateralactivitywithintheMTLinhealthycontrol partici-pants,patientsshowedsignificantlyreducedMTLactivation ipsilesionally(e.g. Detreet al.,1998;Jokeit etal.,2001; Mechanic-Hamiltonetal.,2009).Inacomplexscene encod-ingtasksevenpatients withleftTLE andthreewithright TLEwerestudiedusingasimpleblockdesignfMRIprotocol. Patientsshowedasymmetryofactivationwithgreater acti-vation in thenon-lesional MTL. However, fMRIactivations werenotcorrelatedwithneuropsychologicalperformance, therefore it could not be definitively ascertained if con-tralateralactivationswerespecificallyrelatedtosuccessful memoryencoding(Detreetal.,1998).Jokeitetal.(2001) employed a spatial navigation task in a group of thirty patientswithrefractoryTLE.Asymmetryofactivationwas seen in 90% of patients. Rank correlation was performed between the numberof activatedvoxels within the MTL, outofscannerneuropsychologicalvariables,andWadatest hemispheric memory performance. Correlations between visuospatialmemoryperformanceandrightMTLactivation butnotleftMTLactivationwerefound.LeftMTLactivation wascorrelatedwithWadahemisphericmemoryperformance forvisuallypresentedobjects.Thesecorrelationssuggested that MTL activations during the task were related tothe memoryencodingprocess.

These examplesillustratethat conventional fMRIBOLD activationstypicallyindextheinvolvementofbrainregions inanexperimentaltask.However,tomoredirectlylinkthe engagementofabrainareatosuccessfulmemoryfunction, furthercorrelationwithneuropsychologicalvariables(asin Jokeitetal.,2001)orsubsequentmemoryparadigmsusing, forexample,anevent-relatedanalysisisnecessary(Powell etal.,2007). MVPA, however,offers the importantadded valueofbeingabletoassessthefunctionofMTLstructures attheleveloffine-grained individualmemory representa-tions,whichisamoredirectreflectionofthecapacityofthe MTLstructures.Thismethodprovidestheadditionalbenefit ofbeingabletodistinguishspecificneuralrepresentations withinrelatively smallsub-regions ofthemedialtemporal lobeinindividualpatients,therebyofferinganother dimen-sion to delineating the functional topography of memory representationsinthecontextofTLE.

Conventional fMRI memory encoding studies have also examinedmaterial-specificeffectsandsomehaveobserved memory reorganisation involving the contralesional MTL (Golbyetal.,2002;Powelletal.,2007).InleftTLEpatients subsequentmemory studies usingan event-related analy-sis showed reorganisation of verbal memory encoding to thecontralateralhippocampus(Richardsonetal.,2003)and in right TLE patients reorganisation of visual encoding to involvethelefthippocampus(Powell etal.,2007).Itwill beinterestingforfuturestudiestoemployMVPAinthis con-texttoinvestigatethecapacityforeffectivereorganisation inunilateralTLE.

Ina similarvein,pre-operative conventional fMRI acti-vations have been shown to be important predictors of post-surgical memory outcome in both material-specific memory paradigms (Bonnelli et al., 2010; Powell et al., 2008;Richardsonetal.,2004)andtasksthataremore bilat-erally represented within the MTL (Janszky et al., 2005; Mechanic-Hamiltonetal.,2009)ingroupstudies.Because MVPA indexes the presence of predictable and useful

informationpertainingtospecificmemories,andgiventhat wehaveshownevidenceforfunctionalcapacitywithinthe MTLevenonasinglesubjectbasis,thisleadstothe excit-ingpossibility that pre-operative MVPA couldalso predict post-operativememorychangein patientgroupsbut also, crucially,atthelevelofindividualpatients.

Overall,ourfindingsindicatethatMVPA-fMRIcouldprove ausefulnon-invasivemethodofassessingpre-surgical mem-orycapacitywithintheMTL.Despiteoursmallsamplesize, ourresultsillustratethattheMVPAapproachtofMRImemory studiesin thecontextof TLEcangiveaninsight into hip-pocampalfunctionalreserve.WhilethisMVPA-fMRImethod for assessing memory reserve shows promise, our study reveals clear targets for futureinquiries. Here we exam-inedmemoryrecallin individualswithclear-cutunilateral HS(mostlyleft-sided)andmemoryprofilesthatechoedthe lateralityofthispathology.Thiswasthelogicalfirststepin evaluatingMVPA-fMRIbecauseiftheresultsdidnotconcur withknownpathology,thiswoulddiscourageitsfutureuse. Asitis,MVPAmaybeusefulnotonlyinadditiontoexisting structuralandfunctionalMRIandneuropsychologicaldatain straightforwardcases, butcouldplaya particularlyuseful role in patients in whom existing information is ambigu-ous(Baxendaleetal.,2005).Inthosesituations,MVPA-fMRI could give insights into the memory reserve of the hip-pocampiand surroundingMTL structuresthat mayhave a direct bearing on surgical decisions. Further work is now requiredtoassessthevalidityoftheMVPA-fMRIapproachin largercohorts,andpatientswithunilateralleftandrightHS aswell as more complex cases. It will alsobe important to explore the efficacy of MVPA-fMRI with other memory paradigms, to investigate the concordance of MVPA with clinicaldata,andinparticularthelinkbetweentheabsence ofpatternrecognitionatthelevelofthesclerotic hippocam-pusandpost-surgicalmemorystatus.

Finally, MVPA asit iscommonly implemented including itsusehere,involvestrainingsupportvectormachine clas-sifiers. This requires multiple presentations of the same stimuli in order to accrue enough examples of the brain activity patterns associated with each stimulus for train-ingtobeviable. Assuch,thetechniquedepends onbrain activitybeingstablefromoneinstanceofstimulus presenta-tiontothenext.Certaintypesofdesign,therefore,become morechallenging,includinglearningparadigmsinwhichthe neuralsignaturesofstimulimaybemoredynamic.However, decodingmethodsareconstantlyevolving(Chadwicketal., 2012)anditislikelythatopportunitiesfornewapplications inthecontextofpre-surgicalevaluationofTLEpatientswill arise.

Conflict

of

interest

Noneoftheauthorshasanyconflictofinteresttodisclose.

Acknowledgements

ThisworkwassupportedbytheWellcomeTrustviaaSenior ResearchFellowship(WT084218)toEAM,andaProgramme Grant(WT083148) toJSD. We thankSallie Baxendaleand PamThompsonfortheirneuropsychologicalexpertise,and

JaniceGlensmanandKristjanJohannessonforimaging sup-port.

References

Allwein,E., Shapire,R., Singer,Y., 2000.Reducing multiclassto binary: a unifying approach for margin classifiers. J. Mach. Learn.Res.1,113—141.

Baxendale,S.,1998.Amnesiaintemporallobectomypatients: his-toricalperspectiveandreview.Seizure7,15—24.

Baxendale, S.A., 2000. Carotid amytal testingand otheramytal procedures.In:Oxbury,JM.,Polkey,CE.,Duchowny,M.(Eds.), IntractableFocalEpilepsy:MedicalandSurgicalTreatment.WB Saunders,London,pp.627—636.

Baxendale,S.,2002.TheroleoffunctionalMRIinthepresurgical investigationoftemporallobeepilepsypatients:aclinical per-spectiveandreview.J.Clin.Exp.Neuropsychol.24,664—676. Baxendale,S.,Thompson,P., 2010.Beyondlocalization:therole

of traditionalneuropsychological tests in anage of imaging. Epilepsia51,2225—2230.

Baxendale,S.,Thompson,P.,Duncan,J.S.,2005.Theroleofthe Wadatestinthesurgicaltreatmentoftemporallobeepilepsy: aninternationalsurvey.Epilepsia49,715—729.

Binder,J.R.,Sabsevitz,D.S.,Swanson,S.J.,Hammeke,T.A., Ragha-van,M.,Mueller,W.M.,2008.UseofpreoperativefunctionalMRI topredictverbalmemorydeclineaftertemporallobeepilepsy surgery.Epilepsia49,1377—1394.

Bonnelli,S.B.,Powell,R.H.W.,Yogarajah,M.,Samson,R.S.,Symms, M.R.,Thompson,P.J.,Koepp,M.J.,Duncan,J.S.,2010. Imag-ingmemoryintemporallobeepilepsy:predictingtheeffectsof temporalloberesection.Brain133,1186—1199.

Bonnici, H.M., Kumaran, D., Chadwick, M.J., Weiskopf, N., Hassabis, D., Maguire, E.A., 2012. Decoding representations of scenes in the medial temporal lobes. Hippocampus 22, 1143—1153.

Carr, V.A., Rissman, J., Wagner, A.D.,2010. Imaging the human medial temporal lobe with high-resolution fMRI. Neuron 65, 298—308.

Chadwick,M.J., Hassabis,D.,Weiskopf,N., Maguire,E.A.,2010. Decodingindividualepisodicmemorytracesinthehuman hip-pocampus.Curr.Biol.20,544—547.

Chadwick, M.J., Hassabis, D., Maguire, E.A., 2011. Decoding overlappingmemoriesinthemedialtemporallobesusing high-resolutionfMRI.Learn.Mem.18,742—746.

Chadwick, M.J., Bonnici, H.M., Maguire, E.A., 2012. Decoding informationin thehuman hippocampus:a user’s guide. Neu-ropsychologia50,3107—3121.

Chang,C.C.,Lin,C.J.,2011.LIBSVM:alibraryforsupportvector machines.AVMTransactionsonIntelligentSystems2.

Coughlan, A.K., Oddy, M., Crawford, J.R., 2007. Information ProcessingBattery.BrainInjuryRehabilitationTrust,London. Deichmann,R.,Schwarzbauer,C.,Turner,R.,2004.Optimisationof

the3DMDEFTsequenceforanatomicalbrainimaging:technical implicationsat1.5and3T.Neuroimage21,757—767.

deTisi, J., Bell, G.S., Peacock, J.L., McEvoy, A.W., Harkness, W.F.,Sander,J.W.,Duncan,J.S.,2011.Thelong-termoutcome of adult epilepsy surgery, patternsof seizure remission, and relapse:acohortstudy.Lancet378,1388—1395.

Detre, J.A., Maccotta, L., King, D., Alsop, D.C., Glosser, G., D‘Esposito,M.,Zarahn,E.,Aguirre, G.K.,French,J.A.,1998. Functional MRI lateralisation of memory in temporal lobe epilepsy.Neurology50,926—932.

Dice,L.R.,1945.Measuresoftheamountofecologicassociation betweenspecies.Ecology26,297—302.

Dietterich,T.D.,Bakiri,G.,1994.Solvingmulticlasslearning prob-lemsviaerror-correctingoutputcodes.J.Artif.Intell.Res.2, 263—286.

Duda, O.R.,Hart,P.E., Stork,D.G.,2001. PatternClassification. Wiley,NewYork.

Duncan,J.,2009.Thecurrentstatusofneuroimagingforepilepsy. Curr.Opin.Neurol.22,179—184.

Duvernoy, H.M., 2005. The Human Hippocampus, Functional Anatomy,VascularizationandSerialSectionswithMRI.Springer, Berlin.

Frackowiak,R.S.J.,Friston, K.J.,Frith,C.D.,Dolan,R.J., Price, C.J.,Zeki,S.,Ashburner,J.T.,Penny,W.D.,2004.HumanBrain Function.ElsevierAcademicPress,NewYork.

Golby, A.J., Poldrack, R.A., Illes, J., Chen, D., Desmond, J.E., Gabrieli,J.D.,2002.Memorylateralizationinmedialtemporal lobeepilepsyassessedbyfunctionalMRI.Epilepsia43,855—863. Guyon,I.,Elisseeff,A.,2003.Anintroductiontovariableand

fea-tureselection.J.Mach.Learn.Res.3,1157—1182.

Hamming, R.W., 1950. Error-detectingand error-correcting. Bell Syst.Techn.J.29,147—160.

Hassabis, D., Chu, C., Rees, G., Weiskopf, N., Molyneux, P.D., Maguire,E.A.,2009.Decodingneuronalensemblesinthehuman hippocampus.Curr.Biol.19,546—554.

Haynes,J.D.,Rees,G.,2006.Decodingmentalstatesfrombrain activityinhumans.Nat.Rev.Neurosci.7,523—534.

Hutton,C.,Bork,A.,Josephs,O.,Deichmann,R.,Ashburner,J., Turner,R.,2002.ImagedistortioncorrectioninfMRI:a quanti-tativeevaluation.Neuroimage16,217—240.

Insausti,R.,Juottonen,K.,Soininen,H.,Insausti,A.M.,Partanen, K.,Vainio,P.,Laakso,M.P.,Pitkänen,A.,1998.MRvolumetric analysisofthehumanentorhinal,perirhinal,andtemporopolar cortices.AJNR19,659—671.

Janszky,J., Jokeit,H.,Kontopoulou,K.,Mertens,M.,Ebner, A., Pohlmann-Eden,B.,Woermann,F.G.,2005.FunctionalMRI pre-dictsmemoryperformanceaftertightmesiotemporalepilepsy surgery.Epilepsia46,244—250.

Jokeit,H.,Okujava,M.,Woermann,F.G.,2001.MemoryfMRI later-alizestemporallobeepilepsy.Neurology57,1786—1793. Jones-Gotman,M.,2005.TheroleoftheWadatestinthesurgical

treatmentoftemporallobeepilepsy:aninternationalsurvey—– commentaryonBaxendaleetal.Epilepsia49,725.

Jones-Gotman, M., Smith, M.L., Wieser, H.G., 2008. Intraarte-rial amobarbital procedures. In: Engel Jr., J., Pedley, T.A. (Eds.),Epilepsy:aComprehensiveTextbook.,2nded.Lippincott Williams&Wilkins,NewYork,p.1833.

Kriegeskorte, N., Goebel, R.,Bandettini, P., 2006. Information-basedfunctionalbrainmapping.Proc.Natl.Acad.Sci. U.S.A. 103,3863—3868.

Kriegeskorte,N.,Simmons,W.K.,Bellgowan,P.S.,Baker,C.I.,2009. Circularanalysisinsystemsneuroscience:thedangersofdouble dipping.Nat.Neurosci.12,535—540.

LaConte,S.,Strother,S.,Cherkassky,V.,Anderson,J.,Hu,X.,2005. Support vector machines for temporal classification of block designfMRIdata.Neuroimage26,317—329.

Loring,D.W.,Meador,K.J.,Lee,G.P.,King,D.W.,1992.Amobarbital EffectsandLateralizedBrainFunction:TheWadaTest.Springer Verlag,NewYork.

Loring,D.W.,2005.ShouldwestickwiththeWada?Probing prac-ticablepreferences—–commentaryonBaxendaleetal.Epilepsia 49,722—723.

McKenna, P., Warrington, E.K., 2007. The Graded Naming Test. CambridgeCognitionLimited,Cambridge.

Mechanic-Hamilton,D.,Korczykowski,M.,Yushkevich,P.A.,Lawler, K.,Pluta,J.,Glynn,S.,Tracy,J.I.,Wolf,R.L.,Sperling,M.R.,

French, J.A., Detre, J.A., 2009. Hippocampal volumetry and functional MRIofmemoryintemporallobeepilepsy.Epilepsy Behav.16,128—138.

Norman,K.A.,Polyn,S.M.,Detre,G.J.,Haxby,J.V.,2006.Beyond mind-reading:multi-voxelpatternanalysisoffMRIdata.Trends Cogn.Sci.10,424—430.

Powell, H.W., Richardson, M.P., Symms, M.R., Boulby, P.A., Thompson, P.J., Duncan, J.S., Koepp, M.J., 2007. Reorga-nization of verbal and nonverbal memory in temporal lobe epilepsyduetounilateralhippocampalsclerosis.Epilepsia48, 1512—1525.

Powell,H.W.,Richardson,M.P.,Symms,M.R.,Boulby,P.A., Thomp-son, P.J.,Duncan, J.S.,Koepp,M.J., 2008.Preoperative fMRI predicts memory decline following anterior temporal lobe resection.J.Neurol.Neurosurg.Psychiatry79,686—693. Rausch,R.,2002. Epilepsysurgerywithin thetemporallobe and

its short-term and long-termeffects onmemory. Curr. Opin. Neurobiol.15,185—189.

Richardson,M.P.,Strange,B.A.,Duncan, J.S.,Dolan, R.J.,2003. Preserved verbal memory function in left medial temporal pathology involvesreorganisation of function to right medial temporallobe.Neuroimage20(Suppl.1),S112—S119.

Richardson,M.P.,Strange,B.A.,Thompson,P.J.,Baxendale,S.A., Duncan, J.S.,Dolan,R.J.,2004.Pre-operativeverbalmemory fMRI predicts post-operative memory decline after temporal loberesection.Brain127,2419—2426.

Rissman, J., Wagner, A.D., 2012. Distributed representations in memory: insights from functional brain imaging. Annu. Rev. Psychol.63,101—128.

Salmenpera,T.M.,Duncan,J.S.,2005.Imaginginepilepsy.J. Neu-rol.Neurosurg.Psychiatry76(Suppl.3),iii2—iii10.

Scoville,W.B.,Milner,B.,1957.Lossofrecentmemoryafter bilat-eralhippocampallesions.J.Neurol.Neurosurg.Psychiatry20, 11—21.

Wada,J.,1949.Anewmethodforthedeterminationofthesideof cerebralspeechdominance.Apreliminaryreportonthe intra-carotidinjectionofsodiumamytalinman.IgakutoSeibutsugaki 14,221—222.

Warrington,E.K.,1984.ManualforRecognitionMemoryTest. NFER-Nelson,Windsor,England.

Warrington, E.K., 1996. The Camden Memory Tests. Psychology Press,Hove.

Warrington,E.K.,James,M., 1991.TheVisual Objectand Space Perception Battery(VOSP). Thames Valley TestCo., Bury St. Edmunds,England.

Wechsler,D.,1997.WechslerAdultIntelligenceScaleIII.The Psy-chologicalCorporation,SanAntonio,TX.

Weigl,E.,1927.ZurPsychologiesogenannterAbstrakionsprozesse. Z.Psychol.103,2—45[translatedinJ.Abnorm.Soc.Psychol. 36,3—33].

Weiskopf,N.,Sitaram,R.,Josephs,O.,Veit,R.,Scharnowski,F., Goebel,R.,Birbaumer,N.,Deichmann,R.,Mathiak,K.,2007. Real-timefunctionalmagneticresonanceimaging:methodsand applications.Magn.Reson.Imaging25,989—1003.

Wiebe,S.,Blume,W.T.,Girvin,J.P.,Eliasziw,M.,2001.A random-ized,controlledtrialofsurgeryfortemporal-lobeepilepsy.N. Engl.J.Med.345,311—318.

Yushkevich,P.A.,Piven,J.,Hazlett,H.C.,Smith,R.G.,Ho,S.,Gee, J.C.,Gerig,G.,2006.User-guided3Dactivecontour segmenta-tionofanatomicalstructures:significantlyimprovedefficiency andreliability.Neuroimage31,1116—1128.