Gibson, R.M., Chennu, S., Owen, A.M., Cruse, D. Complexity and familiarity enhance single-trial ...

12 Pages·2013·2.22 MB·English

by  Raechelle M. Gibson

ClinicalNeurophysiologyxxx(2013)xxx–xxx
ContentslistsavailableatScienceDirect
Clinical Neurophysiology
journal homepage: www.elsevier.com/locate/clinph
Complexity and familiarity enhance single-trial detectability
of imagined movements with electroencephalography
Raechelle M. Gibsona,⇑, Srivas Chennub, Adrian M. Owena, Damian Crusea
aTheBrainandMindInstitute,DepartmentofPsychology,WesternUniversity,London,Ont.,Canada
bDepartmentofClinicalNeurosciences,UniversityofCambridge,Cambridge,UK
a r t i c l e i n f o h i g h l i g h t s
Articlehistory:
(cid:2) Weusemachine-learningtoclassifyEEGduringmotorimageryinsamplesofathletes,musicians,and
Accepted23November2013
age-matchedcontrols.
Availableonlinexxxx
(cid:2) Imageryofcomplexactionsandimageryoffamiliaractionscanresultinmorerobustbrainresponses
insomecases.
Keywords:
Motorimagery (cid:2) Ourfindingsmaybeappliedtoimprovebrain-computerinterfacesintendedforusebybehaviourally
Sensorimotorrhythm non-responsivepatients.
Movementcomplexity
Electroencephalography
a b s t r a c t
Vegetativestate
Objective: Wesoughttodeterminewhetherthesensorimotorrhythms(SMR)elicitedduringmotorimag-
ery(MI)ofcomplexandfamiliaractionscouldbemorereliablydetectedwithelectroencephalography
(EEG),andsubsequentlyclassifiedonasingle-trialbasis,thanthoseelicitedduringrelativelysimpler
imaginedactions.
Methods: Groupsofhealthyvolunteers,includingexperiencedpianistsandicehockeyplayers,performed
MI of varying complexity and familiarity. Their electroencephalograms were recorded and compared
usingbrain-computerinterface(BCI)approachesandspectralanalyses.
Results: RelativetosimpleMI,significantlymoreparticipantsproducedclassifiableSMRforcomplexMI.
DuringMIofperformanceofacomplexmusicalpiece,theEEGoftheexperiencedpianistswasclassified
significantlymoreaccuratelythanduringMIofperformanceofasimplermusicalpiece.Theaccuracyof
EEGclassificationwasalsosignificantlymoresustainedduringcomplexMI.
Conclusion: MIofcomplexactionsresultsinEEGresponsesthataremorereliablyclassifiedformoreindi-
vidualsthanMIofrelativelysimpleractions,andfamiliaritywithactionsenhancestheseresponsesin
somecases.
Significance: The accuracy of SMR-based BCIs in non-communicative patients may be improved by
employingfamiliarandcomplexactions.IncreasedsensitivitytoMImayalsoimprovediagnosticaccu-
racyforseverelybrain-injuredpatientsinavegetativestate.
(cid:2)2013InternationalFederationofClinicalNeurophysiology.PublishedbyElsevierIrelandLtd.Allrights
reserved.
1.Introduction emergingfromcoma,VSpatientsretaincyclesofeyeopeningand
closingsimilartothesleep-wakecyclesoffullyawakeandaware
Patientswithdisordersofconsciousness(DOC)arebehaviour- individuals (Multi-Society Task Force on PVS, 1994a,b; Royal
allycharacterizedbyvaryinglevelsofarousalandawarenessmea- CollegeofPhysiciansWorkingGroup,1996;cf.Cruseetal.,2013).
sured primarily by their ability to exhibit reliable responses to Critically,despiteproducingspontaneousmovements,VSpatients
externalstimulation(Jennett,2002;Bernat,2006;Owen,2008).Of areunabletoexhibitanypurposefuloutwardresponsestoverbal
the various conditions includedin the DOC (e.g., coma, the mini- commands, and are thereby diagnosed as ‘unaware’ (Jennett,
mallyconsciousstate(MCS),etc.),thevegetativestate(VS)isone 2002;Owen,2008).ManyVSpatientshavediffusebraininjurythat
ofthemostpoorlyunderstood(Jennett,2002;Owen,2008).After mayincludeinsulttotheperipheralmotorsystem;thesecircum-
stances could lead to an inaccurate diagnosis of VS in a patient
⇑ whoretainsawarenessandcognitivefunction,butlackstheability
Correspondingauthor.Tel.:[phone]x84672;fax:[phone].
E-mailaddress:[email protected](R.M.Gibson). torespondpurposefullyinabehaviouralassessment(Owen,2008).
1388-2457/$36.00(cid:2)2013InternationalFederationofClinicalNeurophysiology.PublishedbyElsevierIrelandLtd.Allrightsreserved.
http://dx.doi.org/10.1016/j.clinph.2013.11.034
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
2 R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx
Infact,researchershavereportedthatsomepatientswhoarediag- impairments,includingtetraplegiaandadvancedamyotrophiclat-
nosedasVScanfollow(e.g.,Owenetal.,2006;Montietal.,2010; eral sclerosis (ALS), have been successfully trained to control
Bardin et al., 2011; Cruse et al., 2011; Goldfine et al., 2011; Naci SMR-based BCIs (Pfurtscheller et al., 2000; Kübler et al., 2005).
and Owen, 2013), or attempt to follow (e.g., Bekinschtein et al., The SMR approach to BCI is therefore a viableoptionforpatients
2011; Cruse et al., 2012), commands by modulating their brain who have been immobile for an extended period, including those
activity,despitebeingunabletofollowcommandswiththeirexter- withadisorderofconsciousness.
nalbehaviour.Thesefindingsraisethepossibilitythatassistivede- DespitethepotentialbenefitsofbedsideEEG-basedBCIsforpa-
vices known as brain-computer interfaces (BCIs) could improve tientsdiagnosedasVSandtheirfamilies,thereissubstantialintra-
diagnostic accuracy in this group by detecting ‘covert’ signs of and inter-subject variability in BCI performance (Wolpaw et al.,
awareness,aswellasby potentiallyofferingthepatientameans 2002;Pfurtschelleretal.,2006;Nacietal.,2012;Grosse-Wentrup
ofcommunication(e.g.,Montietal.,2010;Luléetal.,2013). and Schölkopf, 2013). In many studies of healthy volunteers and
BCIsaredevicesthatcanallowaperson(the‘user’)tooperatea peoplewithseveremotorimpairments,someindividualsaresim-
computerwithoutproducingamotoroutput.Usingmachine-learn- plyunabletoreliablyregulatethebrainsignalsnecessarytooper-
ing techniques, subject-specific patterns of brain activity can be ate a BCI without training (e.g., Guger et al., 2003; Wolpaw and
learnedbyacomputerandsubsequentlyclassifiedintoapredefined McFarland,2004;Cruseetal.,2011;Hammeretal.,2012).Inthe
communicativeoutput.Forexample,thecomputermayoutputthe current work, we propose modifications to the traditional SMR-
response‘‘yes’’whentheuserproducesbrainactivitypatternA,and basedBCIdesignthatmayoptimizeBCIperformanceforbehavio-
outputtheresponse‘‘no’’whentheuserproducesbrainactivitypat- urally non-responsive patients in particular. These modifications
ternB(e.g.,MasonandBirch,2003;Sorgeretal.,2009;Luléetal., applytothenatureofthetaskusedtogeneratetheSMRandthe
2013;Nacietal.,2013).Thecomputeralgorithmmustbetrained nature of the comparisons made during signal classification (see
on a series of trials in which the desired output from the user is alsoCurranandStokes,2003;Curranetal.,2004).
known (the ‘training phase’ of machine-learning classification), InpublishedSMR-basedBCIresearchtodate,usersaretypically
andthentestedontrialsinwhichthedesiredoutputfromtheuser instructedtoimaginemovingtheirhands,feet,ortonguetogener-
isnotknown(the‘testingphase’ofmachine-learningclassification) ate an SMR (e.g., Neuper and Pfurtscheller, 2001; Kübler et al.,
basedonpredefinedfeaturesofthedata(e.g.,powerinagivenfre- 2005; Cruse et al., 2011). With only a few exceptions, users are
quencybandoftheelectroencephalogram,EEG).Fromthetesting askedtoimagineverysimpleactions,suchasrepeatedlysqueezing
phaseofclassification,onecanobtainanaccuracyvaluebasedon oneoftheirhandsintoafist.However,actionsthataremorecom-
thenumberofsuccessfullyidentifiedbrainresponsesand,byexten- plexcouldresultinamorerobustandconsistentSMR(Curranand
sion,correctlyexecutedcommunicativeoutputsfromtheBCI.Cru- Stokes, 2003; Curran et al., 2004). Indeed, there is evidence that
cially, from a clinical perspective, when classification accuracy is compleximagined actions are associated with more robust brain
significantlyabovechance,theindividualisdemonstrablycapable responses than simpler imagined actions. For instance, there is
ofproducingconsistentandappropriatepatternsofbrainactivity convergingevidencefromfunctionalmagneticresonanceimaging
inresponsetocommands,thusprovidingameanstoidentifycovert (fMRI), functional near-infrared spectroscopy (fNIRS), and trans-
command-followingintheabsenceofabehaviouralresponse(Cruse cranial magnetic stimulation (TMS) studies that complex motor
etal.,2011;Owen,2013).Sinceclassificationmustbebothaccurate imageryisassociatedwithgreaterhemodynamicchangeandhigh-
and reliable for successful communication and other BCI output eramplitudemotor-evokedpotentialsthansimplemotorimagery
functions, such as computer mouse cursor control, classification (e.g., Kuhtz-Buschbecketal.,2003;RoosinkandZijdewind,2010;
accuracyandtasksensitivityaretwoofthemostimportantmea- Holper and Wolf, 2011). Similar to previous work, we define
surementsofanyBCI. ‘complex’ motor imagery in this paper as tasks that involve both
AparticularEEGsignalcalledthesensorimotor-rhythm(SMR)is sequences of movements and more than one body part
apracticaloptionforBCIsintendedforusebyVSpatients(Chatelle (e.g., Kuhtz-Buschbecketal.,2003;RoosinkandZijdewind,2010;
etal.,2012;Nacietal.,2012;Grosse-WentrupandSchölkopf,2013). HolperandWolf,2011).Wemustalsoclarifythatourcomplexity
Usingasfewasfoursurfaceelectrodesplacedontheheadoverthe manipulationsinthisworkalwaysinvolve‘‘common’’complexac-
sensorimotorcorticalareas(sitesCP3,CP4,FC3,andFC4fromthe tionsequences;that is, we chose actions that participants would
modified international 10–20 system; Sharbrough et al., 1991), have previously encountered through overt practice (e.g., Studies
one can acquire the SMR as a person kinesthetically imagines 2and3)orcommonknowledge(e.g.,clappingasinStudy1).We
moving a body part. Power decreases known as event-related selected common action sequences to ensure that participants
desynchronizations (ERDs) and power increases known as event- coulddrawfromproceduralmemoryorsemanticknowledgeinor-
related synchronizations (ERSs) in the mu (7–13Hz) and beta dertoimagineeachaction.Thesesortsofknowncomplexactions
(13–30Hz)frequencybandsaretypicallyusedasthesignalfeatures would therefore have lower cognitive demands than novel, com-
for classificationwith SMR-based BCIs (Pfurtscheller and Neuper, plex action sequences that would need to be learned at the time
1997;NeuperandPfurtscheller,2001;Neuperetal.,2009).Unlike ofassessment(e.g.,tappingthefingersinarandomsequencede-
otherEEG-basedBCIparadigms(e.g.,theP300spellerdescribedin finedbytheexperimenterasinpreviouswork;Kuhtz-Buschbeck
Farwell and Donchin (1988)), the imagination tasks used with et al., 2003; Roosink and Zijdewind, 2010; Holper and Wolf,
SMR-basedBCIsimposelowsensorydemandsontheuser.Further- 2011). We hypothesized that more complexactionswould result
more,ofparticularimportanceforpatientsdiagnosedasVSwho,by inmorerobustSMRsand,consequently,higherclassificationaccu-
definition,areunabletofixatetheireyes,SMRBCIsneednotinvolve racythantraditionalSMR-basedBCIimagerytasks.
visualstimulation(Chatelleetal.,2012;Nacietal.,2012;Grosse- Additionally,ithasbeenproposedinpreviousworkthatasking
WentrupandSchölkopf,2013).Finally,itisimportanttoacknowl- userstoimagineactionswhichtheyarefamiliarwithcouldimprove
edgethatchangesinthecorticalmotorsystemfollowingprolonged SMRclassification(CurranandStokes,2003;Curranetal.,2004).In
immobility may prevent some behaviourally non-responsive thispaper,wechosetoexploretheroleofactionfamiliarityinmod-
patients from producing reliable SMRs. Nevertheless, there is ulationoftheSMRbydrawingfromsamplesofexperiencedathletes
evidence that individuals diagnosed with disorders of conscious- andmusicians,giventhattheeffectsoflong-termmotorlearning
ness,includingVSandMCS,canproduceSMRsinmotortasks,even havebeenstudiedextensivelyinthesegroupsalready(seeMünte
afterseveralyearsofimmobility(Goldfineetal.,2011;Cruseetal., etal.(2002),andNakataetal.(2010),forreviews).Whileimagining
2011). Furthermore, patients with chronic and extensive motor actions involving the sport or instrument of their expertise,
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx 3
experiencedathletesandmusiciansproducemorefocusedpatterns classification accuracy and result in more users (and, in future
ofbrainactivation(e.g.,Lotzeetal.,2003;Miltonetal.,2007;Olsson work,patientsdiagnosedasVS)withSMRsthatcouldbedetected
etal.,2008;WeiandLuo,2010)andreportmoreobjectivelyaccurate reliably. This questionwas addressed in three studies of healthy,
imagerythannovices(e.g.,Louisetal.(2012);seealsoRieger(2012)). young adults with an experimental set-up suitable for future
Basedonthelatterfindings,itwasexpectedthatfamiliarimagery clinical work with VS and other non-communicative patients. In
wouldresultinamorereliableSMR,andthushigherclassification Study 1 (Complexity), participants imagined simple hand actions
accuracy,thantraditionalSMR-basedBCIimagery.Ifsupported,this (squeezes) ofthe sorttypicallyusedwithSMR-basedBCIs along-
hypothesiscouldbeextendedtofutureworkwithbrain-injuredpa- side other more complex bimanual actions not commonly used
tientsbyselectingimagerytasksbasedontheskillsandhobbiesthat withBCIs.Itwaspredictedthat,inaccordancewithpriorevidence
thepatienthadpriortotheirinjury.Furthermore,basedonthefind- ofincreasedbrainactivityduringcomplexmotorimagery(Kuhtz-
ingsreviewedpreviouslyregardingtheinfluenceofactioncomplex- Buschbecketal.,2003;RoosinkandZijdewind,2010;Holperand
ityonbrainresponsestomotorimagery,itwasalsohypothesized Wolf,2011),classificationaccuracy(versusrest)wouldbehigher
thatactionsthatwerebothfamiliarandcomplexwouldfurtheren- when the participant was imagining complex actions than when
hance the SMR and increase classification accuracy – thereby imagining relatively simpler actions. In Study 2 (Familiarity),
improvingtheabilitytodetectcovertawarenessinfutureworkwith groups of experienced pianists, experienced ice hockey players,
behaviourallynon-responsivepatients. andage-matchedcontrolswereinstructedtoimaginecompleting
Anotherconvention in SMR-basedBCI research to date is that handsqueezesandactionsfromhockeyandpiano.Itwaspredicted
comparisonsarealmostexclusivelymadebetweentheSMRsgen- that classification accuracy between rest and imagery would be
eratedforvariousimaginedmovements(e.g.,lefthandversusright highest for the athletes and musicians in the action with which
hand, etc.; Neuper and Pfurtscheller, 2001; Guger et al., 2003; theyweremostfamiliar(i.e.,pianistsimaginingplayingthepiano,
Kübler et al., 2005; Pfurtscheller et al., 2006; Cruse et al., 2011). and hockey players imagining playing hockey; see Lotze et al.,
Althoughthesetypesofcomparisonsrenderacceptableclassifica- 2003; Fourkas et al., 2008; Olsson et al., 2008; Wei and Luo,
tionaccuracyinmosthealthypeople(e.g.,Gugeretal.,2003;Bai 2010). In Study 3 (Complexity and Familiarity), the experienced
etal.,2008),asdiscussedinCruseetal.(2012),thesetypesofcom- pianistsfromStudy2imaginedplayingonesimplepieceofmusic
parisonsmaynotbeappropriateforVSpatients.Indeed,aspartof and one relatively more complex piece of music on the piano. It
thestandardbehaviouralassessmenttoolforVSpatients,theComa was expected that classification accuracy would be highest for
RecoveryScale-Revised(CRS-R;KalmarandGiacino,2005),aware- thecomplexpieceversusrestcomparison.Implicationsandmod-
nessisassessedbyapatient’sabilitytoproduceonetypeofaction ificationsforfutureBCIworkwithVSpatientsarediscussed.
followingacommand.If,onthreeoutoffouroccasions,thepatient
is able to successfully follow that command, the patient is diag-
2.Methods
nosed as, at least, minimally conscious (Kalmar and Giacino,
2005;seealsoCruseetal.,2012).Inthiswork,wethereforeper-
2.1.Ethicsstatement
formedcomparisonsthatare moresimilarto behaviouralassess-
ments of awareness than previous SMR-based BCI paradigms.
Allparticipantsgaveinformedwrittenconsent.ThePsychology
Specifically, contrasts were made between one imagined action
Research Ethics Board of Western University (London, ON, CAN)
ofinterestand periodsofrest(mind-wandering), ratherthanbe-
providedethicalapprovalforthestudies.
tweentwodistinctimaginedactions.Inadditiontoprovidingapo-
tentialBCI-basedassessmentofawarenessthatismoresimilarto
thestandardbehaviouralmethodthanpreviousBCIs,wepropose 2.2.Participantsandstimuli
thatthistechniqueismorepracticalforbehaviourallynon-respon-
sivepatientsbecausemaintainingmorethanoneimaginedaction 2.2.1.Study1(Complexity)
inworkingmemorymayimposeexcessivecognitivedemandson Sixteenhealthy,right-handedyoungadultsparticipatedinthe
somepatients. complexitystudy(fivemen;agerange=17–20years;medianage
Asafinalconsideration,itisimportanttonotethatbehavioural of18years).Forthesimpleimaginedmovementphaseofthecom-
assessmentsofawarenesssuchastheCRS-R(KalmarandGiacino, plexitystudy(Study1),theparticipantswereinstructedtoimagine
2005)arebasedonthereliabilityofapatient’sabilitytofollowcom- repeatedlysqueezingtheirlefthand,righthand,orbothhandsfol-
mands.Patientsareevaluatedacrossmultipleassessmentsessions lowingtheauditorycuesof‘‘left’’,‘‘right’’,and‘‘both’’,respectively.
andmultipleattemptstofollowthesamecommandineachsession ForthecompleximaginedmovementsphaseofStudy1,thepartic-
inordertoensurethatanaccurateassessmentofawarenessisob- ipantswereinstructed toimagineeitherplayingtheguitar,clap-
tained. If neuroimaging-based methods are to be used in clinical ping their hands, or juggling using both hands. These tasks were
assessments of awareness, these methods should maximize the cued with the words ‘‘guitar’’, ‘‘clap’’, and ‘‘juggle’’, respectively.
likelihoodofdetectingareliablecommand-followingresponseto Ineachtaskphase,participantswerealsoaskedtoceasetheprevi-
reducetheriskofmisdiagnosingawareness.SMR-basedBCIsnatu- ously-cued mental imagery and mind-wander following the cue
rallyprovideanadditionalmeasureofreliabilitythatisnotauto- ‘‘relax’’.Theorderofthesimpleandcompleximaginedmovement
matically available with standard behavioural assessments, i.e., phases were counter-balanced across participants. All auditory
thestatisticalsignificanceoftheclassificationofthebrainresponse. instructionstimuliwere1sinlength.
Asdescribedabove,whentwobrainstatesproducedbyapatient
canbedifferentiatedwithstatisticalsignificance,itmaybeinferred 2.2.2.Study2(Familiarity)
that the patient possesses a covert ability to follow commands. Forty-eighthealthy,right-handedyoungadultsparticipatedin
Thus,increasingtheaccuracyofBCIclassificationnotonlyincreases the familiarity study. Sixteen participants were experienced ice
theaccuracyofpotentialcommunicationdevices,butalsoincreases hockeyplayers(sevenmen;agerange=18–29years;medianage
thenumberofpatientsinwhomcovertcommand-followingmay of 20years); sixteen participants were experienced pianists (six
bedetectedwhenitispresent. men; age range=18–29years; median age of 20years); and six-
The research question in this work was whether having users teen participants had either limitedor no experienceplaying the
perform motor imagery tasks involving more complex and piano or hockey (eight men; age range=18–28years; median
familiar movements than previous investigations would improve ageof18years).Allhockeyplayershadplayedregular,competitive
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
4 R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx
icehockeyforatleast10years,andallpianistshadformalmusical wereabletorecallbothpiecesfrommemory.Finally,thepianists
trainingandhadplayedandpracticedpianoregularlyforatleast were also instructed to mind-wander following the auditory cue
10years.Therewerenosignificantdifferencesinmeanageoffirst of ‘‘relax’’, and all instructions in Study 3 were 1s in length. The
play experience, mean years of total play experience, or mean Study 3 procedure was always conducted following the Study 2
self-reported hours of regularplay perweek betweenthegroups procedureinordertopreventpianistsfromselectingtheparticular
ofathletesandmusicians,pairwiseps>.51(Bonferronicorrection; musicalpiecesfromStudy3forthepianoactioninStudy2.
see Table 1). As shown in Table 1, the three groups also did not
differsignificantlyinmeanage,sex,handedness(Oldfield,1971), 2.3.Procedure
or imagery ability (Gregg et al., 2010; ps>.34). All participants
wereinstructedtoimaginemakingaslapshot(abimanualaction BeforetheEEGrecordingsession,eachparticipantcompleteda
from hockey), playing a musical piece on the piano using both seriesofshortquestionnaires.AllparticipantscompletedtheEdin-
hands,orsqueezingtheirrighthandintoafistfollowingtheaudi- burgh Handedness Inventory (Oldfield, 1971) and the Movement
torycuesof‘‘hockey’’,‘‘piano’’,and‘‘righthand’’,respectively.Asin Imagery Questionnaire-Revised Second Version (Gregg et al.,
Study1,participantswereaskedtomind-wanderfollowingthecue 2010).ParticipantsinStudies2and3alsocompletedaquestion-
‘‘relax’’,andallinstructionswere1sinlength. naireregardingtheirexperiencesplayinghockey,piano,andother
sportsandinstruments.AttheconclusionofStudies2and3,par-
2.2.3.Study3(ComplexityandFamiliarity) ticipants rated the vividness of their imagined actions using a 5-
Theexperiencedpianists(n=16)fromStudy2completedStudy pointLikertscale(seeTable1).
3inthesameEEGrecordingsession.InStudy3,thepianistswere Allauditorycueswerepre-recordedbyonefemalespeakerand
instructed to imagine playing ascending and descending C-major presentedtotheparticipantusingER-1insertearphones(Etymo¯tic
scales and B-major arpeggios over two octaves using both hands ResearchInc.,ElkGroveVillage,IL).Eachtrialbeganwithanaudi-
following theauditory cues of ‘‘scale’’ and ‘‘arpeggio’’. Thepieces torycueandwasfollowedby5–8sofsilencebeforetheonsetof
ofmusicwereselectedbasedonthecurriculumoftheRoyalCon- thenextauditorycue.The durationofthe silentintervalwas se-
servatoryofMusic(RCM),whichisaprominentmusiceducation lectedrandomlyfromauniformdistributiononeachtrial.Studies
institution in Canada. In the RCM curriculum, piano students are 1and2werecompletedinfourblocksof48trials(12trialsofeach
evaluatedonscalesandthekeyofC-majorfromthe1stgradelevel, instructionperblock);Study3wascompletedinthreeblocksof48
arpeggiosfromthe4thgradelevel,andthekeyofB-majorfromthe trials (16 trials of each instruction per block) as there were only
7thgradelevel(RoyalConservatoryofMusic,2008).Giventhedif- three(ratherthanfour)trialtypesinthelattertask;andeachblock
ferentgradelevelsatwhichtheC-majorscaleandB-majorarpeg- of 48 trials was approximately 6min in duration. All trials were
gioareevaluatedintheRCMcurriculum,onecanconcludethatthe presentedinapseudorandomordersothatnomorethantwocues
B-majorarpeggiorepresentsamoredifficult,i.e., complex,action ofthesametypewerepresentedconsecutively,andthefirsttrialof
thantheC-majorscale.Itisalsoimportanttonotethatallthepia- eachblockwasalwaysanimaginedactiontrial(ratherthana‘re-
nistsinthisstudyreportedhighfamiliaritywithbothpiecesand lax’ trial). Participants were provided with short breaks between
Table1
SummaryofparticipantdemographicsandexperiencemeasuresfromStudies2and3.
Variable Pianists Hockeyplayers Controls Total
M±SD M±SD M±SD M±SD
Demographics
n 16 16 16 48
Sex(#male) 6 7 8 21
Age(years) 20.6±2.9 20.9±3.4 19.4±2.7 20.3±3.0
LQ 70.9±14.7 70.9±21.9 67.8±20.7 69.9±19.0
Hockeyexperiencea
Initialage(years) 1.6±3.0 5.0±1.7⁄,a 2.6±4.2 3.1±3.4
Totalyears 0.9±1.8 15.4±3.4⁄⁄⁄ 0.4±0.9 5.6±7.4
Hoursperweek 0.5±2.0 9.2±4.5⁄⁄⁄ 0.8±2.0 3.5±5.1
Numberofothersportsplayed 1.5±1.2 2.9±1.5⁄ 2.4±1.6 2.3±1.5
Pianoexperienceb
Initialage(years) 5.4±2.0(cid:2),b 1.6±3.4 3.7±6.7 3.6±4.7
Totalyears 14.8±3.6⁄⁄⁄ 0.5±1.2 0.6±1.4 5.3±7.1
Hoursperweek 8.9±3.8⁄⁄⁄ 0.3±0.7 0.6±1.3 3.2±4.7
Numberofotherinstrumentsplayed 1.6±1.0⁄⁄⁄ 0.3±0.5 0.7±0.6 0.9±0.9
MIQ-RSScores
Kinesthetic 5.2±1.0 5.1±1.1 4.8±1.3 5.0±1.1
Visual 5.8±1.0 5.6±1.4 5.4±1.0 5.6±1.1
Self-reportratingsofimagery
Vividnessc 3.9±1.0d 3.7±0.6 3.8±0.7 3.8±0.8
4.2±1.1e
Note: ⁄⁄⁄p<.001; ⁄p<.05; (cid:2)p<.1. M=mean; SD=standard deviation; LQ=laterality quotient handedness measure (Oldfield, 1971); MIQ-RS=Movement imagination
questionnaire-revisedsecondversion(Greggetal.,2010).
a Fourpianistsreportedsomeleisureexperienceplayinghockey(2–5yearstotalexperienceinitiatedatages5–8),andfourcontrolsreportedsomeleisureexperience
playinghockey(1–3yearstotalexperienceinitiatedatages5–13).
b Threehockeyplayersreportedsomeleisureexperienceplayingpiano(1–4yearstotalexperienceinitiatedatages7–10),andfivecontrolsreportedsomeleisure
experienceplayingpiano(0.5–5yearstotalexperienceinitiatedatages5–22).
c Responsestothefollowingquestion:‘‘Pleaseratetheoverallvividnessofyourimaginedactionsduringthetask,suchthat:1=notatallvivid,2=slightlyvivid,
3=somewhatvivid,4=moderatelyvivid,5=veryvivid’’.
d ResponsesfromthepianistsfortheStudy2imagery.
e ResponsesfromthepianistsfortheStudy3imagery.
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx 5
blocksinordertoreducefatigue.Participantswerealsoinstructed theseanalyses,thelogbandpowervaluesoffourfrequencybands
toimaginecompletingeachactionrepeatedlyfromtheoffsetofthe atelectrodesC30andC40weretheclassificationfeatures.Basedon
auditory cue to the onset of the next auditory cue in order to previouswork(Cruseetal.,2011,2012),thefrequencybandswere
account for potential differences in the duration of the imagined 7–13Hz(mu),13–19Hz(low-beta),19–25Hz(mid-beta),and25–
actions. To reduce ocular artefacts, participants were instructed 30Hz (high-beta), for a total of eight features per classification
tokeeptheireyesclosedthroughoutthetasks. analysis (two electrodes(cid:3)four frequency bands). For the single-
trialanalyses,thespectralpowerineachbandwasestimatedwith
2.4.Electrophysiologicaldataacquisitionandpre-processing aslidingHammingwindow(1-s;asrecommendedbyPfurtscheller
andLopesdaSilva,1999)movingin50msstepsusingashort-time
Inallthreestudies,EEGwasrecordedusingtheg.Gammaactive Fouriertransform(MATLABfunction‘spectrogram’).
electrode system (g.tec Medical Engineering GmbH, Austria). In Classificationofeachimaginedaction(e.g.,squeezingtheright
Study 1, EEG was recorded with a four-channel montage housed hand in Study 1, etc.) and the corresponding rest condition was
inanelectrodecap; theelectrodes wereplacedatsitesCP3,FC3, performed using a naïve Bayes classifier (MATLAB’s ‘naivebayes’
CP4, and FC4 (Sharbrough et al., 1991). In Studies 2 and 3, EEG object). Each classification analysis was conducted using ten-fold
was recorded from the same four scalp sites as in Study 1, and cross-validation. For the cross-validation procedure, each partici-
additionalelectrodeswereplacedatsitesTP7,FT7,CPz,FCz,TP8, pant’strialsforonetypeofimaginedactionandtherestcondition
andFT8(Sharbroughetal.,1991).ThereportedanalysesforStud- fromthesameexperimentwereseparatedintotenapproximately
ies 2 and 3 consist of data from only the four electrodes used in equalgroups.ThenaïveBayesclassifierwastrainedonthefeatures
Study1,asthismontagehasbeenpreviouslyshowntoprovidero- ofnineofthesegroups(‘training’),andthentheclassofeachtrial
bust SMR recordings (Guger et al., 2003; Cruse et al., 2012). The in the tenthgroup was predicted in orderto calculatethe classi-
EEG signals were acquired using a g.USBamp amplifier operating fier’s accuracy (‘testing’). Specifically, during training, the naïve
throughaUSB2.0port.Stimulipresentationandphysiologicaldata Bayesclassifierestimatedtheparametersofaprobabilitydistribu-
recordingswereperformedusingaSimulink(cid:3)modelinMATLAB(cid:3) tionpertrainingfeatureperclass;theparameterswerethemean
(The Mathworks, Inc., Natick, MA). Simulink is often utilized in andstandarddeviationofanormaldistribution, thetrainingfea-
BCIapplicationsbecauseitensurestheprecisesynchronizationof tureswerebandpowerperfrequencybandateachelectrode,and
EEGactivityandcueonset/offset(Gugeretal.,2001).Inallthree thetwoclassesweretherestandimagerytrialtypes.UsingBayes’
studies,bipolarsurfaceelectromyographic(EMG)recordingswere Theoremduringtesting,thefeaturesofthetesttrialswereusedto
obtained from both forearms on the ventral surface (placed over calculatetheposterior probabilities for eachclass, and then each
theflexordigitorumprofundus)inordertodetectovertmovements. testtrialwasplacedintheclasswiththehighestposteriorproba-
FollowingpreviousresearchregardingBCIs(e.g.,Gugeretal.,2003; bility (for more information regarding naïve Bayes classification,
Kübleretal.,2005;Pfurtscheller etal.,2006),wedidnotacquire see Jiang et al., 2007). The classification procedure was repeated
electrooculogramstomeasureeyemovementsduringtherecord- tentimessothateachtrialservedasatesttrialinexactlyoneof
ing. Online, the EEG data were filtered from 0.5 to 60Hz with a the tencross-validation folds. Theaverage classificationaccuracy
60Hz notch filter using an infinite impulse response (IIR) digital across the ten folds was then calculated at each time-point, and
Butterworthfiltersetusingtheg.USBampgraphicaluserinterface thetime-courseofthecross-validatedclassificationaccuracywas
(GUI).TheEMGdatawerefilteredfrom5to250Hzwitha60Hz smoothedwithasliding-windowof500mstocontrolforoutliers
notchfilterusingtheg.USBampGUI.TheEEGrecordingswereref- (cf.Cruseetal.,2012).
erencedtotherightearlobewithaforehead(Fpz)ground,andthe To determine the statistical significance of the classification
right elbow (olecranon) was used for the EMG ground. The EMG accuracy,apermutationtestwith1000repetitionsthatcontrolled
andEEGdataweresampledat600Hzwithimpedanceskeptbelow for familywiseerrorwas used (Maris,2004; see also Cruse et al.,
5kXfortheEEGrecording. 2012). For each permutation, the class labels of imagery or rest
Offline,theEEGdataweredown-sampledto100Hz,filteredbe- wererandomlyshuffledacrosstrials,andthecross-validatedclas-
tween0.5and40HzusingtheEEGLABfunction‘pop_eegfilt’,and sificationproceduredescribedpreviouslywasrepeated.Themaxi-
segmentedinto6-sepochstime-lockedtotheonsetoftheauditory mumsmoothedaccuraciesacrossalltime-pointsfromeachofthe
cue. The EEGLAB filter function consisted of a two-step least- 1000repetitionswereusedtoformadistributionrepresentingthe
squares finite impulse response (FIR) filter; in the first step, data expected classification results if the classifier were operating at
werefilteredwithahigh-passcut-offof0.5Hz,andinthesecond chance(thenullhypothesis).Theclassificationaccuracyobtained
step,datawerefilteredwithalow-passcut-offof40Hz.TheEMG fortheparticipant’soriginaldata(i.e.,thedatawiththecorrecttrial
datawererectifiedandthenfilteredwitha10Hzhighpassfilter labels)wasthenevaluatedagainstthisdistributiontocalculatea
usingthesameEEGLABleast-squaresFIRfilterfunctiondescribed familywiseerror-correctedsignificancevaluefortheoriginalclas-
previously. Trials containing physiological artefacts, including sificationresultsateachtime-point.Finally,tocontrolforthemul-
overthandmovementsasevidentfromtheEMG,wereidentified tiplecomparisonsofbandpower(i.e.,onecomparisonforeachtime
by visual inspection and removed. After artefact rejection, the point of imagery versus rest), a control of False Discovery Rate
mediannumberoftrialsincludedineachimageryandrestcondi- (FDR) approach was used (implemented via MATLAB’s ‘fdr’ func-
tionperparticipantwas:Study1–40(range:29–48);Study2–43 tion;BenjaminiandHochberg,1995;Verhoevenetal.,2005).The
(range: 27–48); and Study 3–43 (range: 28–48). Finally, the EEG controlofFDRapproachisknowntoreducetheriskofTypeIerror
data were re-referenced offline to form two bipolar channels without requiring as stringent reductions in power as Bonferroni
(FC3–CP3, FC4–CP4) that are subsequently identified as C30 and procedures(seeVerhoevenetal.(2005)).
C40,respectively;thisbipolarapproachisknowntodetectchanges
in mu and beta power with high accuracy across many people 2.6.EEGspectralanalyses
(Cruseetal.,2012).
Inadditiontothesingle-trialclassificationanalysesofthedata,
2.5.EEGsingle-trialclassificationprocedure theEEGdatafromallthreestudieswereanalyzedusingthesame
spectralanalysisprocedurereportedinCruseetal.(2012).Foreach
Amachine-learningalgorithmwasusedforsingle-trialclassifi- time-pointatC30andC40,spectralpowerestimateswerecalculated
cationoftheEEGdataasdescribedinCruseetal.(2011,2012).For usingaHanningwindow(1-s)time–frequencytransformationvia
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
6 R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx
the‘ft_freqstatistics’functionoftheopen-sourceMATLABtoolbox, the trial numbers were rank-transformed to meet the statistical
FieldTrip (Oostenveld et al., 2011). The time–frequency data at assumptions of the ANOVA). Additionally, a Kruskal–Wallis test
bothelectrodeswerethencomparedbetweentheimaginedmove- wasperformedtocomparetheself-reportedimageryvividnessrat-
mentsandrestusingcluster-basedpermutationtesting(cf.Maris ingsbetweenthegroups.Giventhelackofsignificantdifferences
andOostenveld,2007;Cruseetal.,2012)implementedviaField- in classification accuracy in Study 2, no comparisons were made
Trip. For the cluster-based testing, the time–frequency data for a forthenumberofsignificanttimepointsintheimageryversusrest
given imagery condition and rest (or another imagery condition; comparisons.
seeSection2.7foradescriptionofthecomparisonsforeachstudy) Thegroupspectralanalyseswereconductedwiththetime–fre-
werelog-transformedandthencomparedateachdatapointbya quencydataaveragedacrossalltrialsineachconditionperpartic-
paired-samples t-test. All significant data points (p<.025) were ipant. For Study 1, the time–frequency data for each participant
then arranged into groups, i.e., clusters, based on their temporal wereaveragedacrossalloftheimaginedactionsineachcomplex-
andspectralproximitytoeachother,andthesumofthetvalues itylevel,andthecluster-basedpermutationtestingwasconducted
was calculated for each cluster. To determine the familywise er- between each complexity level and rest, and between the two
ror-corrected significance valuefor each summed t-value cluster, complexity levels (Fig. 2). For Study 2, comparisons were made
a Monte Carlo randomization test that controlled for familywise separatelyforeachfamiliaritygroupbetweeneachimagerycondi-
error was used. In the randomization test, the condition labels tionandrest(Fig.3).Finally,forStudy3,comparisonsweremade
wererandomlypermutedtoremovetask-relateddifferences,and betweeneachofthetwoimageryconditionsandrest,andbetween
theclusteringprocedurewasrepeated1000times.Themaximum the two imagery conditions (Fig. 4). Although we were primarily
summedt-valueclustersfromeachrepetitionwereusedtoforma concerned withtheimageryversusrestcomparisonsbecausewe
distribution, and this distribution was then used to test the null positthatthesecomparisonsaremorepracticalforfutureBCIwork
hypothesis that the original summed t-value cluster (i.e., the with brain-injured patients, we conducted post hoc comparisons
summedt-valueclustercomputedfromthedatawiththecorrect between the imagery conditions in the spectral analyses of the
triallabels)occurredbychance. Studies1and3datatoattempttoidentifytheneurophysiological
correlatesofthecomplexityeffectobservedinthesingle-trialanal-
2.7.Group-levelstatisticalanalyses ysesofStudy3.Wedidnotmakecomparisonsbetweenthefamil-
iarandcontrolimageryconditionsinStudy2becausetherewasno
Forthesingle-trialanalysesoftheEEGdata(seeTable2,Fig.1, evidence of an effect of action familiarity in the single-trial
andSupplementaryTablesS1–S3),allgroup-levelstatisticalanaly- analyses.
ses were conducted using IBM SPSS Statistics version 21.0, and
whereapplicable,theDunn–Sidakcorrectionwasusedforthefol-
low-uptests.ForthespectralanalysesoftheEEGdata(seeFigs.2– 3.Results
4),allstatisticalanalyseswereconductedusingthecluster-based
permutation testing described previously (cf. Maris and Oosten- 3.1.Study1(Complexity)
veld, 2007; Cruse et al., 2012) via custom MATLAB script using
theopen-sourcetoolbox,FieldTrip(Oostenveldetal.,2011). Intermsofthesingle-trialanalysesoftheEEGdata,therewasa
Forthegroupcomparisonsofsingle-trialanalysesoftheStudy1 trendforclassificationaccuracy(imageryversusrest)tobehigher
andStudy3data,severalparametricandnon-parametricrepeated- forthecompleximaginedactions,t(15)=(cid:4)1.963,p=.068,d=0.49
measures statistical tests were used. Two paired-samples t-tests (Simple:M=60.68%,SE=0.74%;Complex:M=62.74%,SE=0.93%).
were used to compare maximum classification accuracy and the Therewasnosignificantdifferenceforthetimeatwhichthemax-
timeatwhichmaximumclassificationaccuracyoccurredrelative imumclassificationaccuracyoccurredbetweenthetwocomplex-
totheonsetoftheinstructionin bothStudies1and 3.Wilcoxon ity conditions, p=0.29. From the familywise permutation tests,
Signed Rank Tests were used to compare the number of time- thereweresignificantlymoretime-pointsatwhichsignificantclas-
points for which a significant classification was obtained in both sificationswereobtainedinthecomplexconditionthaninthesim-
studies, and this test was also used to compare the self-reported plecondition,Z=(cid:4)2.197,exactp=.026,r=.55(Simple:medianof
vividness ratings of the imagined actions from the pianists be- 0significanttime-points[range:0–26];Complex:medianof14.5
tweenStudy2andStudy3.Anexact(ratherthanasymptotic)cal- significanttime-points[range:0–31]).Itisalsoworthnotingthat
culation of the p-value was used with the test of time-points in therewasnosignificantdifferenceinthenumberoftrialsineach
Study1toaccountforthepositiveskewofthecountdata(given complexitylevel,p=.29.
thatmanyparticipantshadzerosignificanttime-points,especially Therewassomevariabilitybetweenandwithinsubjectsforthe
in the simple complexity condition).Finally, the numberof trials single-trial analyses of the Study 1 EEG data. While at least one
includedineachcomplexityconditionofeachstudywascompared simpleimaginedactiontypewasclassifiedsignificantlyfromrest
usingtheFriedmantest. for only four of the 16 participants (25%), at least one complex
Tofurtherillustratethedifferenceintermsofthesignificanceof imaginedactiontypewasclassifiedsignificantlyfromrestforsig-
thesingle-trialanalysesbetweenthecomplexitylevelsinStudies1 nificantly more participants (11% of 16%, or 69%), Fisher’s exact
and 3, participants were assigned to a binary category based on p=.032(two-tailed).Oftheelevenparticipantswhoproducedsig-
whetherornotatleastoneimaginedmovementineachcomplex- nificant responses for the complex imagery versus rest compari-
ityconditionwasclassifiedsignificantlyfromrestforthepartici- sons, three participants produced significant responses for all
pant (0=no significant classifications). The number of threeofthecomplexactions;fiveparticipantsproducedsignificant
participants with at least one significant classification and the responsesfortwoofthethreecomplexactions;andthreepartici-
number of participants with no significant classifications in each pantsproducedasignificantresponseforonlyoneofthecomplex
complexitylevelwerethencomparedusingFisher’sExactTest. actions.Inthesimpleimagerycomparisons,twoofthefourpartic-
For Study 2, 3 (Group: Pianist, Hockey, Control)(cid:3)3 (Action: ipants produced significant responses for all three of the simple
PlayPiano,Slap-shot,Squeeze)mixedanalysesofvariance(ANO- actions;oneparticipantproducedsignificantresponsesfortwoof
VAs)wereusedtocomparetheaveragedmaximumclassification the simple actions; and one participant produced a significant
accuracies, the time at which the maximum accuracy occurred, response for only one of the simple actions. Furthermore, there
and the total number of trials included in each condition (N.B., were some participants in the sample who did not produce any
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx 7
Table2
Summaryofsingle-trialclassificationanalysesforStudies1–3foreachoftheimagerycomparisonsandgroups(whereapplicable).
Study1(Complexity;n=16)
Simpleimagery
RighthandimageryversusRest LefthandimageryversusRest BothhandsimageryversusRest
Acc Time SigTPs SigSub Acc Time SigTPs SigSub Acc Time SigTPs SigSub
M 60.61 3.59 3.63 2 60.15 3.22 1.75 1 61.27 2.85 6.25 4
SE 0.97 0.29 2.50 – 0.70 0.37 0.94 – 0.93 0.33 3.46 –
Compleximagery
JuggleimageryversusRest GuitarimageryversusRest ClapimageryversusRest
Acc Time SigTPs SigSub Acc Time SigTPs SigSub Acc Time SigTPs SigSub
M 62.71 3.79 9.31 8 63.83 3.42 15.06 7 61.67 3.44 6.06 7
SE 1.07 0.24 3.67 – 1.58 0.32 6.19 – 1.23 0.33 2.63 –
Study2(Familiarity;n=48)
PianoimageryversusRest HockeyimageryversusRest HandsqueezeimageryversusRest
Acc Time SigTPs SigSub Acc Time SigTPs SigSub Acc Time SigTPs SigSub
Pianists(n=16)
M 63.31 3.76 16.31 7 62.74 3.27 9.88 8 60.76 3.45 6.19 7
SE 1.49 0.24 5.99 – 1.38 0.35 4.69 – 1.05 0.33 2.96 –
Hockeyplayers(n=16)
M 59.23 3.25 3.88 4 60.34 3.31 5.50 4 59.52 3.35 4.69 4
SE 0.98 0.31 1.95 – 1.64 0.24 2.95 – 1.02 0.28 2.80 –
Controls(n=16)
M 63.12 3.35 12.94 8 63.75 3.18 11.69 9 63.69 3.25 12.06 8
SE 1.49 0.27 5.70 – 1.38 0.26 4.82 – 1.37 0.33 4.53 –
Study3(ComplexityandFamiliarity;n=16)
Simple(Scale)imageryversusRest Complex(Arpeggio)imageryversusRest
Acc Time SigTPs SigSub Acc Time SigTPs SigSub
M 66.34 3.61 23.19 9 69.60 3.51 33.69 13
SE 1.96 0.19 7.18 – 2.03 0.28 7.10 –
Note:Acc=maximumcross-validatedclassificationaccuracy(%);Time=timeofAcc(secondsfollowingonsetofauditorycue);SigTPs=numberoftimepointsforwhich
significantclassificationresultswereobtained;SigSub=numberofparticipantsforwhichsignificantclassificationswereobtained.
Fig.1. Meansmoothed,cross-validatedclassificationaccuracyfromtheEEGsingle-trialanalysesacrosstimeforonesubject(anexperiencedguitarplayer)fromStudy1by
imageryversusrestcomparison.Timeismeasuredrelativetotheoffsetoftheauditoryinstructioncue.Shadedregionsdepict±1standarderrorofthemean(smoothed),and
starsdenotetime-pointswithstatisticallysignificantclassificationresultsforthecorrespondingaccuracytimecourse.
significantbrainresponsesforanyoftheimaginedactions(fourof low-beta band in the complex imagery versus rest comparison
the sixteen participants, or 25% of the sample). Only one partici- (ps<.014). In the simple imagery versus rest comparison, there
pantinthesampleproducedsignificantbrainresponsesforevery wasanERDovertherighthemisphereinthemid-betabandthat
imaginedactioninthesimpleandcomplextaskphases.Theinter- approached statistical significance (p=.050). Additionally, there
and intra-subject variability in classification accuracy is summa- werenosignificantclustersinthesimpleimageryversuscomplex
rized in Table 2 and detailed in the supplementary data tables imagerycomparisons(ps>.10).
(SupplementaryTableS1). Even though the role of action familiarity was not explicitly
FromthegroupspectralanalysesoftheEEGdata(Fig.2),there examined in Study 1, there was one interesting finding in this
werestatisticallysignificantERDsoverthelefthemisphereinthe experimentthatemphasizedtheimportanceofthisfactorandits
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
8 R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx
Fig.2. Averaged,group(n=16)time–frequencyplotsfromthespectralanalysesoftheEEGdataforStudy1(Complexity)averagedacrossthethreeimaginedactionsineach
imagerycondition.Therangeofpowervalues(logratiodifference)thatareplottedis±0.6121.Significantclusters(ps<.014)areoutlinedwithsolidlines;dashedlines
highlightaclusterwithp=.050.Plotsontheleftandrightreflecttheleft-andright-hemisphereEEGchannels(C30andC40,respectively),asindicated.Frequency(Hz)is
indicatedontheverticalaxis,andtimeismeasuredrelativetotheoffsetoftheinstruction.
Fig.3. Averaged,grouptime–frequencyplotsfromthespectralanalysesoftheEEGdataforStudy2(Familiarity)byimageryversusrestcomparisonperfamiliaritygroup
(n=16pergroup).Therangeofpowervalues(logratiodifference)thatareplottedis±0.6121.Significantclusters(ps<.021)areoutlinedwithsolidlines;dashedlines
highlightaclusterwithp=.036(C30forthesimpleimageryminusrestcomparisonforthehockeyplayers).Plotsontheleftandrightineachpairreflecttheleft-andright-
hemisphereEEGchannels(C30 andC40,respectively),asindicated.Frequency(Hz)isindicatedontheverticalaxis,andtimeismeasuredrelativetotheoffsetofthe
instruction.
potentialinfluenceonsingle-subjectperformance.InFig.1,wede- didnotproducesignificantlyclassifiableSMRsforanyofthesim-
pictthetime-courseofthesingle-trialclassificationaccuraciesfor pleimaginedactions;theaveraged,maximumclassificationaccu-
oneoftwoexperiencedguitaristswhoparticipatedinStudy1.In racyobtainedfortheseparticipantsonthesimpleimaginedactions
line with the group trends already reported, these participants was57.58%(SE=1.01%),andtheclassificationresultsforthesim-
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx 9
Fig.4. Averaged,group(n=16)time–frequencyplotsfromthespectralanalysesoftheEEGdatafromStudy3(ComplexityandFamiliarity)bycomparison.Therangeof
powervalues(logratiodifference)thatareplottedis±0.6121.Significantclusters(p<.017)areoutlinedwithsolidlines;dashedlineshighlightclusterswithp=.044(C30)
andp=.038(C40).Plotsontheleftandrightreflecttheleft-andright-hemisphereEEGchannels(C30andC40,respectively),asindicated.Frequency(Hz)isindicatedonthe
verticalaxis,andtimeismeasuredrelativetotheoffsetoftheinstruction.
pleimageryversusrestcomparisonswerenotstatisticallysignifi- FromthegroupspectralanalysesoftheEEGdata(Fig.3),there
cantatanytime-pointforeitherparticipant.Bothparticipantspro- weresignificantERDsforthefamiliarimageryversusrestcompar-
duced SMRs for the instruction to imagine clapping (one of the isonsforboththeexperiencedpianists(ps<0.015)andtheexperi-
actionsinthecompleximagerycondition)thatweresignificantly enced hockey players (ps<0.019). Although the ERDs were
classifiablefromrestforashorttime;significantclassificationre- significant bilaterally (rather than unilaterally) and over a longer
sults occurred for 6–7 time-points, and the smoothed maximum period of time for the pianists, the significant ERDs were similar
classification accuracy for each participant was 64.79% betweenthehockeyplayersandpianistsforthefamiliarimagery,
(SE=3.52%)and 64.47%(SE=4.04%;showninFig.1). Mostinter- inthatbothERDsfeaturedsignificantclustersinboththelow-mu
estingly, however, both participants produced a markedly robust (8–10Hz)andlow-betabands.However,thepianistsalsohadsig-
SMRfortheinstructiontoimagineplayingtheguitarthatwassig- nificantERDsforthehockeyimagery(ps<.017)andsimpleimag-
nificantlyclassifiablefromrestformostoftheepoch(68–69time- ery (ps<0.017) versus rest comparisons, and the hockey players
points) with very high accuracy (maximumaccuracies of 81.10%, had an ERD that approached statistical significance (p=.036) in
SE=3.44%, as shown in Fig. 1 and 7128%, SE=3.97%, no figure thesimpleimageryversusrestcomparison.Furthermore,thecon-
provided). trol group (which consisted of age-matched individuals without
significantexperienceplayinghockeyorpiano)alsoproducedsig-
nificant ERDs for both the piano (p=.021) and hockey imagery
3.2.Study2(Familiarity)
(p=.002)versusrestcomparisons,thoughtheseERDclusterswere
smallerintheirspatiotemporalextentthanthosegeneratedbypia-
Intermsofthesingle-trialanalysesoftheEEGdata,therewereno
nists.Therewerenosignificantclustersinanyoftheothercompar-
significant differences in accuracyfor any of the imaginedaction
isons (ps>0.08). Thus, much like the single-trial analyses, the
versusrestcomparisons,orforanygrouponanyoftheimaginedac-
spectral analyses do not provide strong evidence that there was
tion versus rest comparisons, ps>.44 (refer to Supplementary
anadvantageintermsofSMRdetectionforanygroupregardless
TableS2foranoverviewofclassificationaccuracybyparticipant).
oftheirfamiliarity(orlackthereof)withtheimaginedactions.
The main effect of group on accuracy approached significance
(p=.054),andthiswasdrivenbytherelativelylowoverallclassifi-
cation accuracy of the hockey players (M=59.70%, SE=1.03%) 3.3.Study3(ComplexityandFamiliarity)
compared to the control group (M=63.52%, SE=1.24%; pairwise
p=.054).Therewerenosignificantdifferencesintermsofthetime In terms of classification accuracy, there was a significant
at which maximum classification accuracy occurred for any advantage for the complex imagined action (Complex:
imagined action or any group by imagined action type, ps>.64, M=69.60%, SE=2.03%) compared to the simple imagined action
andtherewasalsonosignificantdifferenceinthenumberoftrials for the pianists-only study (Simple: M=66.34%, SE=1.96%;
includedinanyoftheimaginedactiontypesorrestconditionson t(15)=(cid:4)2.589, p=.021, d=0.65). Supplementary Table S3 pro-
averageorbygroup,ps>.41.Thethreegroupsalsodidnotdiffer vides an overview of classification accuracy by participant. Fur-
in their self-reported vividness ratings of the imagined actions, thermore, there was an advantage for the complex imagined
p=.34, and in motorimageryability asmeasuredbythe MIQ-RS actioncomparedtothesimpleimaginedactioninthatsignificantly
(ps>.545;Greggetal.,2010;seeTable1). more time-points were classified significantly from rest in the
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034
10 R.M.Gibsonetal./ClinicalNeurophysiologyxxx(2013)xxx–xxx
formercomparison,Z=(cid:4)2.510,p=.009(two-tailed),r=.63(Sim- fourparticipantsproducedaSMRthatwassignificantlyclassifiable
ple:medianof9.0significanttime-points[range:0–76];Complex: fromrestforeachofthesimpleimaginedactions.Wepositthat,at
medianof32.0significanttime-points[range:0–74]).Therewasno the group-level, the advantage of the complex imagery was not
significantdifferencebetweenthenumberofparticipantswithsig- statistically significant for classification accuracy because of this
nificant imagery versus rest comparisons in the two complexity variability.Inotherwords,therewasnoadvantageforthecomplex
levels, Fisher’s exact p=0.25. It is important to note that there imageryatthegroup-levelbecausemostparticipants(eightofthe
wasnosignificantdifferenceintermsofwhenthemaximumclas- eleven participants with significant complex imagery versus rest
sification accuracy occurred relative to the onset of the auditory comparisons) only produced significant responses for one or two
cueforthetwoimaginedactions,p=.72,anditisalsoworthnot- ofthecompleximaginedactions,ratherthanforallthreeofthese
ing that the differences between the complexity conditions were actions. The latter observation is well-illustrated anecdotally by
notdrivenbyadifferenceinthenumberoftrialsinanycondition, the classification results of two guitarists who participated in
p=.16, or by a difference between the self-reported vividness of Study1(seeSection3.1andFig.1).Therewasamarkedimprove-
theimaginedactionsinStudy3comparedtoStudy2,p=.10. mentinclassificationaccuracywhentheguitaristsimaginedplay-
TheresultsofthegroupspectralanalysesareshowninFig.4for ing the guitar as part of the complex imagery phase of the
eachcomparison.Comparedwithrest,significantERDs(ps<.017) experimental procedure, but neither guitarist produced a signifi-
occurred in the low-mu and low-beta bands over the left hemi- cantly classifiable brain response for any of the simple imagined
sphereandin thelow-,mid-,andhigh-betabandsovertheright actions.
hemisphere beginning approximately one second after the offset WedesignedStudies2and3tofollowuponthegrouptrendof
oftheinstructiontoimagineplayingthesimplepiece.Forthecom- the advantage for the complex actions in Study 1 and to explore
plex imagery versus rest comparison, a similar response was ob- the role of action familiarity in the imagery paradigm. Study 2,
served with the same time-course, although the ERDs were which included groups of experienced pianists, experienced ice
significant (ps<.011) over both hemispheres throughout the hockeyplayers,andage-matchedcontrols,didnotmeetourexpec-
low-muband,andthroughoutthelow-andmid-betabands.Ina tations in that there was no advantage for any of the imagined
comparison of the two imagery conditions, there was more of a actionsforanygroup,regardlessoftheirfamiliaritywiththeimag-
desynchronization bilaterally for the complex imagery in the inedactions(e.g.,pianistsimaginingplayingpiano,etc.).Compared
low-mu and low-beta bands that approached statistical signifi- to novices, experienced athletes and musicians utilize fewer re-
cance(ps=.044[C30]and.038[C40]). gions of the brain when imagining actions involving the sport or
instrument with which both groups have familiarity (e.g., Lotze
et al., 2003; Fourkas et al., 2008; Olsson et al., 2008; Wei and
4.Discussion Luo, 2010). Although expert brain responses to familiar imagery
arefairlyconsistentwithinandacrossindividuals(e.g.,Langheim
In this paper, we presented a series of three experiments in etal.,2002),theseresponsesdidnotresultinanenhancementof
which movement complexity and familiarity with movements theSMRinStudy2.
weremanipulatedinaSMR-basedimageryparadigm.Thepurpose Interestingly,inStudy3,wefoundanadvantageinSMRclassi-
ofthisworkwastoincreasethelikelihoodofdetectingreliableand ficationforspecific,familiaractions(performanceoftwomusical
robustSMRs,andtherebyimprovehowwellcovertcommand-fol- pieces),suchthatclassificationaccuracywashigherandsignificant
lowingcanbedetectedinbehaviourallynon-responsivepatients. over a longer periodof time when the experiencedpianists from
In future work, we will apply these manipulations to SMR para- Study2imaginedplayingthemorecomplexmusicalpiece.More-
digms used with BCIs intended for communication with severely over,theERDsfromthespectralanalysesofStudy3werestatisti-
brain-injuredpatients,includingindividualsdiagnosedasVS. callysignificantforthelongestperiodoftimeandassociatedwith
InStudy1,imageryofarangeofbimanualsequencesofactions thelargestlog-ratiodifferencesinpowerofalltheimagerycondi-
(‘‘complex imagery’’) resulted in SMRs that were classified from tions with significant ERDs across all three experiments (see
restwithsimilaraccuracyasimageryofthesimplehandsqueezes Figs. 2–4). As anexploratory posthoc test, we alsocompared the
typicallyusedwithSMR-basedBCIs.Therewasagrouptrendthat classificationaccuracyforthepianistsontheinstructiontoimag-
thecomplexactionsusedinthistaskwereclassifiedwithhigher ine playing a musical piece of their choice on the piano (versus
accuracy than the simple hand squeezes typically used with rest)inStudy2withtheclassificationaccuracyforimaginedper-
SMR-basedBCIs,althoughthisresultdidnotreachstatisticalsig- formance of each of the two specific musical pieces (versus rest)
nificance(p=.068).Furthermore,wefoundasignificantadvantage in Study 3. This analysis resulted in a significant effect of move-
for the complex actions in that the SMRs for these actions were menttype, F(15)=16.016,p=.001, g 2=.361,that wasdrivenby
p
classified significantly from rest for a longer period of time than thesignificantlylowerclassificationaccuracyforthepianoimag-
the simple actions. We also found that significantly more single ery in Study 2 compared to the complex piano imagery in Study
subjects produced significantly classifiable SMRs for at least one 3 (p=.003; Complex Musical Piece [Study 3]: M=69.60%,
ofthecomplexactionsthanforatleastoneofthesimpleactions. SE=2.03%; Musical Piece of Choice [Study 2]: M=63.31%,
Overall, the findings of Study 1 align with our prediction based SE=1.49%; other pairwise ps>.06). This finding provides addi-
onpreviousworkthattherewouldbeanenhancementofthebrain tionalsupportforourclaimthatitistheimageryofaspecific,com-
response for the complex imagery (e.g., Kuhtz-Buschbeck et al., plex, and familiar action that leads to an advantage in SMR
2003;RoosinkandZijdewind,2010;HolperandWolf,2011),and classification,ratheranyofthesethreepropertiesindividually.
support our hypothesis that motor imagery that involves more Acombinationoffactorslikelycontributedtothefindingthat
thanonebodypart andsequencesofactionsislikelytoimprove the specific, complex, familiar imagery from Study 3 resulted in
theabilitytodetectcovertcommand-followingorcommunication the most robust brain responses in this work. The finding that
infuturework. thecompleximageryresultedinanenhancementoftheSMRcom-
AnotherinterestingfindingfromStudy1wasthebetween-and paredtothesimpleimagerywithinStudy3alignswellwithprevi-
within-subjectvariabilityforthevariousimaginedactions.Nearly ous work (e.g., Kuhtz-Buschbeck et al., 2003; Roosink and
halfofthesample(sevenoreightofthesixteenparticipants;see Zijdewind, 2010; Holper and Wolf, 2011). Moreover, there was
Table 2) produced a SMR that was significantly classifiable from morepotentialvariabilityinthebrainresponsesbetweenpartici-
rest for each of the compleximagined actions, while only one to pants in the piano imagery for Study 2 versus Study 3 due to
Pleasecitethisarticleinpressas:GibsonRMetal.Complexityandfamiliarityenhancesingle-trialdetectabilityofimaginedmovementswithelectroen-
cephalography.ClinNeurophysiol(2013),http://dx.doi.org/10.1016/j.clinph.2013.11.034