EEG Regression Model for BCI Cursor ControlEEG headset was used to measure brainwave activity...
Transcript of EEG Regression Model for BCI Cursor ControlEEG headset was used to measure brainwave activity...
Background:Brain-ComputerInterface(BCI)systemshavebecomeasourceofgreatinterestintherecentyears.Establishingalinkwiththebrainwillleadtomanypossibilitiesinthehealthcare,robotics,orentertainmentfields.Electroencephalography(EEG)isonepopularnoninvasivetechniquetoestablishaBCI.Inthismethod,electrodesareplacedalongthescalptorecordtheelectricpotentialscreatedfromtheneuronsfiredinthebrain.Throughaparadigmknownas“imaginedbodykinematics,”asubjectcancontrolacomputercursorwithoutmakinganyovertmovements.ThismethodofusingimaginedbodykinematicsfromEEGhasshowntodecreasetrainingtimefromdaystominutescomparedtoothermethods.Thisplatformcouldleadtoimprovingneurorehabilitationprogramsaswellasmoreadvancedprostheticdevicesthatcanbecontrolledbythebrain.Itcanalsobebeneficialtoparalyzedpatientsorthosewithotherneurodegenerativediseasesthatinhibitsmusclemovementsincenoovertmovementsarerequiredtocontrolthecursor.
Objective:ThegoalofthisprojectwastoimprovethepredictionaccuracyforapreviouslydevelopedBCImodelthatusedlinearregressiontopredictcursorvelocityfromasubject’sthoughtsbytestingnewmethodsandnonlinearmodels.
Figure1:RawEEGandEmotiv EPOCheadset
Thetrainingphaseconsistsof5horizontaland5verticaltrialseachlasting1minute.Thesubjectwasinstructedtousetheparadigmofimaginedbodykinematicstotrackthemotionofanautomatedcursorusingtheirdominanthandasiftheywereusingacomputermousewhilemakingnoovertmovements(Figure2).
Figure2:Outlineoftrainingtaskusingonedimensionalmovement
Datafrom33subjectsweretestedusingdifferentcombinationsofthemostimportantchannelsfoundforvelocityprediction(Table1).ItwasdeterminedthatchannelcombinationF7,FC5,T8,FC6,F4,andF8providedthebestaccuracyforpredictinghorizontalvelocity,whileallthechannelswerebestforverticalvelocity.UsingonlyF7andF8achievedacceptableaccuracygivingthepotentialforamorecovenantrealworldapplicationwithasmallerheadset.Futureworkwillincludetestingthesefeaturesinnewmachinelearningmodelstoimprovethepredictionscoresfromthelinearregressionmodel(Figure5).
Table1:Predictionaccuracyusingdifferentchannelcombinations
Figure5:Actualandpredictedvelocitiesfortwotrialshorizontal(top)andtwotrialsvertical(bottom)
• TheNationalScienceFoundation• TheJointInstituteofComputationalSciences• RezaAbiri andSoheil Borhani• Dr.Kwai Wong• XSEDEandSanDiegoSupercomputerCenter
EEGRegressionModelforBCICursorControl
Introduction
Students:JustinKilmarx,DavidSaffo,LucienNgMentor:Xiaopeng Zhao
Acknowledgements
TrainingProtocol
ModelDesign:Eachchannelwastestedindividuallyinourlinearregressionmodeltodeterminethemostimportantchannelsforvelocityprediction(Figure4).
Figure4:Heatmapofpredictionaccuracyforeachchannelduring5horizontaltrials
Cross-Validation:Themodelsweretestedwithtrialwisecrossvalidationwhere4trialswereusedfortrainingand1wasusedfortesting.Thiswasrepeatedforall5combinationsofhorizontalandverticaltrials.Thiswastoensurethemodelwiththebestaccuracyofvelocitypredictionwaschosenastheonetobeusedduringtheonlinecontrol.
Serialvs.Parallel:Aswearegeneratinganewmodelforeveryindividualsubjectandtrialwehavemanycomputationsthatareindependentofeachotherthatcanbedonefasterwithparallelprocessing.Doingthisalsoallowsustotestdifferentmodelsandhyper-parametersatoncereturningresultsfasterthanrunningeverythinginserial.WeusedDask Distributedtosetupournetworkwithaschedulerandworkernodes. ThenetworkwassetuponCometusing4nodes,96workers,4coresperworker.Timesareshownbelow.
ResultsAnEmotiv EPOC14-channelwirelessEEGheadsetwasusedtomeasurebrainwaveactivity(Figure3)whileBCI2000programrecordeddatasuchastheEEGactivityandcursorposition.AllofflineprocessingwasdoneusingPythonprogrammingontheCometsupercomputerinSanDiego.13previouspointsofEEGdatafrommemorywasusedasfeaturestotrainthemodels. Figure3:Channellocations
MaterialsandMethods
Serial AdaBoost 04:30:00Parallel AdaBoost 00:04:29
Features Horizontal Accuracy VerticalAccuracy
AllChannels 70.77% 44.67%
F7, 02,P8,T8,FC6,F4,F8,AF4 71.03% 41.68%
F7andF8 69.93% 25.64
F7,FC5,T8, FC6,F4,F8 72.73% 36.98%