EEG Channel Selection Using Decision Tree in Brain-Computer Interface

Transcription

1 EEG Channel Selecton Usng Decson ree n Bran-Computer Interface Mahnaz Arvaneh * Cunta Guan Ka Keng Ang and Hok Cha Quek * * School of Computer Engneerng Nanyang echnologcal Unversty Sngapore Insttute for Infocomm Research Agency for Scence echnology and Research (A*SAR) Sngapore E-mal: stuma@r.a-star.edu.sg Abstract Selectng approprate channels n Bran-Computer Interface (BCI) applcatons helps to mprove the usablty and the performance of the BCI as some channels are contamnated by nose or contan rrelevant nformaton. hs paper proposes a method of usng decson trees to select approprate channels n EEG-based BCI applcatons. he proposed method selects the best subset of approprate channels by consderng the correlaton nformaton between them usng Decson ree. he performance of the proposed method s compared wth several other methods of channel selecton such as Fsher Crteron Mutual Informaton Support Vector Machne and Common Spatal Pattern coeffcents. he performances of these methods are evaluated n terms of usng publcly avalable BCI Competton IV dataset IIa. Expermental results show that the proposed method outperforms the exstng channel selecton methods specfcally n the case where the number of selected channels s relatvely small. I. INRODUCION A bran-computer nterface (BCI) measures analyzes and decodes bran sgnals drectly to provde a non-muscular means of controllng a devce. hus BCIs enable users wth severe motor dsabltes to use ther bran sgnals for communcaton and control [1] []. here are manly two types of BCIs nvasve and nonnvasve. Electroencephalogram (EEG) s commonly used n nonnvasve BCIs because t s the least expensve compared wth other methods of bran sgnal acquston equpments. However EEG sgnal processng s a challengng problem due to the poor resoluton of EEG and ts mult-channel nature n the acquston of bran sgnals [3]. he use of too few channels may result n nsuffcent nformaton whereas too many channels may nclude nosy and redundant channels that degrade BCI performance. One method to mprove the performance of EEG-based BCI s to use approprate channels on the scalp. hs s because f nosy and redundant channels are excluded computatonal complexty s decreased whle accuracy of the BCI may be ncreased [4]. Moreover the use of a large number of channels s not practcal because t nvolves a longer EEG setup tme. Snce approprate channels may dffer from subect to subect a method of fndng subect-specfc optmal number of approprate channels plays an mportant role n the performance of BCI applcatons. he problem of EEG channel selecton can be consdered as a feature selecton problem. Channel selecton methods n the lterature are manly characterzed as wrapper or flter approaches. In wrapper approaches feature selecton s coupled wth classfcaton algorthm such as the Support Vector Machne (SVM) classfer [4]. In flter approaches feature selecton s ndependent of nducton algorthms. One example s to select features based on certan crtera such as the Mutual Informaton (MI) between channels and class labels [5]. he performances of wrapper-based methods depend on the accuracy of the appled classfer and propertes of the features comng from channels. Although some methods have been proposed to avod retranng classfers [6] wrapper-based feature selecton methods generally nvolved ntensve computatons. In contrast flter-based feature selecton methods are computatonally less ntensve than wrapper approaches but may not select an optmal subset of features [6]. Another EEG-specfc approach uses the Common Spatal Pattern () coeffcents [7]. he algorthm s shown to be effectve n dscrmnatng two classes of EEG measurements n BCI applcatons [8]. However s senstve to outlers and EEG sgnals are generally nosy from varous artfacts. hus channel selecton usng coeffcents may not select an optmal subset of approprate channels. hs paper proposes a method of usng decson trees [9] to select approprate subset of channels for EEG-based BCI applcatons. In the proposed algorthm ntally a mult band sgnal decomposton flter s presented to reduce nose by dentfyng the subect-specfc frequency range and then the most dscrmnatve subset of features s selected by the defned decson tree classfer. Fnally the selected features are ranked accordng to a tree prunng method. Snce the decson tree selects a feature accordng to the results of prevous chosen features selected features would be more relevant and less correlated to each other. he remander of ths paper s organzed as follows: Secton II revews the exstng EEG channel selecton methods based on Fsher Crtera (FC) MI SVM and coeffcents. Secton III explans the proposed method. Secton IV descrbes the experment performed on the publcly avalable BCI Competton IV dataset IIa. Secton V presents the expermental results by comparng the proposed method wth exstng methods. Fnally secton VI concludes ths paper APSIPA. All rghts reserved. 5 Proceedngs of the Second APSIPA Annual Summt and Conference pages 5 30 Bopols Sngapore December 010.

2 II. REVIEW ON CHANNEL SELECION MEHODS he goal of channel selecton s to remove rrelevant or correlated channels n order to mprove the performance of the BCI system. he followng revews exstng channel selecton methods used n BCI applcatons n the lterature: A. Fsher Crtera (FC) he FC determnes how strongly a feature s correlated wth the labels [4] whereby the score R of feature s defned as ( ( X ) ( X )) R( X ) V ( X ) V ( X ) where X + and X - denote the set of trals n two dfferent classes; µ and V respectvely denote the mean and varance of feature. he rank of a channel s smply set to the mean score of the correspondng features. B. Mutual Informaton (MI) In ths method the features that have maxmum MI wth the class labels are ranked as the best features. he MI between nput features X and the class Y={1...N y } s defned as I( X; Y) H( Y) H( Y X ) where N y s the number of classes and H denotes the entropy functon [5]. Entropy s a measure of uncertanty assocated wth a random varable. Gven a data = { 1 d } the entropy of the random varable s defned as H ( ) p( )log p( ) where p() s the probablty functon. he condtonal entropy of two random varables X and Y s defned as H( Y X ) p( x y)log p( y x). xx yy In ths study the Parzen Wndow [10] s used to estmate p(y x). he Mutual Informaton based channel selecton algorthm s descrbed as follows: Step 1: Intalze a set of d features F { f1 f... fd} Step : Compute the MI of features wth the output class I( f ; Y) 1... d f F. Step 3: Select the best features that maxmze ( f ; Y) C. Support Vector Machne (SVM) I he SVM s a classfcaton technque [4] whch performs effcently n a number of real-world problems. he SVM d separates the tranng data X R wth two classes y={-11} d by fndng a hyperplane wth a weght vector w R and an offset b R as shown n (5). A good separaton s acheved by a hyperplane that has the largest dstance to the nearest tranng data ponts of any class. hs dstance s called the functonal margn. (1) () (3) (4) H : R d { 11} x sgn( w. x b) In SVM-based channel selecton the channels are selected usng a Recursve Feature Elmnaton (RFE) method. he RFE method was proposed by Guyon et al. [6] based on the concept of margn maxmzaton. he RFE algorthm begns wth the subset comprsng all the features and elmnates one feature at a tme from the subset. In each teraton the learnng machne f (n our case SVM classfer) s traned on the current subset of features by optmzng a cost functon J to maxmze the margnal functon. For each remanng feature X the change n J resultng from the removal of X s estmated wthout retranng the f. hereafter the feature X v(k) that results n mprovng or least degradng J s removed. hs algorthm s terated tll only the specfed number of features remans. Guyon et al. have presented n [6] that under some condtons the removal of one feature wll nduce a change n the generalzaton error proportonal to gradent of f wth respect to th feature at pont x k gven by m k 1 f ( ) ( xk ). x In SVM classfer f w. So the SVM Recursve x Feature Elmnaton method s descrbed as follows: Step 1: Get w * as the soluton of SVM on the data set restrcted to features. Step : Select top features as ranked by * ( w ). Snce ( w * ) s proportonal to the th feature the best features are those that have greater w * ). ( D. Common Spatal Pattern () he algorthm [8] s an effectve technque to dscrmnate between two classes of EEG data. he algorthm proects the raw sgnal to a spatally fltered sgnal Z as gven n (7) that maxmzes the varance of one class whle mnmzes the varance of the other class. Z WX N Let X R denotes a matrx that represents the EEG of a sngle-tral; N and denotes the number of features and the number of measurement samples per feature respectvely. he rows of proecton matrx W are the statonary spatal flters and the columns of W -1 are the common spatal patterns. he algorthm performs smultaneous dagonalzaton of the covarance matrces of both classes. For each centered and scaled X the estmated covarance matrx n class C (C ) R s gven by ( C) 1 I C I c X X ( C { }) (5) (6) (7) (8) 6

3 where I C denotes the number of trals belongng to class C. he proecton matrx W s computed by smultaneous dagonalzaton of the two covarance matrces gven by W W ( ) ( ) W W (C) where s the dagonal egenvalues and the scalng of W s ( ) ( ) commonly determned such that I [8]. he proposed channel selecton by Wang et al. [7] s defned as follows: Optmal channels for every motor magery task are determned through the maxmums of the absolute value of the concerned spatal pattern. Let SP R and SP H denote th optmal channels of spatal pattern for rght and left hand motor magery respectvely therefore equaton (10) s calculated to obtan overall rankng where vares from 1 to the total number of channels. Fnally snce every channel has been terated twce n CH the lower rank s dscarded. As shown n equaton (10) n ths method channels are parwsely selected from both left and rght motor magery areas. CH CH 1 Fnd( Mn( SP ( ) ( ) Fnd( Max( SP R R SP SP III. MEHODOLOGY he general structure of the proposed EEG channel selecton method s shown n Fg.1. In ths method a multband sgnal decomposton flter s appled to all channels of the multchannel EEG. Subsequently the EEG sgnals are subect-specfcally fltered nto the most relevant frequency range. Fnally the relevant channels are selected usng the proposed decson tree based algorthm. L L )) )) (9) (10) the most relevant sub-bands. he fourth step employs an Ellptc bandpass flter to flter the orgnal unfltered EEG nto the smallest frequency range ncludng all the most relevant sub-bands. B. Decson ree-based Channel Selecton Decson rees are classfers that provde nterpretable solutons. Snce tranng the nducton algorthm and selectng the features are performed smultaneously the decson tree (D)-based feature selecton method s characterzed as an embedded approach [6]. he D comprses a root nternal (non-termnal) decson nodes and a set of termnal nodes or leaves each representng a class. he decson tree based feature selecton method conssts of two phases: 1- Buldng the tree for feature reducton: A tranng data set ncludng all the features s used to buld the tree. Features that do not appear n the tree are dscarded. - ree prunng for feature rankng: he remanng features are ranked backward by prunng the tree. B.1 Buldng the tree In ths work Classfcaton And Regresson ree (CAR) [11] one of the wdely used D algorthms s used. CAR s a technque that uses bnary tree structure (wth only two branches at each nternal node). An example of CAR decson tree s llustrated n Fg. Mult- Channel EEG Sgnal 4-8Hz 8-1Hz Hz Frequency Flterng Spatal Flterng Feature Selecton by MI & Fndng Best Frequency Range Flterng Full Channel EEG nto Selected Interval Fltered EEG Sgnals Feature Reducton By Buldng the Decson ree Fg.1. Proposed EEG channel selecton method Feature Rankng by ree Prunng A. Flterng by Mult Band Sgnal Decomposton Fg. 1 shows that the subect-specfc flter used n our work comprses four progressve steps descrbed as follows: he frst step employs a Chebyshev flter bank to perform bandpass flterng of EEG n multple frequency bands. he second step performs spatal flterng on each of these bands usng the algorthm. hus each par of bandpass and spatal flters yelds features whch are correspondng to the frequency range of the bandpass flter. he thrd step employs a Mutual Informaton-based feature selecton algorthm to select the best dscrmnatve features. Subsequently the best dscrmnatve features ndcate Fg.. Example of CAR decson tree he process of tree buldng starts at the root (frst nternal node) wth the entre tranng dataset beng splt nto two subsets. Smlarly the parttonng of every subset nto two subsubsets s contnued at each nternal node based on a predefned crteron. Accordng to the predefned crteron a test s conducted at each step to fnd out the most sutable feature that gves the best separaton of the tranng samples. hs work uses the gn ndex crteron [11] to buld the tree. he gn ndex provdes a measure of the mpurty degree n a dataset. For data set S the gn ndex s computed usng m p 1 gn( S) 1 (11) where p and m are the probablty of class and the total number of classes respectvely. he mnmum value of the gn ndex occurs when the set conssts of only one class; and the maxmum value occurs when all the classes n the set have equal probablty. 7

4 he goodness of a splt pont s specfed by the gn value. If the decson splts the database S nto sets S1 and S the gn value of the dvded data s gn splt n1 n ( S) gn( S1) gn( S) n n where n s the number of nstances n S. he gn value s evaluated for every possble splt of an attrbute. hs mples that not only a certan feature s chosen but also a splt pont that s used n a node. he procedure s repeated for each feature wth ts own splt ponts and fnally the feature wth the lowest gn value s selected for the next splt. he tree buldng contnues untl all the remanng nstances belong to a same class; or there s no new splttng to mprove the overall accuracy of the tree. B. Prunng the tree In ths work a prunng process based on the overall accuracy of the tree s appled to rank the features. Whle explorng over the nternal nodes from the bottom to the top the procedure checks the overall accuracy of the tree after replacng each nternal node wth a leaf labeled by the hghest represented class. herefore the nodes wth small decreases n performance are known as less mportant nodes. Consequently features are ranked accordng to the mportance of the correspondng nodes. A. Data descrpton IV. EXPERIMENS he EEG data from publcly avalable BCI Competton IV datasets a [13] are used n ths study whch comprses two classes: rght and left hand motor mageres. he EEG was recorded from nne subects usng electrodes per subect. Durng each experment the subect was gven vsual cues that ndcated four motor mageres should be performed: left hand rght hand foot and tongue. 140 sngle-trals of EEG from each class of rght and left hand motor mageres wth the tme segment of 0.5 to.5 seconds after cue were appled n ths study. he EEG data from foot and tongue motor mageres were not used. B. Data processng (1) In ths work the raw EEG s fltered usng the best subectspecfc frequency range extracted by the mult band sgnal decomposton algorthm. Subsequently the covarance of each channel s computed as the feature correspondng to the channel. hese features are used n all the channel selecton methods except -based method because n channel selecton method channels are drectly selected from the spatal patterns (refer to secton II.C). he performances of the dfferent channel selecton methods are evaluated by calculatng the accuracy of the classfcaton usng dfferent number of optmal channels. For ths purpose the common spatal flters are employed to spatally flter the EEG. hen the varances of three frst and three last rows of the fltered sgnals [1] are used as nputs of the SVM classfer. It s noted that the classfcaton accuraces of dfferent methods are evaluated by averaged fold cross-valdaton. V. EXPERIMENAL RESULS For evaluaton purpose the classfcaton accuraces after the channel reducton (the frst phase of our method) were compared wth the results obtaned from: (1) all the channels () three typcally used channels for left and rght motor mageres (C3 C4 Cz). he expermental results for 9 subects are shown n able 1. he frst row presents the averaged folds classfcaton accuraces of full channel EEG. Averaged number of selected channels by the decson tree and acheved accuraces after selectng those channels are ndcated on the second and thrd rows respectvely. Fnally obtaned accuraces by usng only C3 C4 and Cz are presented n the last row. ABLE 1 PERFORMANCE COMPARISON OF DECISION REE BASED EEG CHANNEL REDUCION (CH: CHANNELS ACC: ACCURACY #: NUMBER) Subect Mean±Std Full Ch Acc ±14.9 (%) #Selected Ch Remaned Ch Acc (%) C3 C4 Cz Acc (%) ± ± ±13.8 As can be seen n able 1 the proposed method decreased the number of electrodes on average to 38% (of electrodes) wth sustanng only a drop of 3.63% n accuracy. Interestngly enough the accuraces of selected channels n subects 4 and 6 are even more than the full channel accuraces. It can happen due to removng redundant and nosy channels whch degraded the performance. Accordng to the results the proposed method performs sgnfcantly (mean=6% and p=0.09) better than three typcal channels (C3 C4 and Cz). Furthermore compared wth the full channel the use of only three C3 C4 and Cz channels leaded to a sgnfcant drop (mean=9.4% and p=0.016) n accuracy. As a result although the use of only three C3 C4 and Cz channels certanly allevates the nconvenence of preparaton but t nevtably causes performance drop whch was n our study around 10%. On contrary the proposed method can brng the beneft of reducng the number of channels wth a small drop n accuracy. After performng channel selecton the remanng channels are ranked usng the proposed prunng method. o consder the performance of the proposed rankng method the accuracy of best ranked channels (from to all the remanng channels) were calculated and compared wth four other 8

5 channel selecton methods based on Fsher Crteron (FC) Mutual Informaton (MI) Support Vector Machne (SVM) and Common Spatal Pattern coeffcents (). Fg. 3 depcts the averaged accuracy versus dfferent number of channels selected by 5 dfferent channel selecton methods. hs fgure shows that the classfcaton accuracy of subect was close to random and the results of channel selecton were scattered. Accordng to the appled algorthm ths subect s dentfed as a BCI llterate meanng he cannot use a BCI. Hence we gnored the results obtaned by ths subect and compared the rest. Fz FC3 FC1 FCZ FC FC4 C5 C3 C1 Cz C C4 C6 CP3 CP1 CPZ CP CP4 P1 Pz P POZ FC MI Subect1 51 FC MI 49 SVM 48 D Subect3 Subect Subect Subect9 Accuracy (%) versus Number of Channels 57 Subect 56 Subect Subect Subect8 Average of 9 subects Accuracy (%) versus Number of Channels Fg.3. Comparson of 5 EEG channel selecton methods FC MI SVM D Most Important SVM Fg. 4. Vsualzaton of channels mportance for subect 1 he results of other subects llustrate that proposed method outperformed the other channel selecton methods. As can be seen the proposed method yelded superor classfcaton accuracy n selectng 3 to 6 channels. It mght be due to selectng channels accordng to the results of prevous chosen channels; hence the selected channels would be more relevant and less correlated to each other. Besde the proposed method the -based channel selecton method s capable of selectng more relevant channels compared to FC MI and SVM methods. On the contrary FC and MI methods perform rather a channel rankng than a channel selecton method. herefore they are mostly not as good as the proposed method n selectng a few channels. As t s vsble n Fg.3 a sharp decrease of accuracy around 6 to channels happens for hred SVM method. Vsualzaton of the channel postons accordng to ther ranks may support the analyss of our appled methods. As the expermental paradgm s well known we nvestgated whether the best selected channels were those stuated over or close to the motor areas. Fg. 4 vsualzes the selected channels of subect 1 for fve consdered methods where darker colors show more mportant channels whch are selected earler and lghter ones show less mportant channels. It should be noted that n ths step cross valdaton was not appled. Accordng to Fg. 4 the best channels selected by FC method are neghbor and ust n one sde of the bran (downrght). Consequently selectng a few channels results n poor performance because selected channels are full of redundant nformaton wthout supportng both task actvtes. MI channel postons are slghtly better than FC but stll quet near to each others. In subect 1 SVM recognzed top and down of the bran channels as the most mportant ones whch both parts are not related to motor area. It mght be the reason that n Fg. 3 a sharp drop n accuracy around 8 to channels are seen for hred SVMs. he preference of based method s selectng channels par wsely from both sdes of the bran. he best channels D Least Important 9

6 selected by method are CP4 and CP3 (near motor magery areas) and after a whle some channels from top are also selected. It would be the reason that based method acheved quetly good results n our experments. As t can be seen the decson tree method selected ust some of neghbor electrodes from both sdes of the bran. hus redundant nformaton s reduced and performance s ncreased. In summary vsualzaton presents that the preference of the decson three based method to the other method s selectng channels from neurophysologcal relevant areas and removng redundant and correlated channels. VI. CONCLUSIONS hs paper presents a decson tree-based method for EEG channel selecton n BCI applcatons. he proposed method frst employs a subect-specfc multband flter to flter the EEG then rrelevant channels are removed usng a decson tree. Subsequently the remanng channels are ranked usng a prunng method. Snce the decson tree selects channels based on prevously chosen features the selected channels are more relevant and less correlated to each other. Moreover snce tranng the decson tree and channel selecton are performed smultaneously the proposed method s computatonally effcent. he expermental results showed that the proposed method reduces the averaged number of electrodes from to 8.44 whereas the classfcaton accuracy decreases only 3.63%. Whle f three typcal channels (Cz C3 C4) are used the accuracy drops around 9.6%. A comparatve study of the proposed method wth other channel selecton methods usng Fsher Crteron Mutual Informaton Support Vector Machne and on 9 subects for two motor magery tasks showed that our method outperformed the others n selectng around 3 to 6 channels. A vsualzaton of the selected channels llustrated that the proposed method mproves the results by removng some of the neghborng channels and selectng those from both hemspheres of the bran. [5] L. an D. Erdogmus A. Adam M. Pavel and S. Mathan "Salent EEG Channel Selecton n Bran Computer Interfaces by Mutual Informaton Maxmzaton" n Proc. IEEE/ EMBS ' pp [6] I. Guyon and A. Elsseeff "An Introducton to Varable and Feature Selecton" Journal of Machne Learnng Research vol. 3 no. pp [7] Y. Wang S. Gao and X. Gao "Common Spatal Pattern Method for Channel Selelcton n Motor Imagery Based Brancomputer Interface" n Proc. IEEE/ EMBC' pp [8] B. Blankertz R. omoka S. Lemm M. Kawanabe and K. R. Muller "Optmzng Spatal flters for Robust EEG Sngle-ral Analyss" IEEE Sgnal Processng Magazne vol. 5 no. 1 pp [9] L. Rokach and O. Mamon "op-down nducton of decson trees classfers - a survey" Systems Man and Cybernetcs Part C: Applcatons and Revews IEEE ransactons on vol. 35 no. 4 pp [10] K. K. Ang Z. Y. Chn H. Zhang and C. Guan "Flter Bank Common Spatal Pattern (FB) n Bran-Computer Interface" n Proc. IEEE/ IJCNN ' pp [11] L. Breman Classfcaton and Regresson rees. Boca Raton: Chapman & Hall [1] H. Ramoser J. Muller-Gerkng and G. Pfurtscheller "Optmal spatal flterng of sngle tral EEG durng magned hand movement" IEEE ransactons on Rehabltaton Engneerng vol. 8 no. 4 pp [13]. Data Sets a for the BCI Competton IV ACKNOWLEDGMEN he authors would lke to thank Mr. Hamed Ahmad from Natonal Unversty of Sngapore for hs constructve comments on ths manuscrpt. REFERENCES [1] J. R. Wolpaw N. Brbaumer D. J. McFarland G. Pfurtscheller and. M. Vaughan "Bran-computer nterfaces for communcaton and control" Cln Neurophysol vol. 113 no. 6 pp [] N. Brbaumer "Bran-computer-nterface research: Comng of age" Cln Neurophysol vol. 117 no. 3 pp [3] A. Al-An and A. Al-Sukker "Effect of Feature and Channel Selecton on EEG Classfcaton" n Proc. IEEE/ EMBS' pp [4]. N. Lal M. Schroder. Hnterberger J. Weston M. Bogdan N. Brbaumer and B. Scholkopf "Support vector channel selecton n BCI" IEEE ransactons on Bomedcal Engneerng vol. 51 no. 6 pp