optimizing spatial patterns with sparse filter bands for motor imagery based brain computer...

7/26/2019 Optimizing Spatial Patterns With Sparse Filter Bands for Motor Imagery Based Brain Computer Interface 2015 Journal of Neuroscience Methods

http://slidepdf.com/reader/full/optimizing-spatial-patterns-with-sparse-filter-bands-for-motor-imagery-based 1/7

Journal of Neuroscience Methods 255 (2015) 85–91

Contents lists available at ScienceDirect

Journal of NeuroscienceMethods

journal homepage: www.elsevier .com/ locate / jneumeth

Computational Neuroscience

Optimizing spatial patterns with sparse filter bands for

motor-imagery based brain–computer interface

Yu Zhang a,∗, Guoxu Zhou b, Jing Jin a, Xingyu Wang a, Andrzej Cichocki b,c

a Key Laboratory for Advanced Control andOptimization for Chemical Processes, Ministry of Education, East China University of Science andTechnology,

Shanghai 200237, Chinab Laboratory for Advanced Brain Signal Processing, RIKENBrain Science Institute,Wako-shi, Saitama 351-0198, Japanc System Research Institute, Polish Academy of Sciences,Warsaw 00-901, Poland

h i g h l i g h t s

• This study proposes a sparse filter band common spatial pattern (SFBCSP) for optimizing the spatial patterns.• Experimental results on two public EEG datasets (BCI Competition III dataset IVa and BCI Competition IV IIb) confirm the effectiveness of SFBCSP.• The optimized spatial patterns by SFBCSP give overall better MI classification accuracy in comparison with several competing methods.• Our study suggests that the proposed SFBCSP is a potential method for improving the performance of MI-based BCI.

a r t i c l e i n f o

Article history:

Received 21 May 2015

Received in revised form 3 August 2015

Accepted 5 August 2015

Available online 13 August2015

Keywords:Brain–computer interface (BCI)

Common spatial pattern (CSP)

Electroencephalogram (EEG)

Motor imagery (MI)

Sparse regression

a b s t r a c t

Background:Common spatial pattern (CSP) has been most popularly applied to motor-imagery (MI) fea-

ture extraction for classification in brain–computer interface (BCI) application. Successful application of

CSP depends on the filter band selection to a large degree. However, the most proper band is typically

subject-specific and can hardly be determined manually.

New method: This study proposes a sparse filter band common spatial pattern (SFBCSP) for optimizing

the spatial patterns. SFBCSP estimates CSP features on multiple signals that are filtered from raw EEG

data at a set of overlapping bands. The filter bands that result in significant CSP features are then selectedin a supervised way by exploiting sparse regression. A support vector machine (SVM) is implemented on

the selected features for MI classification.

Results: Two public EEG datasets (BCI Competition III dataset IVa and BCI Competition IV IIb) are used to

validate the proposed SFBCSP method. Experimental results demonstrate that SFBCSP help improve the

classification performance of MI.

Comparison with existing methods: The optimized spatial patterns by SFBCSP give overall better MI clas-

sification accuracy in comparison with several competing methods.

Conclusions: The proposed SFBCSP is a potential method for improving the performance of MI-based BCI.

© 2015 Elsevier B.V. All rights reserved.

1. Introduction

A brain–computer interface (BCI) is an advanced commu-

nication approach that assists to establish the capabilities of

environmental control for severally disabled people (Wolpawet al.,

2002; Gao et al., 2014; Hoffmann et al., 2008; Jin et al., 2015; Zhang

et al., 2012; Rutkowski and Mori, 2015; Chen et al., 2015). BCI can

translate a specific brain activity into computer command, thereby

building a direct connection between human brain and external

∗ Corresponding author. Tel.: +86 21 64253581.

E-mail address: [email protected](Y. Zhang).

device. One of the most popularly adopted brain activities is event-

related (de)synchronization (ERD/ERS) (Pfurtscheller and Neuper,

2001; Li et al., 2013), and can be typically measured by electroen-

cephalogram (EEG). ERD/ERS can be quantified by band-power

changes occurring when subjects do motor-imagery (MI) tasks, i.e.,

imagine their limbs (left hand, right hand and foot) (Pfurtscheller

et al., 2006; Li and Zhang, 2010; Koo et al., 2015).

So far, a large number of methods have been introduced to EEG

analysis for various applications (Li et al., 2015, 2008; Arvaneh

et al., 2013; Zhang et al., 2012, 2014; Cong et al., 2010; Jin et al.,

2013; Zhang et al., 2015b). Common spatial pattern (CSP) is a very

efficient method and has been mostly applied to MI feature extrac-

tion (Ramoser et al., 2000; Higashi et al., 2012). The variance of

http://dx.doi.org/10.1016/j.jneumeth.2015.08.004

0165-0270/© 2015 Elsevier B.V. All rights reserved.

http://localhost/var/www/apps/conversion/tmp/scratch_2/dx.doi.org/10.1016/j.jneumeth.2015.08.004

http://www.sciencedirect.com/science/journal/01650270

http://www.elsevier.com/locate/jneumeth

mailto:[email protected]



mailto:[email protected]

http://crossmark.crossref.org/dialog/?doi=10.1016/j.jneumeth.2015.08.004&domain=pdf

http://www.elsevier.com/locate/jneumeth

http://www.sciencedirect.com/science/journal/01650270




86 Y. Zhang et al. / Journal of NeuroscienceMethods 255(2015) 85–91

Fig. 1. Illustration of the proposed sparse filter band common spatial pattern (SFBCSP) algorithm for motor-imagery classification.

Table 1

Classification errors (%) obtained by the CSP, FBCSP, DFBCSP, and SFBCSP methods, respectively, for BCI Competition III dataset IVa. For each subject, the lowest error is

marked in boldface.

Subject CSP FBCSP DFBCSP SFBCSP

aa 20.11 ± 3.82 9.61 ± 2.14 9.68 ± 2.36 8.46 ± 2.10

al 2.11 ± 1.45 2.18 ± 1.72 1.54 ± 1.21 1.43 ± 1.09

av 29.61 ± 4.16 27.46 ± 6.67 24.86 ± 3.22 22.57 ± 4.87

aw 6.96 ± 2.32 2.79 ± 1.53 2.18 ± 1.38 1.97 ± 1.24

ay 7.86 ± 2.86 5.46 ± 2.88 4.71 ± 2.35 5.31 ± 2.96

Average 13.33 ± 2.92 9.50 ± 2.99 8.59 ± 2.10 7.95 ± 2.45

p-value p < 0.05 p = 0.14 p = 0.27 –

Table 2

Classificationerrors (%) obtained by the CSP,FBCSP,DFBCSP, and SFBCSP methods,respectively, for BCICompetitionIV dataset IIb.For eachsubject,the lowest erroris marked

in boldface.

Subject CSP FBCSP DFBCSP SFBCSP

B0103T 23.44 ± 5.88 22.50 ± 4.61 22.06 ± 4.48 21.85 ± 4.79

B0203T 44.44 ± 6.97 44.06 ± 6.05 42.75 ± 5.89 41.25 ± 5.84B0303T 47.38 ± 8.10 46.25 ± 6.95 44.81 ± 6.50 44.19 ± 6.09

B0403T 1.94 ± 1.50 1.13 ± 1.13 1.19 ± 1.43 1.15 ± 1.23

B0503T 11.81 ± 4.77 9.56 ± 2.42 7.56 ± 2.14 7.94 ± 2.35

B0603T 30.63 ± 5.60 21.94 ± 3.44 18.31 ± 3.17 17.68 ± 3.62

B0703T 16.56 ± 4.85 13.50 ± 2.90 10.50 ± 1.83 9.75 ± 1.25

B0803T 13.44 ± 3.02 11.25 ± 2.78 11.38 ± 2.93 11.13 ± 3.21

B0903T 18.25 ± 4.63 16.12 ± 3.39 15.25 ± 3.33 14.50 ± 3.61

Average 23.10 ± 5.04 20.70 ± 3.74 19.31 ± 3.52 18.83 ± 3.55

p-value p < 0.01 p < 0.01 p < 0.05 –

band-pass filtered signals has been known to be equal to the band-

power. Since CSP finds spatial filters to maximize the variance of

the projected signal from one class while minimizing it for another

class, it provides a natural approach to effectively estimate the dis-

criminant information of MI (Blankertz et al., 2008). However, toguarantee the successful application of CSP to MI classification,

a pre-specified filter band is required to accurately capture the

band-power changes resulting from ERD/ERS (Ang et al., 2008).

Unfortunately, the most proper filter band is typically subject-

specific and can hardly be determined in a manual way. A poor

selection of the filter band may result in low effectiveness of CSP

(Sun et al., 2010).

Although a wide filter band (i.e., 8–30Hz) was usually adopted

for CSP in MI classification, an increasing number of studies sug-

gested that the optimization of filter band could significantly

improve classification accuracy (Ang et al., 2008; Sun et al., 2010;

Lemm et al., 2005; Dornhege et al., 2006; Novi et al., 2007). So

far, two types of approaches have been mainly proposed to fix the

problem of filter band selection. One is simultaneous optimization

of spectral filterswithin theCSP (Lemm et al., 2005; Dornhege et al.,

2006; Higashi and Tanaka, 2013) while another one is selection of

significant CSP features from multiple frequency bands (Ang et al.,

2008; Novi et al., 2007; Thomas et al., 2009).

By extending CSP to state space, common spatio-spectral pat-tern (CSSP) (Lemm et al., 2005) was proposed to optimize a simple

FIR filter by employing a temporal delay within CSP. A further

extension of CSSP is called common sparse spectral spatial pattern

(CSSSP) (Dornhege etal., 2006), which optimizesan adaptive FIRfil-

ter simultaneously with CSP. More recently, (Higashi and Tanaka,

2013) proposed to simultaneously learn multiple FIRfilters andthe

associated spatial weights by maximizing a cost function extended

from CSP.

On theotherhand, an alternative approach is sub-bandcommon

spatial pattern (SBCSP) (Novi etal., 2007). Instead of simultaneously

optimizing a spectral filter within CSP, SBCSP filtered EEG signals

using multiple filter bands and extracted the sub-band CSP fea-

tures for classification with score fusion. Although SBCSP achieved

superior classification accuracy over both CSSP and CSSSP



Y. Zhang et al./ Journal of NeuroscienceMethods 255 (2015) 85–91 87

50

60

70

80

90

100

C l a s s i f i c a t i o n

a c c u r a c y ( % )

a a a l a v a w a y

B 0 1 0

3 T

B 0 2 0

3 T

B 0 3 0

3 T

B 0 4 0

3 T

B 0 5 0

3 T

B 0 6 0

3 T

B 0 7 0

3 T

B 0 8 0

3 T

B 0 9 0

3 T

A v e r a g

e

CSP

FBCSP

DFBCSP

SFBCSP

Fig. 2. Classification accuracies of all subjectsfrom both BCICompetition IIIdataset IVa andBCI CompetitionIV dataset IIb. Higheraccuracyon average wasachieved by the

proposed SFBCSPmethod in comparison to theCSP, FBCSP and DFBCSPmethods.

0.05 0.1 0.15 0.2 0.25

76

77

78

79

80

81

82

83

84

85

86

λ

C l a s s i f i c a t i o n a c c u r a c y ( % )

Fig. 3. Effects of varying the regularization parameter on the classification accu-

racy obtained by the SFBCSP method for subject “B0103T”. The parameter value

giving the highest accuracy was highlighted in a red circle. (Forinterpretation of the

references to color in this figurelegend,the readeris referred to theweb version of

this article.)

(Novi et al., 2007), it ignored the correlation among CSP features

from different filter bands. (Ang et al., 2008) introduced a filter

bank common spatial pattern (FBCSP) to select the optimal filterbands through estimating the mutual information among CSP fea-

turesat several fixedfilter bands. Higherclassificationaccuracywas

achieved by FBCSP over SBCSP. Subsequently, Thomas et al. (2009)

proposed a discriminantFBCSP(DFBCSP)usingFisherratio to select

subject-specific filter bands instead of fixed ones, which enhanced

accuracy of the FBCSP.

In this study,we will focus on further improvingthe secondtype

of approaches, i.e., CSP feature selection from multiple frequency

bands. To this end, we propose a sparse filter band common spa-

tial pattern (SFBCSP) by exploiting sparse regression for automatic

band selection. Recently, sparse regression has been increasingly

applied to BCI and shown its efficiency in ERP feature extraction to

alleviate overfitting (Zhang et al., 2014). In the proposed method,

CSPfeaturesare estimatedon multiple signals that arefiltered from

raw EEG data at a set of overlapping filter bands. Sparse regressionis implemented to supervisedly select the significant CSP features

with class labels. A support vector machine (SVM) with linear ker-

nel is then trained on the selected features to classify MI tasks.

The proposed SFBCSP is validated on two public EEG datasets (BCI

Competition IIIdataset IVa andBCI Competition IV dataset IIb), and

compared with CSP, FBCSP and DFBCSP. Experimentalresults show

SFBCSP achieves superior classification accuracy.

2. SFBCSPfor MI classification

2.1. Feature extraction by CSP

Common spatial pattern (CSP) is an effective method for feature

extraction in discriminating two classes of data. In recent years,CSP has become greatly popular for the extraction of EEG features

related to motor imagery. Consider two classes of EEG samples X i,1and X i,2 ∈ R

C ×P recorded from i-th trial, where C is the number of

channels, and p denotes the number of samples. Both EEG samples

are bandpass filtered within a specified frequency band. Without

loss of generality, we assumeX i,1 and X i,2 have been centered. Spa-

tial covariance matrix of the class l (l=1, 2) is then computed

by

l =1

N l

N li=1

X i,lX T i,l, (1)

where N l is the number of trials in class l. CSP is aimed at learning

linear transforms (also called spatial filters) to maximize the ratioof transformed data variance between the two classes:

max w

J ( w ) = w T 1 w

w T 2 w s.t. w 2 = 1, (2)

where · 2 denotes the l2-norm, and w ∈ RC is a spatial filter. The

maximization of Rayleigh quotient J ( w ) can be achieved by solving

the following generalized eigenvalue problem:

1 w = 2 w . (3)

The eigenvectors are then obtained to form the learned spatial

filters matrix ˜ W . The projection Z of a given EEG sample X is then

given by

Z = ˜ W

T

X . (4)




Generally, only a small numberM of projected signals are used

for feature computation. With the M first and last rows of Z, i.e.,

Zm, m∈ {1, . . ., 2M }, the feature vector is then formed asg = [ g 1, . . .,

g 2M ]T with entries

g m = log

var(Zm)

2M

h=1var(Zh)

, (5)

where var(Z) denotes the variance of each row of Z.

2.2. Sparse filter band CSP

The successful application of CSP to MI classification largely

depends on the frequency band selection for bandpass filtering of

EEG(Blankertzet al., 2008). A good selectionof filter bands cancon-

siderably improve the MI classification accuracy, whereas a poor

one will result in low effectiveness of CSP (Ang et al., 2008; Thomas

et al., 2009). The optimal filter band is typically subject-specific and

can hardly be determined in a manual way. To solve the problem,

this study proposes a sparse filter band common spatial pattern

(SFBCSP) for automatic selection of filter bands.

Instead of directlyusinga wide filterband,we perform bandpass

filtering on rawEEG by using a setof overlapping subbands ( fb1, . . .,

fbK ). These subbands are chosen from the frequency range 4–40

Hz with bandwidth of 4Hz and overlapping rate of 2Hz, i.e., fb1 =

4–8Hz, fb2 = 6–10Hz, . . ., fbK = 36–40 Hz where K = 17. The adopted

settings of bandwidth and overlapping rate are consistent with

those in literature (Thomas et al., 2009). CSP is then implemented

on thefiltered signals at each sub-band to calculatethe correspond-

ingfeaturesby (5). Asa result, 2MK features are extractedfrom each

EEG sample. With the extracted CSP features, we can construct the

following feature set:

G =

g 1,1 . . . g 1,2MK

.... . .

...

g N,1 . . . g N,2MK

, (6)

where g i, j denotes the j-th feature extracted from EEG sample at

i-th trial, and N =N 1 +N 2.

Here, we propose to use the following sparse regression model

(i.e, the well-known Lasso estimate (Tibshirani, 1996)) for signifi-

cant CSP feature selection

u = arg minu

1

2Gu− y 2

2 + u1, (7)

where · 1 denotes the l1-norm, y ∈ RN is a vectorcontaining class

labels {1, 2}, u is a sparse vector to be learned, is a positive reg-

ularization parameter for controlling the sparsity of u, and larger

could result in more sparse u. The coordinate descent algorithm

(Friedmanet al.,2010) is adopted to solvethe optimization problem

in (7). The coordinate-wise update form foru is given by:

u j ← S

N i=1

g i,j( yi − ˜ y( j)i

),

, (8)

where ˜ y( j)i =

d /= j g i,dud is the fitted value excluding the contribu-

tion from g i, j, and S (a, ) is a shrinkage-thresholding operator:

sign(a)(|a| − )+ =

a− if a >

0 if |a| ≤

a+ if a < −.(9)

We can obtain the optimized sparse vector u satisfying (7)

through repeating the update form in (9) f or j= 1, 2, . . . , D, 1, 2,

. . .until convergence. The column vectors in G corresponding to

those zero entries in u are excluded to form a optimizedfeature set

G that is of lower dimensionality in comparison with G. The given

determines the sparsity degree of u, and hence the selection of

CSP features.

Support vector machine (SVM) is adopted and implemented on

the optimizedfeatureset G totraintheclassifier.Foranewtestsam-

ple X ∈ RC ×P , the corresponding subband feature vector g ∈ R2MK

is estimated by CSP. The optimal features are selected according to

the sparse vector u andthenfed into thetrainedSVM forMI classifi-

cation. A simple linear kernel is adopted for the SVM training. Fig. 1

illustrates the proposed SFBCSP algorithm for MI classification.

3. Experimental study

3.1. Public BCI Competition datasets

3.1.1. BCI Competition III dataset IVa

This dataset is recorded for five subjects (named “aa”, “al”, “av”,

“aw”, and “ay”) at 118 electrodes during right hand and foot MI

tasks. For each subject, a total of 280 trials of EEG measurements

areavailable (half foreach class of MI). The visualcue indicating the

MI task at each trial lasted for 3.5 s. The sampling rate was 100 Hz.More details about thedatasetcan be found atwebsite http://www.

bbci.de/competition/iii/.

3.1.2. BCI Competition IV dataset IIb

This dataset is recorded from nine subjects at electrodes C3, Cz

and C4 with sampling rate of 250Hz, during right hand and left

hand MI tasks. For comparison with the results reported in litera-

ture (Thomas et al., 2009), this study only uses the third training

sessions of the dataset, i.e., “B0103T”, “B0203T”, . . ., “B0903T”. For

each subject, a total 160 trials of EEG measurements are available

(half for each class of MI). In each trial, the subject was indicated by

a visual cue to perform MI task for 4.5s. See website http://www.

bbci.de/competition/iv/ for more details about the dataset.

3.2. Evaluation scheme

The effectiveness of the proposed SFBCSP method is tested on

the two above-mentioned datasets, in comparison with the CSP,

FBCSP (Ang etal.,2008) and DFBCSP methods (Thomas et al., 2009).

On each trial, a segment is extracted starting from 0.5 to 2.5 s after

the visual cue. A wide filter band 4–40Hz is adopted for CSP while

17 overlapping subbands as mentioned in section 2.2 are chosenfor

FBCSP, DFBCSP and SFBCSP. For each of the two datasets, a 10× 10-

fold cross-validation is implemented to evaluate the classification

performance. After the optimal CSP features are selected, SVM is

trained to classify the MI tasks. The regularization parameters in

SFBCSPandC in SVM are determined by cross-validation on training

data. The number of CSP filters is set to 2 (i.e.,M =1).

3.3. Experimental results

3.3.1. BCI Competition III dataset IVa

Table 1 presents classification errors derived by the CSP, FBCSP,

DFBCSP and SFBCSP methods, respectively, for the five subjects. All

of the three modified CSP algorithms outperformed the standard

CSP. The proposed SFBCSP method further yielded lower average

errors than those of the FBCSP and DFBCSP methods. The average

error rate reductions achieved by SFBCSP were 40.36 %, 16.32% and

7.45 % in comparison with the CSP, FBCSP and DFBCSP methods,

respectively. The SFBCSP method obtained significant improve-

ment on classification accuracy compared to CSP ( p < 0.05).

http://www.bbci.de/competition/iii/


http://www.bbci.de/competition/iv/

















Fig.4. Sparse vectorsand themost significant spatial filters learned by theSFBCSPmethod for eachof thefive subjectsin BCICompetitionIII dataset IVa.The feature indicated

as 1,2, . . ., 34 correspond to thefilter subbands 4–8Hz, 6–10 Hz,. . ., 36–40Hz, respectively.

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0103T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0203T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0303T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0403T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0503T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0603T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0703T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0803T

5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

B0903T

Feature index

V a l u e

s o f u

Fig. 5. Sparse vectors leaned by the SFBCSP method for each of the nine subjects in BCI Competition IV dataset IIb. The feature indicated as 1, 2, . . ., 34 correspond to the

filter subbands 4–8Hz, 6–10 Hz,. . ., 36–40Hz, respectively.

3.3.2. BCI Competition IV dataset IIb

Table 2 summarizes classification errors for the nine subjects.

The SFBCSP method yielded lower classification errors for most

subjects than those of the other three methods. The classification

performance of SFBCSP was significantly better than those of CSP

( p < 0.01), FBCSP ( p < 0.01) and DFBCSP ( p < 0.05). The average errorreductions achieved by the SFBCSP method were 18.48 %, 9.03 %

and2.49% in comparison with theCSP,FBCSPand DFBCSPmethods,

respectively.

Additionally, the classification accuracies for all subjects from

both the aforementioned datasets are depicted in Fig. 2. In sum-

mary, the proposed SFBCSPmethod yielded higheroverall accuracy

over the CSP, FBCSP and DFBCSP methods for MI classification.

4. Discussion

In the proposed SFBCSP method, the regularization parameter

required by sparse regression played an important role in CSP

feature selection. A too large may exclude useful features from

the model while a too small one could not eliminate redundancy

effectively. In this study, the optimal subject-specific was deter-

mined by cross-validation on training data. Fig. 3 presents an

example to see the effects of varying on the classification accu-

racy for subject “B0103T”. The optimal validation accuracy was

achieved at =0.14 that was chosen for the subsequent test pro-

cedure. A potential problem for the proposed method is thatthe cross-validation procedure is usually time-consuming and is

not applicable for small sample size datasets (Lu et al., 2010).

Instead of cross-validation, Bayesianinferenceprovidesan effective

approach to automatically and quickly estimate the model param-

eters under the so-called evidence framework (MacKay, 1992;

Bishop, 2006). Recently, sparse Bayesian learning (Tipping, 2001)

has been increasinglyappliedto the automatic determination regu-

larization in EEGanalysis (Wu et al., 2011, 2014, 2014; Zhang et al.,

2015a; Zhang and Rao, 2011). In fact, Bayesian estimation of the

optimal may further improve the efficiency of SFBCSP, which will

be investigated in our future study.

Moreover, sparse regularization has been increasingly applied

to EEG analysis, such as feature reduction (Zhang et al., 2014;

Blankertz et al., 2002), channel optimization (Arvaneh et al., 2011)




−3 −2 −1 0 1−4

−3

−2

−1

0

1

CSP

Value of the best feature 1

V a l u e o f t h e b e s t f e a t u r e 2

−3 −2 −1 0 1−4

−3

−2

−1

0

1

FBCSP


V a l u e o f t h e b e s t f e a t u r e 2

−3 −2 −1 0 1−4

−3

−2

−1

0

1

DFBCSP


V

a l u e o f t h e b e s t f e a t u r e 2

−3 −2 −1 0 1−4

−3

−2

−1

0

1

SFBCSP


V

a l u e o f t h e b e s t f e a t u r e 2

Fig. 6. Distributions of the most significant two features obtained by CSP, FBCSP, DFBCSP and SFBCSP, respectively, from subject B0703T.

and trial selection (Zhang et al., 2013). This study exploited sparse

regression for filter band optimization to further improving the

MI classification accuracy. With a properly determined regular-

ization parameter , the optimal sparse vector u can be learned

from (7) to effectively select the significant CSP features extracted

from multiple filter subbands. Fig. 4 presents the sparse vectors

and the most significant spatial filters learned by SFBCSP for each

subject in BCI Competition III dataset IVa. Fig. 5 shows the sparse

vectors learned by SFBCSP for each subject in BCI Competition IVdataset IIb. It can be seen that the distribution of significant filter

bands is sparse and subject-specific. The spatial filter correspond-

ing to the most significant band accurately captured the features in

the sensorimotor areas. The significant feature appeared at multi-

ple subbands for some subjects but at an exact subband for some

other subjects (e.g., “B0403T” and “B0903T”). This indicates that an

effective optimization of filter band forCSP is necessaryto improve

the MI classification accuracy. Fig. 6 depicts distributions of the

most significant two features derived by CSP, FBCSP, DFBCSP and

SFBCSP, respectively, from subject B0703T. All of the three modi-

fied CSP algorithms provided more separable feature distributions

in comparison with the standard CSP. The highest discriminability

of features was achieved by SFBCSP.

In recent years, multiway learning algorithms have showntheir promising potentials for the collaborative optimization in the

spatial, temporal and spectral dimensions of brain signals (Cong

et al., 2013, 2015; Zhang et al., 2013, 2014; Cichocki et al., 2008;

Wu et al., 2014; Zhou et al., 2013; Zhao et al., 2015). A combination

of collaborative multiway optimization and regularization for filter

band selection could further benefit to improve the effectiveness

of CSP for MI-based BCI.

5. Conclusions

In this study,we introduced a sparsefilter band commonspatial

pattern (SFBCSP) algorithm to automatically select the significant

filter bands for improving MI task classification performance in

BCI. In the proposed method, CSP features were first estimated on

multiple signals filtered from raw EEG data at a set of overlapping

filter bands. Significant CSP features were then selected by explo-

iting sparse regression and were used in SVM to classify MI tasks.

Experimental studies on two public EEG datasets (BCI Competition

IIIdatasetIVa andBCI Competition IV dataset IIb) indicatedthat the

proposed SFBCSP yielded higher overall classification accuracy in

comparison with several other competing methods.

Acknowledgements

This study was supported in part by the Nation Nature Sci-

ence Foundation of China under Grant 61305028, Grant 91420302,

Grant 61573142, Grant 61203127, Grant 61201124, Fundamen-

tal Research Funds for the Central Universities under Grant

WH1314023, Grant WG1414005, Grant WH1414022, the Guang-

dong Natural Science Foundation under Grant 2014A030308009,

and JSPS KAKENHI Grant 26730125.

References

Ang K, Chin Z, Zhang H, Guan C. Filter bank common spatial pattern (FBCSP) inbrain–computer interface. In: IEEE International Joint Conference on NeuralNetworks(IJCNN 2008); 2008. p. 2390–7.

Arvaneh M, Guan C,AngK, QuekC. Optimizing the channel selection and classifica-tion accuracy in EEG-based BCI. IEEE Trans Biomed Eng 2011;58(6):1865–73.

Arvaneh M, CuntaiG, Kai K,Chai Q. Optimizing spatial filters by minimizingwithin-class dissimilarities in electroencephalogram-based brain–computer interface.IEEE Trans Neural Netw Learn Syst 2013;24(4):610–9.

Bishop C. Pattern recognition and machine learning. Springer; 2006.Blankertz B, Curio G, Müller K. Classifying single trial EEG: Toward brain computer

interfacing. In: Proc Advances in Neural Information Processing Systems (NIPS01) 14; 2002. p.157–64.

Blankertz B, Tomioka R, Lemm S, Kawanabe M, Müller K. Optimizing spatial filtersforrobust EEGsingle-trial analysis. IEEE Signal Process Mag 2008;25(1):41–56.

Chen L, Jin J, Zhang Y, Wang X, Cichocki A. A survey of the dummy face and humanface stimuli used in BCI paradigm.J Neurosci Methods 2015;239:18–27.

Cichocki A, WashizawaY, Rutkowski T, Bakardjian H, Phan A-H, Choi S, LeeH, ZhaoQ, Zhang L, Li Y. Noninvasive BCIs: Multiway signal-processing arraydecompo-sitions. IEEE Comput 2008;41(10):34–42.

Cong F, Kalyakin I, Huttunen-Scott T, Li H, Lyytinen H, Ristaniemi T. Single-trialbasedindependent component analysis on mismatch negativity in children. Int

J Neural Syst 2010;20(4):279–92.

http://refhub.elsevier.com/S0165-0270(15)00285-X/sbref0005


































































































































































































Cong F, Phan A-H, Astikainen P, Zhao Q, Wu Q, Hietanen J, Ristaniemi T, CichockiA. Multi-domain feature extraction for small event-related ptentials throughnonnegative multi-way array decomposition from low dense array EEG. Int JNeural Syst 2013;23(2):1350006.

Cong F, LinQ, Kuang L, Gong X, Astikainen P, Ristaniemi T. Tensor decomposition of EEGsignals: a brief review. J Neurosci Meth 2015;248:59–69.

Dor nh ege G, Blankert z B, Krauledat M, Losch F, Cur io G , Müller K. Combinedoptimization ofspatialand temporalfiltersfor improvingbrain–computerinter-facing. IEEE Trans Biomed Eng 2006;53(11):2274–81.

FriedmanJ, Hastie T, TibshiraniR. Regularizationpathsfor generalized linear modelsvia coordinate descent. J Stat Softw2010;33(1):1–22.

Gao S,WangY, Gao X,Hong B. Visual and auditory brain–computer interfaces. IEEETrans Biomed Eng 2014;61(5):1436–47.

Higashi H, TanakaT. Simultaneousdesign of FIRfilter banks andspatial patterns forEEG signal classification. IEEE Trans Biomed Eng 2013;60(4):1100–10.

Higashi H, Cich ocki A, Tanaka T. Regularization using geometric informationbetween sensors capturing features from brain signals. In: IEEE InternationalConference on Acoustics, Speech and Signal Processing (ICASSP 2012); 2012. p.721–4.

Hoffmann U, Vesin J, Ebrahimi T, Diserens K. An efficient P300-basedbrain–c omputer interface for disabled sub jects. J Neurosci Methods2008;167(1):115–25.

Jin J, Sellers EW, Zhang Y, Daly I, Wang X, Cichocki A. Whether generic modelworksfor rapid ERP-based BCI calibration. J Neurosci Methods 2013;212(1):94–9.

Jin J, Sellers E, Zhou S, Zhang Y, Wang X, Cichocki A. A P300brain computerinterfacebased on a modification of themismatch negativity paradigm. Int J Neural Syst2015;25(3):1550011.

KooB, LeeH, NamY, KangH,KohC,ShinH, ChoiS. A hybrid NIRS-EEGsystemfor self-pacedbrain computerinterfacewith online motor imagery. J Neurosci Methods2015;244:26–32.

Lemm S, Blankertz B, Curio G, Müller K. Spatio-spectral filters for improving theclassification of singletrial EEG. IEEE Trans Biomed Eng2005;52(9):1541–8.

Li J, Zhang L. Bilateral adaptation and neurofeedback for brain computer inter-facesystem. J Neurosci Methods 2010;193(2):373–9.

Li Y, Guan C, Li H, Chin Z. A self-training semi-supervised SVM algorithm and itsapplication in an EEG-based brain computer interface speller system. PatternRecognit Lett 2008;29(9):1285–94.

Li J, L iang J, Zhao Q, L i J, Hong K, Zhang L. Design of assistive wheelchair systemdirectly steered by human thoughts. Int J Neural Syst 2013;23(3):1350013.

Li J , St ru zik Z, Z hang L, Cich ocki A. Feature learning from incomplete EEG withdenoising autoencoder. Neurocomputing 2015;165:23–31.

Lu H, EngH, Guan C, Plataniotis K, VenetsanopoulosA. Regularized common spatialpattern with aggregation for EEG classification in small-sample setting. IEEETrans Biomed Eng 2010;57(12):2936–46.

MacKay D. Bayesian interpolation. Neural Comput 1992;4(3):415–47.NoviQ, Guan C,Dat TH,XueP. Subband common spatial pattern forbrain–computer

interface. In: Proceedings of 3rd International Conference Neural Engineering.EEEEngineeringin Medicine and Biology Society. (EMBS); 2007. p. 204–7.

Pfurtscheller G, Neuper C. Motor imagery and direct brain–computer communica-tion. Proc IEEE 2001;89(7):1123–34.Pfurtscheller G, Brunner C, Schlögl A, Lopes da Silva F. Murhythm (de), synchro-

nization and EEG single-trial classification of different motor imagery tasks.NeuroImage 2006;31(1):153–9.

Ramoser H, Muller-Gerking J, PfurtschellerG. Optimal spatial filtering of single trialEEGduring imagined hand movement. IEEE Trans Rehab Eng2000;8(4):441–6.

Rutkowski TM, Mori H. Tactile and bone-conduction auditory braincomputer inter-facefor vision andhearingimpairedusers.J NeurosciMethods2015;244:45–51.

Sun G, Hu J, Wu G. A novel frequency band selection method for common spatialpattern in motor imagery basedbrain computerinterface. In: IEEEInternational

Joint Conference on Neural Networks (IJCNN 2010); 2010. p. 1–6.ThomasK, GuanC, Lau C,VinodA, Ang K. A new discriminativecommon spatial pat-

tern method for motor imagery brain–computer interfaces. IEEE Trans BiomedEng 2009;56(11):2730–3.

Tibshirani R. Re gr ess ion s hr inkage and s election via t he lasso . J R St at S oc B1996;58(1):267–88.

Tipping M. Sparse Bayesian learning and the relevance vector machine. J MachineLearn Res 2001;1:211–44.

Wolpaw JR, Birbaumer N, McFarland DJ, Pfutscheller G, Vaughan TM.

Brain–computer interfaces for communication and control. Clin Neurophysiol2002;113(6):767–91.

Wu W, Chen Z, Gao S, Brown E. A hierarchical bayesian approach for learn-ing sparse spatio-temporal decompositions of multichannel EEG. NeuroImage2011;56(4):1929–45.

WuW, WuC,GaoS,LiuB,Li Y,Gao X. Bayesian estimationof ERP components frommulticondition and multichannel EEG. NeuroImage 2014;88:319–39.

Wu W, Chen Z, Gao X, Li Y, Brown E, Gao S. Probabilistic common spatial pat-terns for multichannel EEG analysis. IEEE Trans Pattern Anal Mach Intell2014;37(3):639–53.

Wu Q, Zhang L, Cichocki A. Multifactor sparse feature extraction using convolutivenonnegative tucker decomposition. Neurocomputing 2014;129:17–24.

Zhang Z, Rao B. Sparse signal recovery with temporally correlated source vec-tors using sparse Bayesian learning. IEEE J Sel Top Signal Process 2011;5(5):912–26.

ZhangY, Zhao Q,JinJ, WangX, CichockiA. A novel BCIbased onERP components sen-sitiveto configural processing of human faces. J Neural Eng2012;9(2):026018.

Zhang Y,JinJ, QingX, WangB,WangX. LASSO basedstimulus frequency recognitionmodel for SSVEP BCIs. Biomed Signal Process Control 2012;7(4):104–11.

ZhangY, Zhou G,Jin J,WangM, Wang X,CichockiA. L1-regularizedmultiway canon-ical correlation analysis for SSVEP-based BCI.IEEE TransNeuralSyst Rehabil Eng2013;21(6):887–96.

Zhang Y, Zhou G, Zhao Q, Jin J, Wang X, Cichocki A. Spatial-temporal discriminantanalysisfor ERP-basedbrain–computerinterface. IEEETransNeuralSyst RehabilEng 2013;21(2):233–43.

Zhang Y,Zhou G,Jin J,ZhaoQ, WangX, Cichocki A. Aggregation of sparse linear dis-criminant analysis for event-related potential classification in brain–computerinterface. Int J Neural Syst 2014;24(1):1450003.

Zhang Y, Zhou G, Jin J, Wang X, Cichocki A. Frequency recognition in SSVEP-based BCI using multiset canonical correlation analysis. Int J Neural Syst2014;24(4):1450013.

Z hang Y, Xu P , Che ng K, Yao D. Multivariate synchronization index for frequencyrecognition of SSVEP-based brain–computer interface. J Neurosci Methods2014;221:32–40.

Y. Zhang, G. Zhou, J. Jin, Q. Zhao, X. Wang and A. Cichocki Sparse Bayesian classifi-cation of EEG for brain–computer interface, IEEE Trans Neural Netw Learn Syst(2015), In press.

ZhangY, Zhou G,Jin J, Wang X, Cichocki A. SSVEP recognition using common featureanalysis in brain–computer interface. J Neurosci Methods 2015b;244:8–15.Zhao Q, Zhou G, Zhang L, Cichocki A, Amari SI. Bayesian robust tensor factoriza-

tion for incomplete multiway data. IEEE Trans Neural Netw Learn Syst 2015.,http://dx.doi.org/10.1109/TNNLS.2015.2423694, in press.

Zhou G, Cichocki A, Xie S. Accelerated canonical polyadic decomposition byusing mo de r eduction . I EE E Trans Neu ral Net w Learn S ys t 2 01 3; 24 (1 2):2051–62.
























































































































































http://localhost/var/www/apps/conversion/tmp/scratch_2/dx.doi.org/10.1109/TNNLS.2015.2423694





















































































































































































































































































































































































































































































































































































































































































































































































optimizing spatial patterns with sparse filter bands for motor imagery based brain computer...

Documents