Академический Документы
Профессиональный Документы
Культура Документы
Computational Neuroscience
h i g h l i g h t s
This study proposes a sparse lter 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) conrm the effectiveness of SFBCSP.
The optimized spatial patterns by SFBCSP give overall better MI classication 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 August 2015
Keywords:
Braincomputer 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) feature extraction for classication in braincomputer interface (BCI) application. Successful application of
CSP depends on the lter band selection to a large degree. However, the most proper band is typically
subject-specic and can hardly be determined manually.
New method: This study proposes a sparse lter band common spatial pattern (SFBCSP) for optimizing
the spatial patterns. SFBCSP estimates CSP features on multiple signals that are ltered from raw EEG
data at a set of overlapping bands. The lter bands that result in signicant CSP features are then selected
in a supervised way by exploiting sparse regression. A support vector machine (SVM) is implemented on
the selected features for MI classication.
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
classication performance of MI.
Comparison with existing methods: The optimized spatial patterns by SFBCSP give overall better MI classication 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 braincomputer interface (BCI) is an advanced communication approach that assists to establish the capabilities of
environmental control for severally disabled people (Wolpaw et 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 specic brain activity into computer command, thereby
building a direct connection between human brain and external
device. One of the most popularly adopted brain activities is eventrelated (de)synchronization (ERD/ERS) (Pfurtscheller and Neuper,
2001; Li et al., 2013), and can be typically measured by electroencephalogram (EEG). ERD/ERS can be quantied 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
efcient method and has been mostly applied to MI feature extraction (Ramoser et al., 2000; Higashi et al., 2012). The variance of
86
Fig. 1. Illustration of the proposed sparse lter band common spatial pattern (SFBCSP) algorithm for motor-imagery classication.
Table 1
Classication 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
aa
al
av
aw
ay
Average
p-value
20.11 3.82
2.11 1.45
29.61 4.16
6.96 2.32
7.86 2.86
13.33 2.92
p < 0.05
9.61 2.14
2.18 1.72
27.46 6.67
2.79 1.53
5.46 2.88
9.50 2.99
p = 0.14
9.68 2.36
1.54 1.21
24.86 3.22
2.18 1.38
4.71 2.35
8.59 2.10
p = 0.27
SFBCSP
8.46
1.43
22.57
1.97
5.31
7.95
2.10
1.09
4.87
1.24
2.96
2.45
Table 2
Classication errors (%) obtained by the CSP, FBCSP, DFBCSP, and SFBCSP methods, respectively, for BCI Competition IV dataset IIb. For each subject, the lowest error is marked
in boldface.
Subject
CSP
FBCSP
DFBCSP
B0103T
B0203T
B0303T
B0403T
B0503T
B0603T
B0703T
B0803T
B0903T
Average
p-value
23.44 5.88
44.44 6.97
47.38 8.10
1.94 1.50
11.81 4.77
30.63 5.60
16.56 4.85
13.44 3.02
18.25 4.63
23.10 5.04
p < 0.01
22.50 4.61
44.06 6.05
46.25 6.95
1.13 1.13
9.56 2.42
21.94 3.44
13.50 2.90
11.25 2.78
16.12 3.39
20.70 3.74
p < 0.01
22.06 4.48
42.75 5.89
44.81 6.50
1.19 1.43
7.56 2.14
18.31 3.17
10.50 1.83
11.38 2.93
15.25 3.33
19.31 3.52
p < 0.05
band-pass ltered signals has been known to be equal to the bandpower. Since CSP nds spatial lters 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 discriminant information of MI (Blankertz et al., 2008). However, to
guarantee the successful application of CSP to MI classication,
a pre-specied lter band is required to accurately capture the
band-power changes resulting from ERD/ERS (Ang et al., 2008).
Unfortunately, the most proper lter band is typically subjectspecic and can hardly be determined in a manual way. A poor
selection of the lter band may result in low effectiveness of CSP
(Sun et al., 2010).
Although a wide lter band (i.e., 830 Hz) was usually adopted
for CSP in MI classication, an increasing number of studies suggested that the optimization of lter band could signicantly
improve classication 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 x the
problem of lter band selection. One is simultaneous optimization
SFBCSP
21.85
41.25
44.19
1.15
7.94
17.68
9.75
11.13
14.50
18.83
4.79
5.84
6.09
1.23
2.35
3.62
1.25
3.21
3.61
3.55
of spectral lters within the CSP (Lemm et al., 2005; Dornhege et al.,
2006; Higashi and Tanaka, 2013) while another one is selection of
signicant 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 pattern (CSSP) (Lemm et al., 2005) was proposed to optimize a simple
FIR lter by employing a temporal delay within CSP. A further
extension of CSSP is called common sparse spectral spatial pattern
(CSSSP) (Dornhege et al., 2006), which optimizes an adaptive FIR lter simultaneously with CSP. More recently, (Higashi and Tanaka,
2013) proposed to simultaneously learn multiple FIR lters and the
associated spatial weights by maximizing a cost function extended
from CSP.
On the other hand, an alternative approach is sub-band common
spatial pattern (SBCSP) (Novi et al., 2007). Instead of simultaneously
optimizing a spectral lter within CSP, SBCSP ltered EEG signals
using multiple lter bands and extracted the sub-band CSP features for classication with score fusion. Although SBCSP achieved
superior classication accuracy over both CSSP and CSSSP
87
100
90
80
70
CSP
FBCSP
60
DFBCSP
SFBCSP
e
ag
er
Av
90
3T
3T
B0
3T
80
B0
70
3T
B0
60
3T
B0
50
3T
B0
40
3T
B0
30
3T
B0
20
B0
10
3T
ay
B0
aw
av
aa
al
50
Fig. 2. Classication accuracies of all subjects from both BCI Competition III dataset IVa and BCI Competition IV dataset IIb. Higher accuracy on average was achieved by the
proposed SFBCSP method in comparison to the CSP, FBCSP and DFBCSP methods.
86
85
84
83
82
81
80
79
78
77
76
0.05
0.1
0.15
0.2
0.25
Fig. 3. Effects of varying the regularization parameter on the classication accuracy obtained by the SFBCSP method for subject B0103T. The parameter value
giving the highest accuracy was highlighted in a red circle. (For interpretation of the
references to color in this gure legend, the reader is referred to the web version of
this article.)
1
Xi,l XTi,l ,
Nl
Nl
l =
(1)
i=1
wT 1 w
wT 2 w
s.t. w2 = 1,
(2)
(3)
X.
Z=W
(4)
88
var(Zm )
2M
gm = log
h=1
(5)
var(Zh )
g1,1
.
..
...
g1,2MK
..
..
.
G=
gN,1
...
(6)
gN,2MK
where gi,j denotes the j-th feature extracted from EEG sample at
i-th trial, and N = N1 + N2 .
Here, we propose to use the following sparse regression model
(i.e, the well-known Lasso estimate (Tibshirani, 1996)) for signicant CSP feature selection
1
u = arg min Gu y22 + u1 ,
2
u
(7)
N
i=1
(j)
(j)
gi,j (yi y i ),
(8)
where y i =
g u is the tted value excluding the contribud=
/ j i,d d
tion from gi,j , and S(a, ) is a shrinkage-thresholding operator:
a if a >
0
if |a|
sign(a)(|a| )+ =
a
+
if a < .
(9)
satisfying (7)
We can obtain the optimized sparse vector u
through repeating the update form in (9) for j = 1, 2, . . . , D, 1, 2,
89
Fig. 4. Sparse vectors and the most signicant spatial lters learned by the SFBCSP method for each of the ve subjects in BCI Competition III dataset IVa. The feature indicated
as 1, 2, . . ., 34 correspond to the lter subbands 48 Hz, 610 Hz, . . ., 3640 Hz, respectively.
B0103T
B0203T
B0303T
B0403T
B0503T
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0
5 10 15 20 25 30
0
5 10 15 20 25 30
B0603T
0
5 10 15 20 25 30
B0703T
B0803T
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0
5 10 15 20 25 30
0
5 10 15 20 25 30
Feature index
0
5 10 15 20 25 30
5 10 15 20 25 30
B0903T
0
5 10 15 20 25 30
Values of u
5 10 15 20 25 30
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
lter subbands 48 Hz, 610 Hz, . . ., 3640 Hz, respectively.
effectively. In this study, the optimal subject-specic was determined by cross-validation on training data. Fig. 3 presents an
example to see the effects of varying on the classication accuracy for subject B0103T. The optimal validation accuracy was
achieved at = 0.14 that was chosen for the subsequent test procedure. A potential problem for the proposed method is that
the cross-validation procedure is usually time-consuming and is
not applicable for small sample size datasets (Lu et al., 2010).
Instead of cross-validation, Bayesian inference provides an effective
approach to automatically and quickly estimate the model parameters under the so-called evidence framework (MacKay, 1992;
Bishop, 2006). Recently, sparse Bayesian learning (Tipping, 2001)
has been increasingly applied to the automatic determination regularization in EEG analysis (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 efciency 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)
90
FBCSP
Value of the best feature 2
0
1
2
3
4
3
2
1
0
Value of the best feature 1
DFBCSP
CSP
1
0
1
2
3
4
3
2
1
0
Value of the best feature 1
1
0
1
2
3
4
3
2
1
0
Value of the best feature 1
SFBCSP
2
1
0
Value of the best feature 1
1
0
1
2
3
4
3
Fig. 6. Distributions of the most signicant 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 lter band optimization to further improving the
MI classication accuracy. With a properly determined regularization parameter , the optimal sparse vector u can be learned
from (7) to effectively select the signicant CSP features extracted
from multiple lter subbands. Fig. 4 presents the sparse vectors
and the most signicant spatial lters 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 IV
dataset IIb. It can be seen that the distribution of signicant lter
bands is sparse and subject-specic. The spatial lter corresponding to the most signicant band accurately captured the features in
the sensorimotor areas. The signicant feature appeared at multiple 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 lter band for CSP is necessary to improve
the MI classication accuracy. Fig. 6 depicts distributions of the
most signicant two features derived by CSP, FBCSP, DFBCSP and
SFBCSP, respectively, from subject B0703T. All of the three modied 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 shown
their 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 lter
band selection could further benet to improve the effectiveness
of CSP for MI-based BCI.
5. Conclusions
In this study, we introduced a sparse lter band common spatial
pattern (SFBCSP) algorithm to automatically select the signicant
lter bands for improving MI task classication performance in
BCI. In the proposed method, CSP features were rst estimated on
91