SSVEP-based BCI Classification Using Power Cepstrum Analysis

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 2

SSVEP-based BCI classification using power

cepstrum analysis
Yeou-Jiunn Chen, Aaron Raymond Ang See and
Shih-Chung Chen
The power cepstrum-based parameters for steady-state visually evoked
potential (SSVEP) is proposed. To precisely represent the characteristics of frequency responses of a visually stimulated electroencephalography (EEG) signal, power cepstrum analysis is adopted to estimate
the parameters in low-dimensional space. To represent the frequency
responses of SSVEP, the log-magnitude spectrum of an EEG signal
is estimated by fast Fourier transform. Subsequently, the discrete
cosine transform is applied to linearly transform the log-magnitude
spectrum into the cepstrum domain, and then generate a set of coefcients. Finally, a Bayesian decision model with a Gaussian mixture
model is adopted to classify the responses of SSVEP. The experimental
results demonstrated that the proposed approach was able to improve
performance compared with previous approaches and was suitable
for use in brain computer interface applications.

Introduction: Motor neuron disease weakens the muscles that consummates into loss of voice and voluntary controls of the patients various
limbs, hence detaching them from the outside world [1]. To address
the situation, the augmentative and alternative communication (AAC)
system could actually help patients with neuromuscular impairments
to communicate with other people or devices by sending messages or
commands. Thus, various types of interfaces based on voices or gestures
have been widely used in AAC systems [2, 3]. However, these interfaces
were not suitable for neuromuscular impairments. Recently, the brain
computer interfaces (BCIs) are used to design communication systems
by using electroencephalography (EEG) analysis [47]. It sends messages or commands to the external world through direct measurement
of brain activity, thus it is a suitable interface for neuromuscular impairments. Therefore, to develop a suitable BCI is benecial for patients
with neuromuscular impairments so as to communicate with other
people or devices.
For BCI approaches, to determine the users response, the steady-state
visually evoked potential (SSVEP)-based BCI applies visual stimulus at
specic frequencies and then the electrical activity of the brain is
measured [4]. SSVEPs are advantageous in BCIs because of their excellent signal-to-noise ratio and relative immunity to artefacts. Therefore, to
achieve the best performance of BCIs, many researchers have focused
on selecting stimulator devices and stimuli properties [5]. However, in
practice, the frequency response of a user not only contains many
noises but also a reduction in the level of measured electrical activity.
The accuracy would be greatly reduced, thus to develop a sensitive
parameter would improve the performance of the SSVEP-based BCI.
To actually classify the frequency responses, the coherence and nonstationary signal analysis methods have been used in BCIs [6]. These
approaches are computationally complex and cannot be applied to realtime systems. Another approach, the thresholding method, uses the spectral magnitude estimated by fast Fourier transform (FFT) to classify the
frequency responses [7]. As the threshold is dependent on specic users,
it is difcult to apply to other users. Furthermore, the frequency
responses of a user are usually not only noisy but also lower than the
threshold. Thus, it would greatly reduce the accuracy of the systems
in real-time applications.
To efciently reduce computational time, the parameter dimensions
should be reduced without losing the precision of representing the
characteristics for frequency responses. For data compression, the discrete cosine transform (DCT) had been successfully used to compress
EEG data [8]. DCT then could efciently represent the characteristics
of signals in low-dimensional space. In addition, the cepstrum coefcient derived from the log-magnitude spectrum has been widely used
in speech technologies [9]. It can preserve most information in its spectral envelope. As the log-magnitude spectrum contains the major information of frequency responses, the integrating DCT and cepstrum
would be able to effectively represent the signicant information of
the log-magnitude spectrum in low-dimensional space.
In this Letter, a SSVEP-based BCI using a power cepstrum is proposed to precisely represent the characteristics of frequency responses
in low-dimensional space. To precisely represent the characteristics of
the frequency responses, the power cepstrum analysis integrating FFT
and DCT is proposed to derive coefcients in low-dimensional space.

Thus, it could preserve the accuracy of BCIs and be applied to real-time


systems. The Bayesian decision model with the Gaussian mixture model
(GMM) is adopted to classify frequency responses by using cepstrum
coefcients.
Power cepstrum analysis: To precisely represent the characteristics of
frequency responses in low-dimensional space, the power cepstrum
analysis is proposed in this Letter. For an input frequency response
with N samples, x(n), the FFT is adopted to estimate the log magnitude
of spectrum X(k) and is dened as


N
1


X (k) = log


x(n)ei2pk(n/N )

(1)

n=0

According to the frequencies of visual stimulations, a frequency


band limited log magnitude of the spectrum is extracted from X(k)
and denoted as X (m). When the responses are accurately stimulated,
a peak envelop should be produced at various frequencies.
Subsequently, DCT is applied to linearly transform X (m) into the cepstrum domain and form the cepstrum coefcients C, which can be
derived as
C(k) =

 
 
p
1
m+ k
X (m) cos
2
M
m=1

M
1


(2)

where M is the number of frequency bins in X (m).


Bayesian decision model: To accurately classify the frequency
responses, the Bayesian decision model is adopted in this Letter. For
a cepstrum coefcient C(k), a class of frequency responses k with greatest posterior probability can be found by using Bayes rule

vk = arg max P(vi |C )


vi

= arg max
vi

P (v i )
P(C |vi )
P(C)

(3)

As the denominator P(C ) is a constant term for all classes and the prior
probability P(i) is treated as a uniform distribution, (3) can be rewritten
as

vk = arg max P(C |vi )


vi

(4)

To precisely represent the distributions, the conditional probability


P(C|i) is modelled by the GMM with J mixtures and can be derived as
P(C |vi ) =

J





wj N C mj , sj

(5)

j=1

where wj, j and j are the mixture weight, mean and the standard
variance associated with the jth Gaussian
J component. wj is subject to
the following constraint: wj 0 and
j=1 wj = 1. These parameters
of the GMMs are estimated by the expectation maximisation
algorithm [10].
Experimental results: To evaluate the proposed approach, ve subjects
(four males and one female) aged between 21 and 23 participated in this
study. The frequency responses were stimulated using a liquid crystal
display screen and the visual stimulator (as shown in Fig. 1) ickering
at ve frequencies from 6 to 10 Hz with 1 Hz increments. Then, the
EEG signals were measured using electrodes placed at the Oz, A1 and
A2 (ground) in accordance with the international EEG 10-20 system.
The sampling rate and the frame size were set to be 1 kHz and 1 s,
respectively. Subsequently, using 4096 points FFT, the log-magnitude
spectrum can be estimated and band limited from 5 to 11 Hz. The
number of cepstrum coefcients used range from 6 to 10 coefcients.
As each subject had 60 epochs for each specic frequency, the
number of mixtures in GMM was set to be two mixtures.

ELECTRONICS LETTERS 8th May 2014 Vol. 50 No. 10 pp. 735737

Conclusions: In this Letter, a power cepstrum analysis using DCT is


proposed to improve the efciency of SSVEP-based BCI classication.
DCT was successfully applied to linearly transform the log-magnitude
spectrum into the cepstrum domain in low-dimensional space. Hence,
the cepstrum could preserve most information of frequency responses
and then effectively reduce the computational time. According to the
experimental results, the proposed approach improved the performance
of the SSVEP-based BCI system and outperformed other approaches.
Moreover, it obtained high accuracy in low-dimensional space.
Therefore, the proposed approach is suitable for implementing the
SSVEP-based BCI system in a real-time platform. In future, the cepstrum coefcients can be integrated to provide the rejection mechanism
for epochs without messages or commands.

go forward

8 Hz

turn left

7 Hz

turn right

9 Hz
dive

elevate

10 Hz

6 Hz

Fig. 1 Multiple visual stimulators used to induce SSVEP response

In the study, a subject (male) was randomly selected as training data


and other subjects were treated as testing data. The detailed results with
different numbers of cepstrum coefcients are shown in Table 1. The
accuracy of 8 Hz (87%) was lower than that of other frequencies.
The reason is that the noise is high and the magnitude of the frequency
response is low in our database. However, the average accuracy was
92.58% with seven cepstrum coefcients, thus the cepstrum coefcients
can efciently represent the characteristics of frequency responses in
low-dimensional space. Hence, the computational time can be efciently
reduced. The experimental results demonstrated that the proposed
approach would be benecial for real-time-based BCI applications.

Table 1: Accuracies (%) of proposed approach


Frequency (Hz)
6
7
8
9
10
Average

95.42
91.67
87.08
97.50
91.25

95.00
91.67
86.67
97.08
92.08

95.00
92.92
88.33
93.75
92.50

95.42
94.17
85.83
95.00
92.50

90.75 92.58 92.50 92.50 92.58

To objectively compare the proposed approach, the log-magnitude


spectrum and the thresholding methods were selected. Subsequently,
the proposed approach used seven cepstrum coefcients. The experimental results are shown in Fig. 2 and the proposed approach exhibited
the best performance. When the induced responses were not clear in our
database, the proposed approach still can preserve acceptable performance. Furthermore, the error reduction rates were 45.80 and 25.00%
for 6 and 9 Hz, respectively. Thus, the proposed approach had the greatest improvement, when the induced responses were clear enough.
Consequently, the experimental results demonstrated that the proposed
approach is useful for AAC systems.
100

accuracy, %

95
90
85
80
proposed approach
log magnitude spectrum

75

thresholding method
70

8
frequency, Hz

The Institution of Engineering and Technology 2014


19 January 2014
doi: 10.1049/el.2014.0173
One or more of the Figures in this Letter are available in colour online.
Yeou-Jiunn Chen, Aaron Raymond Ang See and Shih-Chung Chen
(Department of Electrical Engineering, Southern Taiwan University of
Science and Technology, No. 1, Nan-Tai Street, Yungkang District,
Tainan 710, Taiwan)
E-mail: chung@mail.stust.edu.tw
References

Number of cepstrum coefcients


6
7
8
9
10
95.83
89.17
87.08
96.25
85.42

Acknowledgment: This work was supported by the National Science


Council, Taiwan, under grant numbers NSC100-2632-E-218-001MY3 and NSC 102-2221-E-218-001.

1 Usanova, L.D., Usanova, A.D., and Skripal, A.V.: Analysis of effect of


audiovisual stimulation on parameters of encephalogram and rate of
pulsation wave propagation in humans, Biomed. Eng., 2012, 46, (1),
pp. 2528
2 Hawley, M.S., Cunningham, S.P., Green, P.D., Enderby, P., Palmer, R.,
Sehgal, S., and ONeill, P.: A voice-input voice-output communication
aid for people with severe speech impairment, IEEE Trans. Neural
Syst. Rehabil. Eng., 2013, 21, (1), pp. 2331
3 Fu, Y.F., and Ho, C.S.: A fast text-based communication system for
handicapped aphasiacs. Proc. 5th Int. Conf. Intelligent Information
Hiding and Multimedia Signal Processing, Kyoto, Japan, September
2009, pp. 583594
4 Zhu, D., Bieger, J., Garcia Molina, G., and Aarts, R.M.: A survey of
stimulation methods used in SSVEP-Based BCIs, Comput. Intel.
Neuro., 2010, 2010, Article ID 702357
5 Liu, Q., Chen, K., Ai, Q., and Xie, S.Q.: Review: recent development
of signal processing algorithms for SSVEP-based brain computer interfaces, J. Med. Biol. Eng., 2013, accepted for publication, doi: 10.5405/
jmbe.1522
6 Chen, S.C., See, A.R., Hou, C.J., Chen, Y.J., Liang, C.K., Hou, P.Y.,
and Lin, W.K.: Coherence validation of alternative sleep EEG electrode placements using wavelet transform, J. Med. Biol. Eng., 2013,
doi: 10.5405/jmbe.1511
7 Luo, A., and Sullivan, T.J.: A user-friendly SSVEP-based braincomputer interface using a time-domain classier, J. Neural Eng.,
2010, 7, (2), doi: 10.1088/17412560/7/2/026010
8 Zhao, H., and Malik, H.: Audio recording location identication using
acoustic environment signature, IEEE Trans. Inf. Forensics Sec., 2013,
8, (11), pp. 17461759
9 Chen, Y.J., and Wu, J.L.: An objective approach to identify spectral
distinctiveness for hearing impairment. Math. Probl. Eng., 2013,
Article ID 760258, 9p, doi: 10.1155/2013/760258
10 Chen, S.C., See, A.R., Chen, Y.J., Yeng, C.H., and Liang, C.K.: The
use of a brain computer interface remote control to navigate a recreational device, Math. Probl. Eng., 2013, Article ID 823736, 8p,
doi:10.1155/2013/823736

10

Fig. 2 Performance of different approaches

ELECTRONICS LETTERS 8th May 2014 Vol. 50 No. 10 pp. 735737

You might also like