You are on page 1of 5

SICE Annual Conference 2011

September 13-18, 2011, Waseda University, Tokyo, Japan

A VEP-based Spelling System with Sequential Stimulus using Binary BCH


Codes
Takashi Mizoguchi, Jun Irie, Tomonari Yamaguchi and Katsuhiro Inoue
Department of Information Technology, Kyushu Institute of Technology, Fukuoka, Japan
(Tel: +81-948-29-7743; E-mail: mizoguchi@kiri.ces.kyutech.ac.jp, inoue@ces.kyutech.ac.jp)

Abstract: It is known that electroencephalograph (EEG) signals show specic responses with a event; e.g., visual
stimulus, cognition or motor imagery. Especially users can control an electronic device by classifying and interpreting the
EEG signals; such a system is called brain-computer interface (BCI). In this study, we tried to construct the BCI system,
based on transient visual evoked potentials (VEPs) and its visual stimuli ash in conformity with the sequence of binary
Bose-Chaudhuri-Hocquenghem (BCH) codes.

Keywords: BCI, visual evoked potentials, BCH codes, Coding theory.

1. INTRODUCTION the background of a nature picture. Users can choose a


target by just gazing what users want; e.g., handicapped
Brain-computer interface (BCI) allows users to com- user can open a door with watching the ashing door on
municate by classifying and translating neurophysiologic the screen.
signals into inputs for applications; such as a spelling de- We proposed a system that multiple stimuli turn on and
vice, wheelchair or prosthesis[1]. BCIs which use a P300 off respectively and simultaneously, not using all pixels
event-related potential (ERP) or visual evoked potentials on the screen[7][8]. A transient-VEP was introduced to
(VEPs) of the electroencephalogram (EEG) have been realize an input with looking at a ashing alphabet let-
much studying recently. In other words, BCIs are ex- ter. Besides that, with a view to decreasing discomfort
pected as a communication tool for those who have amy- of SSVEP, non-consecutive and random stimulation pat-
otrophic lateral sclerosis (ALS) or locked-in syndrome. terns were generated by Bose-Chaudhuri-Hocquenghem
Since the use of P300-based BCI for communication (BCH) codes. The BCH codes are one of the highest er-
was rst introduced by Farwell and Donchin[2], P300- ror correcting codes in a capability of rectifying and can
BCI has been revealed high accuracy[3]. However P300- maximize the Hamming distance between every pair of
BCI requires users to continue gazing at a target character sequences[9]. The safer BCI system was constructed us-
attentively until the system nds it, users can not use the ing them. We made an experiment to conrm the repro-
system a long time. The reason is that some cognition duction of characteristic VEPs fundamentally. The goal
task is needed to evoke P300. of this paper was to detect the target letter which user
On the other hand, several VEP studies have sought would like to input by extracting VEPs and to investigate
to achieve valuable VEP-based BCIs[4]-[6]. It is known the effect of the stimulation patterns generated by BCH
that VEP is the potential difference evoked by visual codes on our spelling system.
stimulation and recorded from electrodes over a scalp
of the occipital lobe; i.e., the visual area. VEP is di- 2. EXPERIMENTAL PARADIGM
vided broadly into two categories; transient VEP by low- 2.1 Subjects
frequency stimulus and steady-state visual evoked poten-
Five healthy subjects (S1, S2, S3, S4 and S5) partici-
tial (SSVEP) by high-frequency stimulus. SSVEP needs
pated in this experiment. The average age was 25.6 years
users to gaze rapidly blinking ashes on a PC monitor and
old (SD=8.53, range= 21-40 years old). All subjects
can detect gazing target in a short time[4]. The other ad-
reported normal or corrected-to-normal vision. The sub-
vantage is that it does not require a cognition task, differ-
jects provided written informed consent and were paid for
ing markedly from P300-BCIs. However there is a danger
their efforts.
of bringing about the symptoms (a indisposition, dizzi-
ness or light epilepsy). Compared with SSVEP, transient 2.2 Stimuli
VEP has no danger of appearing such symptoms. The nine letters are displayed on a monitor which used
Recently, G. Bin et al. have developed high perfor- a black background in three kinds (1 1, 2 2 and 3 3
mance BCI system using code modulation VEP, where matrix); see Fig. 1 for a 3 3 character matrix. , Those
binary pseudorandom sequences were used to modulate letters are corresponded with Latin alphabet [A... I] and
different visual stimuli[6]. The system is very effective henceforth we use the sight stimulation instead of letter.
as a spelling system and designed to set 32 stimuli over a During the experiment, the sight stimulation ashes in
whole PC monitor. It is a weak point of the study that the conformity to binary sequence decided by binary (15, 5,
system needs to use a full screen. That is, a background 7) BCH codes. Stimuli sequences consist of 15 bits of
image is not considered by the interface. For instance, binary {0,1}. The kinds of stimulation sequences used in
there are some alternatives sparsely on the screen having this experiment are nine which is equivalent with from A

- 1603 - PR0001/11/0000-1603 400 2011 SICE


to I. In case of 1, the sight stimulation ashes on. If bit 3. METHODS
0, it does not ash. The sight stimulation turns on for
50 ms and off for 50 ms. Average signal is one of the most useful techniques
to extract VEPs from EEG signals and the thing which
2.3 EEG acquisition and Experimental procedure a segment of EEG data taken out at trigger time is aver-
EEG was measured using an electrode cap with 17 aged. In this session, the following processing method
Ag/AgCl electrodes following the extended international is based on average signal. The EEG data of a sin-
10-20 system; see Fig. 2 for recorded electrodes place- gle trial, 60 sequences and 15 bits, was divided before-
ment. The non-invasive EEG signals were amplied hand into a training dataset and a test dataset in a ra-
with g.USBamp (24 Bit biosignal amplication unit, g.tec tio of two to one on the basis of 3-folds cross valida-
Medical Engineering GmbH, Austria) and digitized at a tion. Using the training data obtained with average sig-
sampling rate of 512 Hz. In addition, a trigger channel nal of original EEG data equal to 40 sequences, we gen-
was also recorded to synchronize EEG data with visual erated three types of templates ({0,1}, {00,01,10,11}
stimulus. and {000,001,010,011,100,101,110,111}). A template
Each subjects task was to be seated in a comfort- matching technique was used after generating these tem-
able armchair about 1m away from the monitor (Pro- plates. Let us, for the moment, call them k-bits tem-
Lite E2210HDS-B, MouseComputer Co.,Ltd., Japan) in plate. Then 2k templates are obtained; e.g., two template
shielded room. The task needs subjects to focus attention are generated as 0 and 1 respectively when k = 1. In
on instructed sight stimulation for a minute. Hereinafter this paper, we employed k = 1, 2 and 3 and the data
the instructed sight stimulation is referenced to as the tar- length of every template M is 102[point] (equivalent to
get. For ve subjects, we carried out 18 trials with 1 1 200 ms), 154[point] (300 ms) and 205[point] (400 ms) in
matrix, 8 trials with 2 2 matrix and 18 trials with 3 3 practice.
matrix. These 44 trials per a day are dened as one ses- The acquired EEG data are denoted as xn (t), n =
sion and the total becomes three sessions. 1, , N and N is a number of acquired data. Let
averaged signals with k-bits template denote y(t) =
[y1 (t), , y2k (t)] (i = 1, , 2k ). Now dene yi (t):
N
1
yi (t) = xn (t), i = 1, , 2k (1)
N n=1

where i denote ith index for k-bits template and xn (t) is


M N.
Next, we examine the correlation coefcient vector
T
P = [1 , , 2k ] between x(t) and the generated tem-
plate yi (t) for a segment of EEG data.
M

yi (t) x(t)
Fig. 1 Example of stimulus screen, sight stimulation i = t=1
(2)
matrix is 3 3; we set up a illuminometer at the M

M

lower left to obtain an exact trigger time. yi (t) yi (t) x(t) x(t)
t=1 t=1

where i = 1, , 2k .
Then, i represents index of the templates bits for
which maximizes the correlation coefcient among all of
templates. The correlation is calculated until restore bi-
nary sequences for all test data.
i = arg max i (3)
i

where i stands for estimated index and gives k-bits tem-


plates i th bits.
Furthermore, by the minimum Hamming distance de-
coder, the output bits are corrected to some binary se-
quences which had decided in advance and the estimated
target can be obtained. Now the test dataset is separated
into four segments which consist of the ve estimated tar-
gets. The processing length was 7.5 second since one
Fig. 2 Electrode positions and labels in the 10-20 sys-
cycle was equivalent to 1.5 second. Finally, we can rec-
tem.
ognize the target which a user stared at with a decision

- 1604 -
by majority of the estimated ve. The outline of this al- repeated stimuli at rapidly ashing of 5 Hz and over[10].
gorithm is shown at Fig. 3. Also the bit error rate (BER) In our experimental system, the stimuli of this segment
dened as dij /15 was used for evaluating output binary were ashing at 10 Hz. The frequency of 10 Hz satis-
sequences (where dij is the Hamming distance between ed with a condition of evoking SSVEP. Then the bit fol-
sequences i and j). lowing the four consecutive ashes was investigated by
the template matching. Table 1 shows the error rate of
Outline of algorithm which the output bit, an original bit is 0, is judged as
Input: The single trial (corresponding to 60 sequences 1 using 1-bit template. From Table 1, it was obvious
and divided into the training and the test), that the four consecutive ashes inuenced S1. Looking
a priori information of used binary sequences, toward the other subjects, it was conrmed that the error
and the number of bits k. rate showed a different tendency. The tendency should be
Output: The target character examined thoroughly. We will refer to the BER of whole
Initialization: Give a signicance to non-ash or ash bits at following section.
with {0,1}
Method:
1. Generate 2k bits template obtained by the training
data.
2. For iteration number c =1:number of the test data
Examine the test with 2 templates using cor-
k

relation coefcient.
Decide temporal binary sequence conforming
to bits of templates similar to the test.
Correct the temporal bits to the nearest binary
sequence by the Hamming distance.
3. Judge the target by a majority decision.
Fig. 3 Algorithm of detecting the target.

In addition, the occipital region of Oz, known to be


most inuenced by visual information, was adopted as
analyzed channels, and we analyzed the bipolar deriva-
tion with Oz-Pz.
Fig. 4 Superimposed averaged waveforms of subject S1
4. RESULTS AND DISCUSSIONS to every number of the sight stimulation; for stimula-
4.1 Response waveform including VEPs tion pattern of A.
The superimposition of averaged waveform shows al-
Table 1 Error rate of which the output bit is judged as
most similar shape to every number of the sight stimula-
1 using 1-bit template for all subjects in 2 2; the bit
tion; see Fig. 4 for subject S1. However, the electroneg-
error rate is shown with parentheses and the values are
ative wave of nearby 100 ms became unstable. The am-
calculated as a percentage.
plitudes which appeared in a single light source (1 1)
and did not come out in multiple light sources (3 3) Subject Session1 Session2 Session3
give the example. In time of a single light source, the S1 56.7 (41.8) 56.7 (43.1) 65.4 (43.1)
potential difference occurs with P75, N100 and P145, the S2 45.0 (46.7) 45.8 (45.5) 44.2 (46.5)
latencies are 75 ms, 100 ms and 145 ms respectively after S3 50.0 (45.3) 63.3 (45.3) 44.2 (45.8)
stimulus. When multiple light sources, we had already S4 52.5 (39.3) 39.2 (43.0) 30.8 (43.1)
conrmed it difcult to nd N100 in VEPs between stim- S5 41.7 (47.0) 50.8 (47.5) 51.7 (48.3)
ulus and non-stimulus at our previous study[8]. The rea- [%]
son of this phenomenon is that the neighbors of target
inuenced VEPs component. Regarding our experiment,
some subjects reported their steadily gaze was fairly hard 4.2 Discrimination of target to sight stimulation
because the neighbors surrounding a target obstructed it Fig. 5 , Fig. 6 and Fig. 7 show that k = 1 can ob-
in multiple light sources. That has left an unresolved is- tain comparatively high accuracy in 2 2 and the ac-
sue which includes the way to gaze steadily at multiple curacy of k = 3 is the highest in 3 3. It is think-
light sources. able that the accuracy of 3-bits template was resulted
In the same study[8], it was conrmed that the SSVEP from a processed length of templates. The VEPs wave-
components were extracted using independent compo- form of multiple light sources became more complex than
nent analysis when the target ashed on the sequences that of a single light source as described above. In our
which had four consecutive stimuli part. The response method, the template can reect the inuence of mul-
waveform of SSVEP is a sine wave-like and evoked by tiple light sources since the template includes complex

- 1605 -
response signal of VEPs which had subsequent latency spite of being guaranteed the minimum Hamming dis-
when k is larger. The required number of average times tance may give a lower effect to the discrimination. In
for VEPs was from about a hundred to two hundred[10] this study, the BCH codes we selected can correct 3 error
and an adequate number of times could not be obtained bits within 15 bits[9]. As shown in Table 2, the binary
in a present method when k 4. Also the number of sequences will be corrected perfectly when the BER is
average times for {0,1} reached about from four hundred under 20%; i.e., we can detect the target completely if the
twenty to four hundred eighty when k = 1. It is the most error bits of one sequence are under 3 bits within 15 bits.
important that the necessary and sufcient number of av- In future work, a precise decision of binary sequences
erage times has been given on recording evoked poten- will be needed as well as the optimal codes besides rear-
tials. There is a strong possibility that the result of aver- ranging above factors properly.
age signal was obscure using the large number of average The required length of detecting a target was 7.5 sec-
signal times; thus, it may bring the decline of accuracy. ond although a majority decision should be investigated
From the above, we can draw the following discussion: how many majorities are appropriate on our method in
it is necessary that users would receive the adjustment of detail. Mainly calculation was only correlation between
stimulation patterns and a necessary and sufcient train- an original signal and templates. We could make special
ing data should be obtained in advance as P300-based mention of obtaining a sort of results with the method
spelling device considering a practical interface. which is a simple calculation.

Table 2 Bit error rate of all subjects for all of bit


template in 2 2 and 3 3; calculated as a percentage.

22 33
Subject 1 bit 2 bits 3 bits 1 bit 2 bits 3 bits
S1 46.2 42.4 41.2 43.7 43.9 41.0
S2 46.2 42.4 44.8 47.8 46.8 46.5
Fig. 5 The result of {0,1} pattern (k = 1) in 2 2 and S3 45.2 46.5 42.0 45.8 46.0 43.9
3 3, all subjects. S4 41.8 44.1 40.0 42.5 42.2 39.5
S5 47.6 47.6 46.8 47.3 47.6 46.9
[%]

5. CONCLUSIONS
In this study, we tried to construct the EEG typewriter
system based on VEPs by applying template matching.
Fig. 6 The result of {00,01,10,11} pattern (k = 2) in The system was designed to stimulate users in conformity
2 2 and 3 3, all subjects. with binary sequences by BCH codes.
There seems to be a similarity in each of averaged
waveform although the discrepancy of shape of VEPs ex-
ists. Thus the characteristic VEPs were extracted in our
system. Also as for multiple sight stimulations, the prob-
lems of the inuence from other sight stimulations were
found.
By calculation of correlation between templates and
processed signal, it was conrmed that subjects could
choose the target with up to about 74% of recognition ac-
Fig. 7 The result of {000, 001, 010, 011, 100, 101, 110,
curacy in four classes and about 59% in nine classes using
111} pattern (k = 3) in 2 2 and 3 3, all subjects.
the processed length of 7.5 second. The output binary
Table 2 shows the BER for all subjects when all of sequences were inaccurate because the BER showed a
bit templates were used. From Table 2 the BER of 3- higher error rate than 20%. We will investigate the appro-
bits template is the smallest. It means that the output bi- priate method of a precise decision of binary sequences,
nary sequences are the most accurate; i.e., 3-bits template which will be able to output more accurate bits. It also
can restore the sequence which is obtained by our method requires consideration from BCH codes. More practical
to accurate one, compared with the other template. Em- BCI system using VEP will be constructed by developing
ploying the sequences which resemble the other ones in this research further.

- 1606 -
REFERENCES
[1] J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G.
Pfurtscheller, and T. M. Vaughan, Brain-computer
interfaces for communication and control, Clin.
Neurophysiol., vol. 113, pp.767-791, 2002.
[2] L. A. Farwell and E. Donchin, Talking off the top
of your head: toward a mental prosthesis utilizing
event-related brain potentials, Electroenceph. Clin.
Neurophysiol., vol. 70, no. 6, 1988.
[3] C. Guger, S. Daban, E. Sellers, C. Holzner, G.
Krausz, R. Carabalona, F. Gramatica, and G.
Edlinger, How many people are able to control a
P300-based brain-computer interface (BCI)?, Neu-
rosci. Letters, vol. 462, pp.94-98, 2009.
[4] P. Martinez, H. Bakardjian, and A. Cichocki,
Fully online multicommand brain-computer in-
terface with visual neurofeedback using SSVEP
paradigm, J. Comput. Intell. Neurosci., 2007.
[5] Y. Wang, R. Wang, X. Gao, B. Hong, and S. Gao,
A practical VEP-based brain-computer interface,
IEEE Trans. Neural Syst. Rehab. Eng., vol. 14, no.
2, 2006.
[6] G. Bin, X. Gao, Y. Wang, Y. Li, B. Hong, and S.
Gao, A high-speed BCI based on code modulation
VEP, J. Neural Eng., vol. 8, no. 2, 2011.
[7] J. Irie, T. Yamaguchi, K. Omori, and K. Inoue,
Feature extraction of visual evoked potentials us-
ing state-space model, SICE Annu. Conf. 2010, pp.
5457, 2010.
[8] T. Yamaguchi, J. Irie, M. Fujio, and K. Inoue, De-
sign method of morphological lter fo EEG signal
analysis in brain computer interface, IIC-11, pp.21-
26, 2011.
[9] H. Miyagawa, Y. Iwatare, and H. Imai, Coding
Theory, Shokodo, 1973.
[10] T. Takahashi, and Y. Kuroiwa, Vision and Clinical
EEG, Shinkoh-igaku,1995.

- 1607 -

You might also like