The combination of CCA and PSDA detection methods in a SSVEP-BCI system

Abstract

In recent years, based on the steady-state visual evoked potential (SSVEP) brain-computer interfaces (BCIs) have generated significant interest, due to their shorter calibration times and higher information transfer rates. Target identification is the core signal processing task in BCIs. Power spectral density analysis (PSDA) and canonical correlation analysis (CCA) are the most popular and widely used classification methods in SSVEP-BCI systems. In this paper, we first combined these two methods for detecting the SSVEP signals. Moreover, we compared the proposed method with PSDA, CCA method, respectively. The results showed that the proposed method can improve the accuracy and the transfer rate of BCIs.

Full text

The Combination of CCA and PSDA Detection Methods in a
SSVEP-BCI System
Ruimin Wang1,WenWu
1, Keiji Iramina2, Sheng Ge1,3
1School of Electronic Engineering and Optoelectronic Technology, Nanjing University of Science and Technology,
Nanjing, China
2Graduate School of Systems Life Sciences, Kyushu University, Fukuoka, Japan
3Key Laboratory of Child Development and Learning Science of Ministry of Education, Research Center for
Learning Science, Southeast University, Nanjing, China
shengge@seu.edu.cn
Abstract— In recent years, based on the steady-state visual
evoked potential (SSVEP) brain-computer interfaces (BCIs)
have generated significant interest, due to their shorter cal-
ibration times and higher information transfer rates. Target
identification is the core signal processing task in BCIs. Power
spectral density analysis (PSDA) and canonical correlation
analysis (CCA) are the most popular and widely used classifi-
cation methods in SSVEP-BCI systems. In this paper, we first
combined these two methods for detecting the SSVEP signals.
Moreover, we compared the proposed method with PSDA, CCA
method, respectively. The results showed that the proposed
method can improve the accuracy and the transfer rate of BCIs.
Index Terms— Brain-computer interface (BCI), steady-state
visual evoked potential (SSVEP), canonical correlation analysis
(CCA), power spectral density analysis (PSDA)
I. INTRODUCTION
Brain-computer interfaces (BCIs) support direct commu-
nication and control between brain and external devices
without any use of peripheral nerves and muscles [1]. In
recent years, based on the steady-state visual evoked potential
(SSVEP) BCIs have generated significant interest, due to
their shorter calibration times and higher information transfer
rates. It has been proved that when people gazing at a visual
stimulus modulated at a frequency higher than 6 Hz, the
visual cortex will generate oscillatory response, which has
the same fundamental frequency as the driving stimulus [2].
A practical BCI system should provide a sufficiently
high transfer rate, which relies on the number of selectable
targets, the speed and accuracy of target identification. Target
identification is the core signal processing task in BCIs.
There are many researches about SSVEP detection meth-
ods. The most popular and widely used methods are power
spectral density analysis (PSDA) and canonical correlation
analysis (CCA) [2-5]. PSDA is estimated from the users
EEG signal using a fast Fourier transform within a time
window. The frequency corresponding to the peak of PSD
is taken as the visual stimulus frequency. CCA is a relatively
new approach that first applied in SSVEP based BCIs is
at 2007 [3]. CCA is a multivariable statistical method for
measuring the maximized correlation between two sets of
data. One set of data is EEG signals from multiple channels.
Another set of data is reference signals depending on the
stimulus frequency. EEG signals are used to calculate the
CCA coefficients with all reference signals. The frequency
with the largest coefficient is taken as the visual stimulus
frequency. CCA had been described by Lin et al. [3] and Bin
et al. [4].
Lin et al. [3] and Hakvoort et al. [5] had compared CCA
and PSDA in SSVEP BCIs. Experimental results of Lin et
al. on multiple subjects showed that CCA-based detection
method achieved higher recognition accuracy than PSDA
approach [3]. Hakvoort et al. concluded that when the use
of harmonic frequencies is desired, the CCA-based detection
method is preferred over the PSDA-based detection method
[5].
CCA-based detection method more focuses on the correla-
tion between two sets of data, while PSDA method is mainly
to examine the PSD of signals. The classification results of
these two methods are related to each other in some degree,
but not completely agree. For future research it would be
interesting to combine CCA and PSD in SSVEP detection
[5]. In order to more effectively use the signal features, in
this research, we first combined these two detection methods
for classifying SSVEP signals.
II. METHODS
A. Combination Method of CCA with PSDA
In this research, to use both the correlations between EEG
signals and references and PSD features of EEG signals, we
proposed using the PSDA’s classification results to correct
the CCA’s for improving the accuracy.
First, the CCA-based method is used to classify the brain
patterns. After that the PSD values are calculated. If the
EEG signals have the maximum correlation coefficients with
one of the visual stimuli, but the corresponding PSD value
is smallest, the classification result will be invalid, none
result will be output. Otherwise, the classification result will
be execute. In brief, The CCA classification results with
978-1-4799-5825-2/14/$31.00 ©2014 IEEE
Proceeding of the 11th World Congress on Intelligent Control and Automation
Shenyang, China, June 29 - July 4 2014
2424

minimum PSD value will be invalid.
For example, there are four types of stimuli A, B, C, and
D. If the target stimuli is A, there are 4 possible outcomes
(Table 1):
1) The CCA classified result is A, and the corresponding
PSD is in the top 3 values, the classification result is A.
2) The CCA classified result is A, but the corresponding
PSD is smallest, the classification result will be invalid, none
result will be output.
3) The CCA classified result is incorrect B (or C, D),
the corresponding PSD is in the top 3 values, the wrong
classification result B (or C, D) will be output.
4) The CCA classified result is incorrect incorrect B (or
C, D), the corresponding PSD is smallest, the classification
result will be invalid, none result will be output.
The worst situation is the No.2 outcome, but its obviously
that the probability is very low.
TAB LE I
TYPE SIZE FOR PAPERS FOUR POSSIBLE OUTCOMES
(TARGET STIMULUS IS A)
NO. CCA Classification Result PSD Value Sort Outcome
1 A Top 3 Maximum A
2 A Minimum NONE
3B(orC,D) Top 3 Maximum B(orC,D)
4B(orC,D) Minimum NONE
There are four classes of stimuli. The CCA classification results with
minimum PSD value will be invalid.
B. Experiment Design
Four arrows flickering at different frequencies were dis-
played on a LCD monitor (screen refresh rate 60 Hz) as
visual stimuli (Fig. 1). In order to produce stable frequencies,
the frequencies should be chosen as divisors of the standard
refresh rate [6].
Fig. 1. Visual stimuli.
Lee et al. [7] confirmed that using the high duty-cycle
flickers in SSVEP-based BCI can not only improve the
visual comfort of flickers, but can also retain high SSVEP
amplitude.
Based on the above reasons, 7.5, 8.57, 10, and 12 Hz were
selected as flicker frequencies. The flicker duty cycles were
set at 87.5, 85.7, 83.3, and 80%, respectively.
Four healthy subjects (one female and three male) with
an average age of 23 years participated in this study. All
subjects had normal or corrected to normal vision and had
no experiences on BCIs. During the experiments, the subjects
were seated comfortably in a chair approximately 60 cm in
front of the LCD monitor. The visual angles of each arrow
were 3.3 degrees in the length and 2.3 degrees in the width.
The EEG signals were recorded using a SynAmps2 (Neu-
roScan Inc.) system at a sampling rate of 1000 Hz and a
band pass filter between 0.1 and 100 Hz. Subjects wore a
64-channel electrode cap, the electrode placement followed
the standard 10-20 system. Three channels (O1, O2, and Oz)
at the area of occipital were selected as the data sources.
The reference was set as the linked mastoid electrode,
and the ground electrode was placed between FPz and
Fz. Electrooculography (EOG) was recorded to remove the
EOG artifacts. Vertical EOG was monitored with a bipolar
placed above and below the left eye. Horizontal EOG was
monitored with a bipolar left-to-right outer canthus. ICA
method is adopted to remove the EOG artifact. During the
EEG recording, electrode impedances were kept below 10
KΩ.
C. Experiment Protocol
Fig. 2 demonstrates the experiment protocol. Every trial
started with 1-second-long cue, the target stimulus. After the
cue, four arrows begin to flicker for 5 s. Each run included
40 trials (10 trials for each target). In total, three runs (120
trials) were recorded per subject. Between runs, there were
breaks of about 2 min.
Before the experiment, a short pre-experiment was carried
out in order to introduce the protocol to the subjects.
Fig. 2. Experimental protocol.
III. RESULTS
The information transfer rate (ITR, bits per minute) is
a standard measure of BCIs. It depends on the number of
2425

selectable targets, the speed and accuracy of target identifi-
cation. The ITR is equal to the bit rate Bmultiplied by speed
of selection S(trial per minute) [8], it can be expressed as
equation (1).
ITR =B∗S. (1)
The bit rate Bcan be expressed as equation (2)
B=log2N+Plog
2P+(1−P)log2
(1 −P)
(N−1).(2)
where, Nis the number of possible selections, Pdenotes the
probability that the desired choice is actually selected.
If using the CCA or PSDA method, there are no trials will
be abandoned, so the speed of selection
S=60
T.(3)
where, Tdenotes a trial time (s).
However, when using the proposed combination method, in
some situations, the classification results will be abandoned,
then
S=60
T
∗(1 −Pc).(4)
Pcdenotes the probability of invalid trials.
Fig. 3 and Fig. 4 depict the SSVEP recognition accuracies
for each of the 4 subjects and the averaged accuracy obtained
by the CCA, PSDA, and the combination method of CCA
with PSDA at various time window lengths (TWs) from 1 to
5 s, respectively. Fig.5 and Fig.6 show the ITRs of the three
methods for all the subjects and the mean result, respectively.
It’s obvious that the combination method achieved higher
accuracies and higher ITRs than the CCA and PSDA when
the TW lengths are longer than 2 s (includes 2 s). Meanwhile,
our experiment also confirmed that the CCA-based detection
method performs significantly better than PSDA-based detec-
tion method.
IV. DISCUSSION
In this research, we proposed the combination method of
CCA with PSDA to recognize the stimulus frequency for
SSVEP-based BCI. In this method, the PSD values were used
to correct the CCA’s for improving the accuracy. Although
the higher accuracies were achieved at the expense of aban-
doning some trials, the ITRs were increased (Fig.5, Fig.6).
We can conclude that the proposed method is a promising
way of enhancing the performance of SSVEP-based BCIs.
11.5 2 2.5 3 3.5 4 4.5 5
20
25
30
35
40
45
50
55
60
65
70
S1
TW (s)
Accuracy (%)
1 1.5 2 2.5 3 3.5 4 4.5 5
20
25
30
35
40
45
50
55
60
65
70
S2
TW (s)
11.5 2 2.5 3 3.5 4 4.5 5
20
25
30
35
40
45
50
55
60
65
70
S3
TW (s)
Accuracy (%)
1 1.5 2 2.5 3 3.5 4 4.5 5
20
30
40
50
60
70
80
90
100
S4
TW (s)
CCA-based detection PSDA-based detection Combination of CCA & PSDA
Fig. 3. SSVEP recognition accuracies of the 4 subjects derived by the CCA,
PSDA, and the combination method of CCA with PSDA, respectively.
1 1.5 2 2.5 33.5 4 4.5 5
20
25
30
35
40
45
50
55
60
65
70
TW (s)
Accuracy (%)
CCA−based detection
PSDA−based detection
Combination of CCA & PSDA
Fig. 4. Averaged SSVEP recognition accuracy derived by the CCA, PSDA,
and the combination method of CCA with PSDA, respectively.
11.5 2 2.5 3 3.5 4 4.5 5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
S1
TW (s)
ITR (bit/min)
1 1.5 2 2.5 3 3.5 4 4.5 5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
S2
TW (s)
11.5 2 2.5 3 3.5 4 4.5 5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
S3
TW (s)
ITR (bit/min)
1 1.5 2 2.5 3 3.5 4 4.5 5
0
2
4
6
8
10
12
14
16
18
20
S4
TW (s)
CCA-based detection PSDA-based detection Combination of CCA & PSDA
Fig. 5. ITRs of the 4 subjects derived by the CCA, PSDA, and the
combination method of CCA with PSDA, respectively.
2426

1 1.5 2 2.5 3 3.5 4 4.5 5
0
1
2
3
4
5
6
TW (s)
ITR (bit/min)
CCA−based detection
PSDA−based detection
Combination of CCA & PSDA
Fig. 6. Averaged ITR derived by the CCA, PSDA, and the combination
method of CCA with PSDA, respectively.
The proposed combination method is using the CCA-
based method to classify the brain patterns, after that using
the PSDA-based method to decide whether the classification
results are valid or invalid. This combination method also
applies to joint other detection methods, such as SVM [9]
or LASSO [10]. But there is a drawback that it reduced
the speed of selection. Maybe there are better combination
methods without abandoning the doubtful results. Our next
research is how to combine SSVEP detection methods with-
out reducing the speed of selection.
An increase in the number of possible selections is another
effective approach to improve ITR [2]. In our experiment,
there were only 4 possible selections, compared with other
SSVEP-BCIs using more target selections [11], our transfer
rates need to improve. So, future studies involving more
subjects and more target selections may be required to
confirm the combination method of CCA with PSDA.
V. C ONCLUSION
In this paper, we first proposed the combination method of
CCA with PSDA to detect the SSVEP signals. Meanwhile we
designed a multi-channel SSVEP-based BCI system to exam-
ine the proposed method. The experiment results showed that
our proposed method achieved better accuracies and transfer
rates than CCA and PSDA. Moreover, our experiment also
confirmed that the CCA-based detection method performs
significantly better than PSDA-based detection method. The
proposed combination method can be widely used in the
SSVEP-based BCI system.
ACKNOWLEDGMENT
This work was supported by the National Natural Science
Foundation of China (No.51007040) and the Fundamental
Research Funds for the Central Universities, China (No.
3250183202).
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, no. 6, pp. 767-791, June 2002.
[2] Y. J. Wang, R. P. Wang, X. R. Gao, B. Hong, and S. K. Gao, A practical
VEP-based brain-computer interface, IEEE Trans. Neural Syst. Rehabil.
Eng., vol. 14, no. 2, pp. 234-239, June 2006.
[3] Z. L. Lin, C. S. Zhang, W. Wu, and X. R. Gao, Frequency recognition
based on canonical correlation analysis for SSVEP-based BCIs, IEEE
Trans. Biomed. Eng., vol. 54, no. 6, pp. 1172-1176, June 2007.
[4] G. Y. Bin, X. Y. Gao, Z. Yan, B. Hong, and S. K. Gao, An online
multi-channel SSVEP-based braincomputer interface using a canonical
correlation analysis method,” J. Neural Eng., vol. 6, no. 4, p. 046002,
June 2009.
[5] G. Hakvoort, B. Reuderink, and M. Obbink, Comparison of PSDA
and CCA detection methods in a SSVEP-based BCI-system, Technical
Report TR-CTIT-11-03, Centre for Telematics and Information Tech-
nology, University of Twente, Enschede, The Netherlands, 2011.
[6] I. Volosyak, H. Cecotti, and A. Graser, Impact of frequency selection on
LCD screens for SSVEP based brain-computer interfaces, in IWANN
2009, Part I, LNCS, vol. 5517, J. Cabestany, F. Sandoval, A. Prieto, and
J. M. Corchado, Eds. Heidelberg: Springer, 2009, pp. 706-713.
[7] P. L. Lee, C. Yeh, Y. S. Cheng, C. Y. Yang, and G.Y. Lan, An SSVEP-
based BCI using high duty-cycle visual flicker, IEEE Trans. Biomed.
Eng., vol. 58, no. 12, pp. 3350-3359, December 2011.
[8]X.R.Gao,D.F.Xu,M.Cheng,andS.K.Gao,ABCI-based
environmental controller for the motion-disabled, IEEE Trans. Neural
Syst. Rehabil. Eng., vol. 11, no. 2, pp. 137-140, June 2003.
[9] S.N. Resalat, V. Saba, F. Afdideh, and A. Heidarnejad, High-speed
SSVEP-based BCI: Study of various frequency pairs and inter-sources
distances, IEEE-EMBS Int. Conf. Biomedical and Health Informatics,
pp. 220-223, Jan 2012.
[10] Y. Zhang, J. Jin, X. Y. Qing, B. Wang, and X. Y. Wang, LASSO based
stimulus frequency recognition model for SSVEP BCIs, Biomed. Signal
Process. Control, vol. 7, no. 2, pp. 104-111, March 2012,
[11] Y. Wang, Y.T. Wang, and T.P. Jung, Visual stimulus design for high-
rate SSVEP BCI, Electron. Lett., vol. 46, no. 15, pp. 1057-1058, July
2010.
2427