You are on page 1of 5

International Journal of Signal Processing Systems Vol. 1, No.

1 June 2013

A Comparison of EEG Processing Methods to


Improve the Performance of BCI
Arjon Turnip, Demi Soetraprawata, and Dwi E. Kusumandari
Technical Implementation Unit for Instrumentation Development Indonesian Institute of Sciences, Bandung,
Indonesia Email: {arjo001, demi001, esti001}@lipi.go.id

Abstract—Electroencephalogram (EEG) recordings provide study, we focused on the use of the ERP-P300 properties.
an important means of brain-computer communication, but The P300 represents the unpredictable stimuli presented
their classification accuracy is limited by unforeseeable in an oddball paradigm, in which low-probability targets
signal variations due to artifacts or recognizer-subject are mixed with high-probability ones. For this paradigm,
feedback. In this paper, we propose a comparison of
the subject is told to respond to a rare stimulus that occurs
processing method (i.e., NPCA, JADE, and SOBI) entailing
time-series EEG signals. Finally, the promising results randomly and infrequently among other, frequent stimuli
reported here (up to 94% average classification accuracy [8]. The presence, magnitude, topography, and time of
and 36.4% improvement of maximum average transfer the response signal are often used as metrics of cognitive
rate) reflect the considerable potential of EEG for the function in decision making processes. In this paper, a
 comparison of extraction method (i.e., NPCA, JADE, and
continuous classification of mental states.
SOBI) entailing time-series EEG signals is proposed. In
Index Terms—brain computer interface (BCI), classification order to examine the performance (i.e., accuracy and
accuracy, transfer rate, NPCA, JADE, SOBI, transfer rate) improvements of the proposed method, a
electroencephalogram (EEG)
classification using back-propagation neural networks
(BPNN) which has been well developed in the field of
speech recognition is applied.
I. INTRODUCTION
Many people with severe motor disabilities require II. METHODS
alternative methods for communication and control.
The data set used in this study was obtained from the
Numerous studies over the past two decades show that
website of the EPFL BCI group [8]. The data have been
scalp-recorded electroencephalography (EEG) activity
recorded according to the 10-20 international standard
can be the basis for non-muscular communication and
from the 32 electrode configurations [9]. Each recorded
control systems. With production of advanced bio-
signal has a length of 820 samples with a sampling rate of
instruments for recording and amplifying the signals as
2048 Hz (the EEG was down-sampled from 2048 Hz to
well as cheap and powerful personal computers, this
32 Hz by selecting each 64th sample from the band pass-
dream was realized and Brain-Computer Interface (BCI)
filtered data). A six-choice signal paradigm was tested
was developed [1], [2]. Brain activity can be measured
using a population of five disable- and four able-bodied
using EEG. By extracting specific components from
subjects. The subjects were asked to count silently the
human brain activity and linking this brain activity to
number of times a prescribed image flashed on a screen.
specifically developed algorithms, an interface between a
Four seconds after a warning tone, six different images (a
computer and the users’ brain is created. Signals from the
television, a telephone, a lamp, a door, a window, and a
brain are processed to extract specific features that reflect
radio) were flashed in a random order [8]. Each flash of
the user’s intentions. Today there exist various techniques
an image lasted for 100 ms, and for the following 300 ms
by which to accomplish this [3]-[7]. The user’s brain is
no image was flashed (i.e., the inter-stimulus interval was
now coupled to a computer or external device, which
400 ms). Each subject completed four recording sessions.
allow communication or controlling devices directly,
Each of the sessions consisted of six runs with one run for
without implementing any motor action.
each of the six images. Our goal is to discriminate all
There are various properties in EEG that can be used
possible combinations of the pairs of mental tasks from
as a bases for BCI such as rhythmic brain activity (i.e.,
delta, theta, alpha, and beta) [8], event-related potentials each other using the corresponding EEG signals.
(ERPs), event-related de-synchronization (ERD) and It is difficult to compare the performances of the BCI
event-related synchronization (ERS) [9]. In the present systems, because the pertinent studies present the results
in different ways. However, in the present study, the
 comparison was made based on the accuracy and the
Manuscript received February 20, 2013; revised March 18, 2013;
accepted April 1, 2013. This research was supported by Indonesian transfer rate. The speed of a particular BCI is affected by
Institute of Sciences, Indonesia. the trial length, that is, the time needed for one selection.

©2013 Engineering and Technology Publishing 63


doi: 10.12720/ijsps.1.1.63-67
International Journal of Signal Processing Systems Vol. 1, No. 1 June 2013

This time should be shortened in order to enhance a any i.e., WF(M)WT is diagonal. Matrix F is a linear
BCI’s effectiveness in communication. The transfer rate combination of terms of the form wi wi , w is a column
T
depends on both the speed and the accuracy of selection T
and expressed as. If a trial has N possible selections and of W. WF(Mi )W is made as diagonal as possible for
each selection has the same probability of being the different combination of Mi and i  1, , k . The
desired selection, and if P denotes the probability that the T
diagonality of matrix Q  WF(Mi )W can be measured
desired choice is actually selected, then the probability
for the remaining (undesired) selections being selected as the sum of the squares of off-diagonal elements:
will be (1-P)/(N-1). The bit rate (bits/trial) of each Observe n-dimentional data
1

selection can then be expressed as [10] vector x according to: x=As Determining whitening matrix (P)

      Determining the pre-separating matrix (V)


Applying the learning rule:

 
T
(1) P ( k  1 )  P ( k )   ( k )[ I n u ( k ) u ( k )] P ( k )

   Applying the learning rule:

V ( k  1)  V ( k )   ( k ) x ( k )[ x( k )  V ( k ) x ( k )]
Determine the whitening vector (u) by:
u(k) P(k)x(k)
Determine pre-separating vector (x) by:
where N is a number of possible selections of the target x(k)V(k)x(k)
and P denotes the probability that the desired choice is No
E{P (k)P T(k)}=1
No

actually selected. The transfer rate (bits per minute) is Is x ( k ) decorrelated with
Yes

equal to b multiplied by the average speed of selection S i E{x(k)x


T
( k )} Separate the independent components

(trial per minute, which is equal to the reciprocal of the Yes


2 3
1 2
average time required for one selection). Therefore, based
Estimate the mixing matrix (A)
on the data sets information, the desired output signal is Determining separating matrix (W)
Calculate the estimating matrix (Q)
developed. Applying the learning rule:
W ( k  1)  W ( k )   ( k ) f  y ( k ) [ u
T
(k) f
T
 y ( k ) ]W ( k ) Applying the learning rule:
T
Q ( k  1)  Q ( k )   ( k )[ x Ö( k )  Q ( k ) y ( k )] y (k)
III. EXTRACTIONS AND CLASSIFICATION METHOD
Is W converged to a No
Observed the observed signals by:
desired value? x Ö( k )  Q ( k ) y ( k )

A. Nonlinear Principle Component Analysis Yes


Determine the correlation
Let x(k) represent n-dimensional vectors which Separate independent componets
coefficient of signals
according to: y(k)=W(k)u(k)
correspond to the n continuous time series from the n Is the value 2
3
EEG channels. Then xi(k) corresponds to the continuous acdepted?

sensor readings from the ith EEG channel. Because


End
various underlying sources are summed via volume
conduction to give rise to the scalp EEG, each of the xi(k)
are assumed to be an instantaneous linear mixture of n Figure 1. Nonlinear principal component cnalysis flowchart.

unknown components or sources si(k), via the unknown


mixing matrix A [7]  . Minimization of sum of squares of diagonal
x(k)  As(k)  n(k) (2) elements is same as maximization of sum of squares of
diagonal elements [10], [11].
T M 
where x(k)  [x (k), x (k), , x (k)]
1 2 M
M N
R is a noisy
  (3)
sensor vector of EEG signals, A with entries aij
is an unknown MxN mixing matrix, Joint approximate diagonalization of F (M i ) can be
s(k)  [s (k), s (k), , s N (k)]T
2
R
N
is an unknown
1
obtained by maximizing J JADE . Choice of the matrix Mi
source vector signals, and would be to take the eigen matrices of the cumulant.
T M
n(k)  [n (k), , n (k)]  R is a vector of additive After algebraic manipulations, the above equation
1 M becomes
noises. The objective of this work is to design a feed-
forward neural network and an associated adaptive ( ) 2 (4)
learning algorithm that enable estimation of the source  
( , , , )

s(k) and identification of the mixing matrix A and when the above equation is minimized the sum of squares
separating matrix W with a good tracking abilities for
of cross-cumulants of yi is also minimized.
time variable systems. These objective can be achieve
using the flowchart in the Fig. 1. C. Second-Order Blind Identification
B. Joint Approximate Diagonalization of The Second-Order Blind Identification (SOBI) is
Eigen Matrices applied as a blind source separation to EEG data collected
Joint approximate diagonalization of eigen matrices based visual stimulation with the goal of performing
(JADE) is based on diagonalization of cumulant matrices. classification of event-related potentials (ERPs) elicited
In the case of JADE the matrix W diagonalizes F(M) for under different stimulation condition. SOBI uses the EEG

©2013 Engineering and Technology Publishing 64


International Journal of Signal Processing Systems Vol. 1, No. 1 June 2013

measurements x(k) and nothing else to generate an elements in the ith row vector of G, maxj gji does the
unmixing matrix W that approximates A-1 in the Eq. 2, maximum value among the elements in the ith column
and the vector of the estimated component values y(k) in vector of G. When the perfect separation is achieved, the
the Fig. 1. Sensor space projections, which indicate the performance index is zero. In practice, the values of
effect of a given component on all sensors are given by performance index around 10-2 gives quite a good
ˆ 1 performance.
the estimated mixing matrix  . The SOBI algorithm
proceeds in two stages: First, the sensor signals are zero- D. Backpropagation Neural Networks Classifier
meaned and presphered as follows [12]
Neural networks have been proposed in the fields of
    (5) neural sciences following research into the mechanisms
and structures of the brain. The back-propagation
The angle brackets  denote an average over time, so algorithm allows exponential acquisition of input-output
the subtraction guarantees that y will have a mean of zero. mapping knowledge within multilayer networks. If a
The matrix B is chosen so that the correlation matrix of pattern is submitted and its classification is determined to
T be erroneous, the current least mean-square classification
y, y(k) y(k) , becomes the identity matrix. This is error is reduced. The error is expressed as [13]
accomplished by moving to the PCA basis using

    (6)

   (11)

where d j denotes the desired output of node j


where i are the eigenvalues of the correlation matrix,
corresponding to input xi , n is the number of training
     (7)
patterns and y j (xi , wij ) denotes the vector output of the
and U is the matrix whose columns are the corresponding networks corresponding to input xi and weight matrix 
eigenvalues, that is, the PCA components of x. The . During the association or classification phase,
second stage, one constructs a set of matrices that, in the
correct separated basis, should be diagonal. A set of time the trained neural network itself operates in a feed-
forward manner. The error is therefore a function of the
delay values  s is chosen to compute symmetrized
weights of the input and output layers. The back-
correlation matrices between the signal y(k) and a propagation algorithm is a gradient descent method
temporally shifted version: minimizing the mean square error between the actual and
    (8)
target outputs of a multilayer perceptron. Using the
sigmoid nonlinearity
T (12)
where sym(M)  (M  M ) / 2 is a function that takes f neti 
an asymetric matrix and returns a closely related  enet
symmetric one. This symetrization discards some the back-propagation algorithm consists of the following
information, but the problem is already highly valid, steps: First, initialize all weights and node offsets to small
albeit slightly weaker, constraints on the solution. The random values. Second, present continuous input vector
rotation of V that jointly diagonalizes all of them is xi and specify desired output d j . The output vector

calculated by minimazing      , the sum elements are set to zero values except for that
corresponding to the class of the current input. Third,
of the squares of the off-diagonal entries of the matrix calculate the actual output vector y using the sigmoid
products RV through an iterative process. The final nonlinearity. Fourth, adjust the weights by
estimate of the separation matrix is wij t  wij t  j xi (13)
T
WV B (9) where  j is the sensitivity of node j. Fifth, repeat the
which is used to derive the separated component of y(k). To steps from the second step. A better approach is a cross-
measure the performance of algorithms, we use the validation technique, which stops training when the error
performance index (PI) as in [11], [12] defined by on a separate validation set reaches a minimum. We
    observe records a vectors of EEG signals 
    T
   
   
  (10) [x1 (t), x2 (t), , xm (t)] from a multiple-input/ multiple-
 

    
output nonlinear dynamical system. The objective is to find
   
an inverse system, termed a reconstruction system with
where G is the global transformation matrix from s to y,
back-propagation neural networks (BPNN), in order to
gij is the (i,j) -element of the global system matrix G=HW
estimate the primary input source of brain signals
and maxj gij represents the maximum value among the

©2013 Engineering and Technology Publishing 65


International Journal of Signal Processing Systems Vol. 1, No. 1 June 2013

s(t)  [s (t),s 2 (t), ,s n


(t)]
T corresponding to particular Subject 6 reported that he accidentally concentrated on
1
the wrong stimulus during one run in session 1 [8].
stimulus, which are represented by  Shown alongside the classification accuracies using
T BPNN for all of the subjects, in Table I, are the
[ y1 (t), y2 (t), , yn (t)] .
corresponding 92% confidence intervals which the
IV. RESULTS AND DISCUSSION classified based NPCA extraction gives the best
accuracies and transfer rates with smallest standard
In brain research, the ability to measure single-trial deviations
ERPs is one important step toward the understanding of -100

how the relative timing of neuronal activity can affect NPCA

learning and how memory of a particular experience can JADE

be encoded rapidly with a single or very few exposures. SOBI

In the present study, a BPNN classifier was used. In order


to cope with nonlinearly separable problems, additional
layers of neurons placed between the input layer and the

Index ePerformanc
-101
output neuron are needed, leading to the multilayer
perceptron architecture. Performance is measured
according to the specified performance function such as
iteration speed and signal noise to ratio (SNR) criteria
[14]. The robustness of the each algorithm (NPCA, JADE,
and SOBI) was evaluated by comparing its separation
performance as shown in Fig. 2. After 500 iteration, the
2

-10
-4 0 4 8 12 16 20 24 28 32 36
NPCA algorithm perform slightly better performance and SNR [dB]

reach around 10-2 PI after 2500 iteration. This values Figure 3. Comparison of performance index of the NPCA, JADE, and
indicate that NPCA give shortest extraction time compare SOBI algorithms as a function of signal to noise ratio (SNR).
to others algorithm. Fig. 3 shows typical performances of 100 100 50

the three algorithms discussed in this paper. At high SNR, 90

80
90

80
45

40

all tested algorithms perform very well. At low SNR, one 70 Accuracy 70 35
Accuracy(%)

Trans ferrate(bits /min)


can observe that the NPCA gives better performance than
60 60 30

50 50 25

the other algorithms in most SNR ranges. In 0 - 4dB range 40 Transfer rate 40 20

JADE is worse than the others.


30 30 15

Subject 1 Subject 2
20 20 10
10 10 5

0.4 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 0

NPCA 0
Time (s)
0 50

0.35 JADE 100


90 45
50 100
90
50
45

SOBI 80 40 80
NPCA
SOBI 40

0.3
Accuracy JADE
70 35 70 35
Accuracy(%)

Trans ferrate(bits /min)


60 30 60 30
0.25 50 25 50 25

40 20 40 Transfer rate 20
PI(k)

0.2 30 15 30 15

Subject 3 Subject 4
20 10 20 10

0.15 10 5 10 5

0 0 0 0

0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50

0.1 Time (s) Time (s)

0.05
Figure 4. Comparison of classification accuracy and transfer rate plots
(averaged over eight electrode configurations) obtained with BPNN
500 1000 1500 2000 2500 3000 3500 4000 classifier for disabled subjects.
00
100 100 50
90 90 45

Number of iterations k 80 80 40

70 70 35
Transferrate(bits/min)
Accuracy(%)

Figure 2. Evolutions of PI(k) of the NPCA, JADE, and SOBI 60 60 30

algorithms. 50

40
50

40
25

20

30 30 15

The data sets for subject 5 were not included in the 20


10
Subject 6
20
10
Subject 7
10
5

simulation since the subject misunderstood the 0

0
5 10 15 20 25 30 35 40 45 50 0

0
5 10 15 20 25 30 35 40 45 0

50

instructions given before the experiment. Comparative 100

90
100

90
50

45

plots of the classification accuracies and transfer rates 80 80


NPCA
SOBI 40

(obtained with the BPNN classifier method and averaged


70 Accuracy JADE
70 35
Transferrate(bits/min)
Accuracy(%)

60 60 30

over eight electrode configurations) for the disable- (S1 - 50

40
50

40 Transfer rate
25

20

S4) and able-bodied subjects (S6 - S9) are depicted in Fig. 30 30 15

4 and Fig. 5, respectively. All of the subjects, except for 20


10
Subject 8
20
10
Subject 9
10
5

subjects 6 and 9, achieved an average classification 0 0 0

accuracy of 100% after ten blocks of stimulus


0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Time (s) Time (s)

presentations were averaged (i.e., 24.5 s). The reason for Figure 5. Comparison of classification accuracy and transfer rate plots
the poorer performance of subject 9 might be fatigue. (averaged over eight electrode configurations) obtained with BPNN
classifier for able-bodied subjects.

©2013 Engineering and Technology Publishing 66


International Journal of Signal Processing Systems Vol. 1, No. 1 June 2013

TABLE I. AVERAGE CLASSIFICATION ACCURACY (%). [6] A. Turnip and K.-S. Hong, “Classifying mental activities from
EEG-P300 signals using adaptive neural network,” Int. J. Innov.
Subject SOBI JADE NPCA Comp. Inf. Control, vol. 8, no. 9, pp, 6429-6443, 2012.
[7] A. Turnip, K.-S. Hong, and M.-Y. Jeong, “Real-time feature
S1 94.00 93.00 95.75 extraction of P300 component using adaptive nonlinear principal
S2 92.25 94.5 94.00 component analysis,” BioMedical Engineering OnLine, vol.10:83,
2011.
S3 94.00 95.5 95.00
[8] U. Hoffmann, J.-M. Vesin, and T. Ebrahimi, “An efficient P300-
S4 93.00 94.25 94.50 based brain–computer interface for disabled subjects,” Journal of
S6 90.50 91.85 92.55 Neuroscience Methods, vol. 167, no. 1, pp. 115-125, 2008.
[9] E. W. Sellers, D. J. Krusienski, D. J. McFarland, T. M. Vaughan,
S7 94.75 94.00 96.20
and J. R. Wolpaw, “A P300 event-related potential brain-
S8 94.75 96.00 97.75 computer interface (BCI): The effects of matrix size and inter
S9 93.70 92.95 94.00 stimulus interval on performance,” Biological Psychology, vol.
73, no. 3, pp. 242-252, 2006.
Average (S1–S4) 93.30.8 94.31.0 94.50.7 [10] A. Hyvarinen, J. Karhunen, E. Oja, Independent Component
Average (S6-S9) 93.42.0 93.71.7 94.00.5 Analysis, John Wiley & Sons, Inc, 2001.
Average (all) 93.41.4 94.01.3 94.30.6 [11] A. Cichocki and S. Amari. Adaptive Blind Signal and Image
Processing. New York, USA: Wiley, 2002, pp. 161-162.
[12] A. Belouchrani, K. Abed-Meraim, J. F. Cardoso, and E. Moulines,
“A blind source separation technique using second-order
V. CONCLUSION statistics,” IEEE Trans. Signal Processing, vol. 45, no. 2, pp. 434-
444, 1997.
The results presented in this study show that compared [13] Y. Yang, S. S. Ge, T. H. Lee, and C. Wang, “Facial expression
recognition and tracking for intelligent human-robot interaction,”
with the JADE and SOBI algorithms, a better extraction Intelligent Service Robotics, vol. 1, no. 2, pp. 143-157, 2008.
result can be obtained when using the NPCA algorithm [14] S. Choi and A. Cichocki, “Blind separation of nonstationary
for single-trial ERPs based on the P300 component from sources in noisy mixtures,” Electronics Letters, vol. 37, no. 1, pp.
specific brain regions. With NPCA extraction, the data 61-62, 2001.
indicate that a P300-based BCI system can communicate
at the rate around 25.4 bits/min and 36.5 bits/min for the
disable- and able-bodied subjects, respectively. The Arjon Turnip received the B.Eng. and M.Eng. degrees
average of 100% classification accuracy is achieved after in Engineering Physics from the Institute of Technology
Bandung (ITB), Indonesia, in 1998 and 2003,
nine blocks (average) for disabled subjects and after respectively, and the Ph.D. degree in Mechanical
fourteen blocks (average) for able-bodied subjects. These Engineering from Pusan National University, Busan,
results indicate that the system allowed several disabled Korea, under the World Class University program in
users to achieve transfer rates significantly beyond those 2012. He is currently work in the Technical Implementation Unit for
Instrumentation Development, Indonesian Institute of Sciences,
reported previously in the literature. Indonesia as a research coordinator. He received Student Travel Grand
Award for the best paper from ICROS-SICE International Joint
ACKNOWLEDGMENT Conference 2009, Certificate of commendation: Superior performance
in research and active participation for BK21 program from Korean
This research was supported by the tematic program government 2010, and JMST Contribution Award for most citations of
(No. 3425.001.013) through the Bandung Technical JMST papers 2011. His research areas are integrated vehicle control,
adaptive control, nonlinear systems theory, estimation theory, signal
Management Unit for Instrumentation Development processing, brain engineering, and brain-computer interface.
(Deputy for Scientific Services) and the competitive
program (No. 079.01.06.044) through the Research
Center for Metalurgy (Deputy for Earth Sciences) funded Demi Soetraprawata received the Bachelor degree in
by Indonesian Institute of Sciences, Indonesia. Engineering Physics from National University of Jakarta
(UNAS) and the Master degree in Instrumentation and
Control from Institute of Technology Bandung (ITB),
REFERENCES Indonesia. He is currently work in the Technical
[1] F. Aloise, F. Schettini, P. Arico, F. Leotta, S. Salinari, D. Mattia, Implementation Unit for Instrumentation Development
F. Babiloni, and F. Cincotti, “P300-based brain-computer (UPT BPI), Indonesian Institute of Sciences (LIPI), Indonesia as a
interface for environmental control: An asynchronous approach,” Chairman. Several activities that he had attended related to the
Journal of Neural Engineering, vol. 8, no. 2, pp. 025025, 2011. instrumentation are about analytical instrumentation at Queensland
[2] A. Belitski, J. Farquhar, and P. Desain, “P300 audio-visual University of Technology (QUT) and medical instrumentation at the
speller,” Journal of Neural Engineering, vol. 8, no. 2, pp. 025022, Royal Brisbane Hospital, in 1990 and 1992, respectively, Brisbane,
2011. Australia. His research areas are control engineering, instrument
[3] H. Cecotti, and A. Graser, “Convolutional neural networks for technology, engineering measurements, and instrumentation for
P300 detection with application to brain-computer interfaces,” calibration and metrology.
IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 33, no. 3, pp. 433-445, 2011.
[4] Z.-G. Che, T.-A. Chiang, and Z.-H. Che, “Feed-forward neural Dwi Esti Kusumandari received the B.Eng. degrees in
networks training: A comparison between genetic algorithm and Engineering Physics from the Institute of Technology
back-propagation learning algorithm,” Int. J. Innov. Comp. Inf. Bandung (ITB), Indonesia, in 1999, and the M.Eng.
Control, vol. 7, no. 10, pp. 5839-5851, 2011. degree in Biomedical Engineering from the same
[5] J. Escudero, R. Hornero, D. Abasolo, and A. Fernandez, institute in 2006. She is currently work in the Technical
“Quantitative evaluation of artifact removal in real Implementation Unit for Instrumentation Development,
magnetoencephalogram signals with blind source separation,” Indonesian Institute of Sciences, Indonesia as a junior
Annals of Biomedical Engineering, vol. 39, no. 8, pp. 2274-2286, researcher. Her research interest are in biomedical engineering, image
2011. processing, signal processing, and instrumentation.

©2013 Engineering and Technology Publishing 67

You might also like