0% found this document useful (0 votes)
418 views

Feature Extraction of Eeg Signals Using Power Spectral Entropy

This paper presents an approach that performs EEG feature extraction during imagined right and left hand movements by using power spectral entropy (PSE) it acquires good classification results with the time-variable linear classifier. The maximal accuracy achieves 90%. The results show that the PSE is a sensitive parameter for EEG of imaginary hand movements.

Uploaded by

Chua Bee Lin
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
418 views

Feature Extraction of Eeg Signals Using Power Spectral Entropy

This paper presents an approach that performs EEG feature extraction during imagined right and left hand movements by using power spectral entropy (PSE) it acquires good classification results with the time-variable linear classifier. The maximal accuracy achieves 90%. The results show that the PSE is a sensitive parameter for EEG of imaginary hand movements.

Uploaded by

Chua Bee Lin
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

2008 International Conference on BioMedical Engineering and Informatics

Feature Extraction of EEG Signals Using Power Spectral Entropy

Aihua Zhang, Bin Yang, Ling Huang


Lanzhou University of Technology, Lanzhou 730050
E-mail:[email protected]

Abstract extract the features of different EEG signals, such as


power spectral estimation, AR model[2, 3]and wavelet
Brain-Computer Interfaces (BCI) use electroenc- transform (WT)[4] etc. Extracting band energy by
ephalography (EEG) signals recorded from the scalp power spectral estimation needs determine subject-
to create a new communication channel between the specific frequency band, AR model needs determine
brain and an output device by bypassing conventional appropriate order of the model, and wavelet analysis
motor output pathways of nerves and muscles. One of needs select proper mother wavelet. In order to extract
the most important components of BCI is feature features which can best reflect the different mental
extraction of EEG signals. How to rapidly and reliably tasks, a new EEG feature with power spectral entropy
extract EEG features for expressing the brain states of (PSE) is proposed in this paper. PSE based-on
different mental tasks is the crucial element for exact Shannon information theory is an index of complexity
classification. This paper presents an approach that measure for an uncertain system. It has a good metrical
performs EEG feature extraction during imagined effect for the change of nonlinear dynamic states, and
right and left hand movements by using power spectral that it just needs a little data, so this method is well
entropy (PSE). It acquires good classification results suited the analysis of imaginary EEG signals [5, 6].
with the time-variable linear classifier. The maximal Interrelated work includes that using it for studying the
accuracy achieves 90%.The results show that the PSE extent of focal ischemic cerebral injury [7], monitoring
is a sensitive parameter for EEG of imaginary hand the stages of sleep [8] and studying relationship between
movements. The method is simple and quick and it age and cardiovascular system [9].
provides a promising method for on-line BCI system. In this paper, the values of PSE are calculated as
feature of EEG signals during imagined right and left
1. Introduction hand movements. Time-variable classifier is used for
classification of signals. Finally, the satisfactory result
Patients in a late stage of amyotrophic lateral is obtained with the highest classification accuracy
sclerosis (ALS) or suffering from a locked-in 90%. The result shows that the PSE is a sensitive
syndrome are not able to produce any voluntary muscle parameter for EEG signals of imaginary hand
movements [1]. Brain–computer interfaces (BCI) are movements. The method is simple and quick and it
systems that allow us to translate human intentions provides a promising method for on-line BCI system.
reflected by suitable brain signals into a control signal In the following sections, the description of the data
for an output device such as a cursor on computer set, PSE algorithm and time-variable linear classifier
screen, wheel chair or artificial limb. They do not rely are presented in Section 2. Section 3 presents
on muscular activity and can therefore provide a new classification results and discussion. A brief conclusion
communication and control channel for those patients. is given in Section 4.
The present work mainly focuses on off-line research
of BCI. It was achieved correct recognition for 2. Methods
different mental tasks by extracting relevant features of
EEG signals for expressing the brain states. 2.1. Description of the data set
One of the most important components of BCI is the
EEG feature extraction procedure. How to rapidly and The data is coming from a set of EEG data set in
reliably extract EEG features is the crucial element for BCI competition 2003 which is provided by Graze
exact classification and translation of mental tasks. University of Technology. It was recorded from a
Recently, a variety of methods has been widely used to normal subject (female, 25years) during a feedback

978-0-7695-3118-2/08 $25.00 © 2008 IEEE 1359


435
DOI 10.1109/BMEI.2008.254

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 21,2010 at 00:13:55 UTC from IEEE Xplore. Restrictions apply.
session. The subject sat in a relaxing chair with
armrest. The task was to control a feedback bar by
means of imaginary left or right hand movement. The
experimental is described in detail in [10, 11]. The aim
of signal processing was to distinguish the mental tasks 1 2 3

that imagined left or right hand movement. The

5 cm
experiment consists of 7 runs with 40 trials each. All C3 1 Cz 2 C4 3

runs were conducted on the same day with several


minutes break in between. The order of left and right
cues was random. Given are 280 trials of 9s length.
The first 2s was quiet, at t=2s an acoustic stimulus
indicates the beginning of the trial, the trigger channel
went from low to high, and a cross was display for 1s; (a)
then at t=3s~9s, an arrow (left or right) was displayed
as cue (see Fig.1). At the same time the subject was
asked to move a bar into the direction of the cue. The
feedback was based on AAR parameters which were
combined with a discriminate analysis into one output
parameter. The recording was made using a G.tec
amplifier and a Ag/AgCl electrodes. Three bipolar
EEG channels (anterior ‘+’, posterior‘-’) were 0 1 2 3 4 5 6 7 8 9 sec
measured over C3, Cz, and C4 (see Fig.1). The EEG
was sampled with 128 Hz and it was filtered between Feedback period with Cue

0.5 and 30Hz. To prevent any systematic effect due to Trigger


Beep
the feedback, the training (140 trials) and testing (140
trials) data were randomly selected from 280 trails. (b)
2.2. Power Spectral Entropy Figure1. (a) Electrode position. (b) Timing scheme.

Power spectral entropy is information entropy that is 1) The Discrete Fourier Transform X (ωi ) of signal
able to quantify the spectral complexity of an uncertain
system. It is defined in [12]. For an uncertain system, can be obtained by FFT; where ωi is the frequency
let’s assume a random variable X as states of the point of the number i .
system, the value of X is 2) Calculate its power spectral density (PSD)
1
X = {x1 , x2 ,", xn } Pˆ (ωi ) =
2
(n ≥ 1) (1) X (ωi )
N
The corresponding probability is ˆ
3) Normalize P (ωi ) , and obtain power spectral
P = {p1, p2 ," , pn } 0 ≤ pi ≤ 1 , i = 1,2," , n (2) density distribution function
Under constraints Pˆ (ωi )
pi =
n
∑ Pˆ (ω )
∑ pi = 1
i
(3) i
i =1 4) Use equation (4), the PSE can be worked out.
Therefore, the information entropy of the system can PSE can be interpreted as a measurement of the time
be expressed as uncertainty in frequency domain. In power spectrum of
n EEG signals, when the spectrum peak is narrow, its
H = −∑ pi ln pi (4) entropy value is small. It indicates that the signal has
i =1 an obvious concussive rhythm, that is to say, if wave is
The time-series signals become power spectrum by orderliness, its complexity is small; whereas, the
FFT transform, and the information entropy of power spectrum peak is more smoothness, and its entropy
spectrum is called power spectral entropy. The value is more greatness. Therefore, the PSE can reflect
algorithm can be summarized as: the spectra structure of EEG signals.

1360
436

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 21,2010 at 00:13:55 UTC from IEEE Xplore. Restrictions apply.
2.3. Classification
1.25

Assume xk (k = 1,2," , N ) as feature vector of imagined left hand movement


imagined right hand movement
1.2
training data set, in which N1 belongs the sample of

power spectral entropy


imaginary left hand movement, defined as subset χ1 ,
1.15

and in which N 2 belongs the sample of imaginary 1.1

right hand movement, defined as subset χ 2 .The 1.05

individual within-class scatter matrix and the total 1


within-class scatter matrix is defined by
S = ∑ ( x − µ )( x − µ ) i = 1,2
T
(5) 0.95
i i i
x∈χ i 0.9
1 2 3 4 5 6 7 8 9
S w = S1 + S 2 (6) time(s)
where µ i is the mean vector of the class i . (a)
It can be obtained the transform vector w with 1.24
imagined left hand movement
maximal between-class distance and minimal within- 1.22 imagined right hand movement

class variance by Fisher criterion function and 1.2

power spectral entropy


Lagrange multiplier method: 1.18

w = Sw
−1
(µ1 − µ 2 ) (7) 1.16

Hence, linear transformation can be described as: 1.14

y = w x − 0.5( w, µ1 + w, µ 2
T
) (8) 1.12

1.1
To improve classification accuracy of time t0 , a
1.08
time-variable linear classifier is proposed in this paper 1.06

and in which the information of time t < t0 is used. 1.04


1 2 3 4 5 6 7 8 9
The time-variable classifier is defined as: time(s)
D (t0 ) = ∑ ct y (t ) + y (t0 ) (9) (b)
t <t0 Figure2. The time course of power spectral entropy for
imagining left (solid) and right (dashed) hand movements
where ct is a time-variable coefficient. (a) Channel C3 (b) Channel C4
The classification is conducted as follows: 0.9

left hand if D (t0 ) > 0 0.85


D (t0 ) ∈  (10)
right hand if D(t0 ) < 0
0.8
classification accuracy

0.75

3. Results and Discussion


0.7

0.65

This section presents experimental results on the 0.6

EEG data set which was used in the BCI competition 0.55

2003. The result was obtained by using the feature of 0.5

PSE and time-variable linear classifier. 0.45


To gain the time course of feature vector, EEG was
0.4
divided into 1 second segments (128 sample points) 1 2 3 4 5 6 7 8 9
time(s)
with overlapping data, and the PSE of 8~24Hz for
Figure3. The time course of classification accuracy with
each time segment is calculated with sliding time power spectral entropy for imagining left and right hand
windows stepped by one sample point. Because movements
channel Cz shows its independence of the motor
imagery, only channels C3 and C4 were used for unilateral movement but also movement imaging
feature extraction. It was demonstrated that not only activates primary sensorimotor areas, whereby

1361
437

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 21,2010 at 00:13:55 UTC from IEEE Xplore. Restrictions apply.
generally a circumscribed “event-related desynchron- Science Foundation of China (No. 30670529) and the
ization” (ERD) is characteristic for the contralateral, Chunhui Plan of Ministry of Education of China
and an “event-related synchronization” (ERS) for the (Z2005-2-62007)
ipsilateral hemisphere [13]. This fact is a foundation for
distinguish left and right hand movements. References
Fig.2 displays the PSE on channels C3 and C4
corresponding to imagine left hand movement. It [1] R. Herbert, M. G. Johannes, and G. Pfurtscheller,
shows difference in PSE distribution between the two “Optimal Spatial Filtering of Signal Trial EEG during
tasks. For channel C3 [Fig.2 (a)], the values of PSE by Imagined Hand Movement”, IEEE Trans .Rehab. Eng. , Dec.
imaginary left hand is lower than that of right hand. 2000, vol. 8(4), pp. 441-446.
Similar conclusion can be drawn from Fig.2 (b) for
channel C4. The time change regulation of PSE is [2] B. O. Peters, G. Pfurscheller, and H. Flyvjberg,
“Automatic differentiation of multi-channel EEG signals” ,
accordance with ERD/ERS. And the PSE has a good IEEE Trans. Biomed. Eng. , 2001, vol. 48(1), pp.111-116.
separability for the two tasks. So the PSE on channels
C3 and C4 can be used as feature of EEG signals [3] D. Coyle, G. Prasad, and T. M. McGinnity, “A time-
during imagined right and left hand movements. series prediction approach for feature extraction in a Brain-
According to the analysis above, the classification Computer Interface”, IEEE Trans. Neur. Syst. Rehabil. Eng. ,
accuracy can be obtained with the feature of PSE and Dec. 2005, vol.13, pp. 461-467.
time-variable linear classifier. For training data set, the
feature vectors were extracted at every sampling point, [4] V. Bostanov, “BCI Competition 2003-Data sets Ib and
with the sliding window size being 128 samples and IIb: Feature extraction from Event-Related brain potentials
with the continuous wavelet transform and t-Value
the interval is one sample. The time-variable linear scalogram”, IEEE Trans. Biomed. Eng. , June 2004, vol.51,
classifier was trained with these feature vectors. Then pp. 1057-1061.
the feature vectors of testing data set was calculated by
similar method and classified by trained classifier. [5] S. M. Pincus, and A. L. Goldberger, “Physiological time-
From Fig.2, it shows obviously difference in PSE of series analysis: what does regularity quantify?”, American
imaginary left and right hand movements during 4.5s Journal of Physiology,1994, 1643-1656.
to 7s. So the accumulation of information was started
from 4.5s in the algorithm of classification. The time [6] H. Z. Qi, and L. L. Cheng, “Study on Information
course of classification accuracy is shown in Fig.3. It Entropy of EEG Signals during Motor Imaginary”, Chinese
Journal Biomedical Engineering, Feb. 2007, vol.26 (1), pp.
shows the maximal accuracy achieves 90% around 74-77.
t=6.7s. Subject was quiet at first 2s, and at t=3s
beginning to imagine the mental tasks. Because of the [7] H. J. Wu, H. Zhang, and C. X. Cheng, “Application of
accumulation of information, the accuracy achieves power spectral entropy to the noninvasive detection of focal
max around t=6.7s. ischemic cerebral injury”, Journal of Biomedical Engineer-
ing, 2003, vol. 20 (2), pp.229-232.
4. Conclusion
[8] H. Liu, W. X. He, and X. P. Chen, “EEG time-series
analysis using nonlinear dynamics for sleep monitoring”,
This paper proposes new feature extraction and Journal of Jiangsu University: Natural Science, 2005, vol.26
classification methods for EEG signals during imaged (2), pp.174-177.
left and right hand movements. The Experimental
results have shown that the PSE can be used as one of [9] D. Z. Zhang, “Relationship between age and power
the feature of mental tasks successfully. Furthermore, spectral entropy of blood pressure waveform”, Chinese
we applied time-variable classifier for classify and Journal of Medical Physics, 2005, vol.22 (4), pp.606-607.
gained a good result. The result demonstrates that the
proposed algorithm is reasonable and effective for [10] A. Schlögl, K. Lugger, and G.Pfurtscheller, “Using
classification of imaginary left and right hand Adaptive Autoregressive Parameters for a Brain-Computer
movements. High classification accuracy and quick Interface Experiment”, Proceedings of the 19th Annual
calculation speed of the algorithm make it becoming a International Conference IEEE/EMBS, 1997, pp.1533-1535.
feasible choice for a practical BCI system.
[11] G. Pfurtscheller, C. Neuper, A. Schlögl, et al, “Separabi-
lity of EEG signals recorded during right and left motor
Acknowledgment imagery using adaptive autoregressive parameters”, IEEE
Trans. Rehab. Eng ., 1998, vol. 6(3), pp.316-25.
This work was supported by the National Natural

1362
438

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 21,2010 at 00:13:55 UTC from IEEE Xplore. Restrictions apply.
[12] J. N. Kapur, and H.K. Kesavan, Entropy optimization
principles with applications, Academic Press, San Diego,
1992.

[13] G. Pfurtscheller, C. Neuper and C. Guger, “Current


Trends in Graze Brain-Computer Interface (BCI) Research”,
IEEE Trans. Rehab. Eng., vol. 8(2), Jun.2000, pp. 216-219.

[14] G. Pfurtscheller, C. Neuper, D. Flotzinger and M. Preg-


enzer, “EEG-based discrimination between imagination
ofright and left hand movement”, Electroencephalography
and clinical Neurophysiology, 1997, pp.642-651.

[15] C. Neuper, and G. Pfurtscheller, “Event-related


dynamics of cortical rhythms: frequency specific features and
functional correlates”, International Journal of Psychophy-
siology, 2001, vol.43 (1), pp.41-58.

1363
439

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 21,2010 at 00:13:55 UTC from IEEE Xplore. Restrictions apply.

You might also like