Вы находитесь на странице: 1из 4

Selection of EEG signal features for ERD/ERS

classification using genetic algorithms

Andrzej Majkowski1, Marcin Kołodziej1, Dariusz Zapała2, Paweł Tarnowski1, Piotr Francuz2, Remigiusz J. Rak1,
and Łukasz Oskwarek1,
1
Warsaw University of Technology, Institute of Theory of Electrical Engineering,
Measurement and Information Systems, ul. Koszykowa 75, 00-662 Warszawa, Poland
2
The John Paul II Catholic University of Lublin, Al. Racławickie 14, 20-950 Lublin, Poland,

Abstract— The article presents the use of genetic algorithm Measurement of changes in the activity level of sensorimotor
(GA) to select and classify ERD/ERS patterns. One hundred EEG rhythms were carried out with GES300 system (Electrical
twenty eight channel EEG signal was used in the experiments. The Geodesics, Inc. Eugene, OR, USA), comprising Net Amps 300
signal was recorded for 40 people, during the process of imagining made with the use of a 128-channel cap with passive electrodes
right and left hand movements. Feature extraction was performed HydroCel Geodesic Sensor Net (resistance<30kΩ). Data
using frequency analysis (FFT) with the resolution of 1Hz. So the sampling was defined at 500Hz. During EEG measurement, the
features were spectral lines associated with particular electrodes. stimuli were displayed on a 17 inches LCD monitor screen, with
The selection of features, calculated for all people, was made with resolution of 1280×1024 pixels. The subjects were seated at a
GA. The fitness function used in GA was EEG signal classification
distance of 60cm from the monitor.
error calculated using LDA classifier and 5-CV test. The average
accuracy of the classification for all people in 8-30Hz band was The procedure of recording EEG signals, during the
0.85, while for the top 10 results 0.92. movement imagery, was developed by modifying the paradigm
proposed by Hwang, Kwon and Im (2009) [7]. During the
Keywords—genetic algorithm; GA; EEG, ERD/ERS; feature recording, each user performed a total of 180 trials (90 trials for
selection; classification; electroencephalograph; LORETA; left hand and 90 trials for right hand movement imagining). The
I. INTRODUCTION trials were randomly displayed for each participant.
Communication between man and machine throughout EEG III. METHODS
signals (brain-computer interface: BCI) is one of the biggest
challenges in signal processing. Such an interface can improve A. EEG Preprocessing and Feature Extraction
the standard of living of people with severe motor disabilities. The Common Average Reference method (CAR) was used
Disable persons, who cannot move, can think about moving arm, to reduce interference and artifacts on each electrode [8]. The
leg and in this way generate stable motor-related EEG signals, Butterworth notch filter 48Hz-52Hz of 4-th order was used to
so called desynchronization/synchronization - ERD/ERS. remove the 50Hz network frequency. No other pre-processing
methods were implemented. Any signals from selected
The fundamental problem in all BCIs is the proper electrodes or fragments of signals containing artifacts were
“understanding” of EEG signals [1]. There are three, commonly removed.
used, typical stages of EEG signal analysis: feature extraction,
feature selection and classification [2]. Many methods of feature Feature extraction was performed using Fast Fourier
extraction are known. Very often frequency analysis is used. It Transform (FFT) for each "trial" individually. For this purpose,
turns out that an interesting method of optimization that can be each trial (4.5s) was divided into one-second windows with 50%
used to select the best features is the genetic algorithm GA [3]– overlap. Then for each window the spectrum with a resolution
[5], [6]. The purpose of this publication is to present the results of 1Hz was calculated. Spectra calculated for individual
of the selection of EEG signal features obtained using GA, windows were averaged.
calculated for a large number of people, for the classification of
ERD/ERS patterns. TABLE I. FEATURES USED IN GENETIC ALGORITHM

Frequency Number
II. MATERIALS band [Hz] of features
Comments
The experiment was carried out with 40 subjects, aged from 1- 40 5120 40 Spectrum lines ×128 electrodes
20 to 27 (M = 22.60; SD = 1.72; 20 women, one left-handed). 1-50 6400 50 Spectrum lines ×128 electrodes
All the subjects were volunteers who gave their consent to take 1-100 12800 100 Spectrum lines ×128 electrodes
7-30 2944 23 Spectrum lines ×128 electrodes
part in the study and declared they were not taking medication
or other psychoactive substances on a permanent basis. The
Ethics Committee of the Institute of Psychology approved the The averaged spectrum represented features of the
study. individual trial. Ninety such trials were used for imagining the
right hand movement and ninety trials for imagining the left IV. RESULTS
hand movement. For genetic algorithms we used 1-40Hz, Table II shows the average classification results obtained for
1-50Hz, 1-100Hz, 7-30Hz spectral bands in which features all subjects using spectral lines, as features, corresponding to the
(spectral lines) were calculated. The features on which the bands: 1-40Hz, 1-50Hz, 1-100Hz, 7-30Hz. The average
genetic algorithm was trained are summarized in Table 1. Such classification accuracy for 10 participants with the best results
a large number of features makes it impossible to directly learn were presented in Table III. The average classification accuracy
the classifier, hence the need for their selection [9]. for the 10 participants with the worst results were presented in
B. Genetic Algorithm Table IV.
The genetic algorithm block diagram, used for feature TABLE II. AVERAGE CLASSIFICATION ACCURACY FOR ALL 40 PEOPLE
selection, is shown in Fig. 1.
Frequency band 1-40Hz 1-50Hz 1-100Hz 7-30Hz
Average accuracy 0.79 0.84 0.84 0.85
Standard deviation 0.085 0.054 0.049 0.050

TABLE III. AVERAGE CLASSIFICATION ACCURACY FOR 10 BEST RESULTS


Frequency band 1-40Hz 1-50Hz 1-100Hz 7-30Hz
Average accuracy 0.91 0.92 0.90 0.92
Standard deviation 0.055 0.039 0.047 0.041

TABLE IV. AVERAGE CLASSIFICATION ACCURACY FOR 10 WORST


RESULTS

Frequency band 1-40Hz 1-50Hz 1-100Hz 7-30Hz


Average accuracy 0.70 0.79 0.79 0.81
Standard deviation 0.021 0.011 0.013 0.022

In Figures 2, 3, 4 the most commonly chosen EEG features


for 70 runs of GA for one of the participants are presented. The
parameters of the GA for this case were: K=8 chromosomes
(desired features), population size N=190. The results indicate
that 8 to 12Hz spectral lines for electrodes located on the motor
cortex (C3 and C4 - numbers 32 and 104, respectively) were
Fig. 1. Block diagram of the genetic algorithm.
most commonly selected by the algorithm. Average
The first step of the algorithm (Initialize population) was the classification accuracy for all runs was 0.89, while the highest
creation of a startup population of N=190 individuals, containing classification accuracy was 0.96. This demonstrates the good
K=20 chromosomes, representing features (FFT spectrum lines). choice of the features for the ERD/ERS signal classification
Chromosomes for the input population were generated task.
randomly.
An important element of a genetic algorithm is to define the
fitness function (Evaluation), as a measure of selection quality
of the features. As a fitness function, the EEG signal
classification error was used. Classification was performed
using the linear discriminant analysis (LDA) and the 5-fold
cross-validation test (5-CV) [10].
In next step each individual was evaluated using "roulette
wheel” algorithm [11] and on this basis, the best N individuals
were selected for the next round (Selection). In this way a new
population of children was created.
For the created population, the stages of crossing and
mutation were performed (Crossover and Mutation) [12]. The
probability parameter for the crossing operation was 0.8 and for
the mutation 0.01. As a result of these operations, there was an
exchange of features in the originally generated population.
The selection, crossing and mutation stages were repeated Fig. 2. Number of feature choices (spectral line associated with an electrode)
290 times. As a result of GA, "best" features were obtained for for 70 GA runs.
the ERD/ERS classification.
Fig. 3. Number of feature choices associated with particular electrodes. Fig. 5. Distribution of t-values calculated for one user.

Fig. 4. Number of feature choices associated with particular frequencies.


Fig. 6. Sum of t-values determined for each frequency.

V. DISCUSSION
The obtained classification results are satisfactory and show
that genetic algorithm is very well suited to indicate the most
significant features for ERD/ERS recognition. For all
participants, the best classification accuracy (0.85) was obtained
for a limited band 7-30Hz. The best frequency subband foe
ERD/ERS classification is 8-12Hz (Fig. 2-4). Similar results
were obtained for other users.
For comparison, an experiment for another feature selection
method t-test [13] was also performed. Fig. 5 presents a
distribution of t-values calculated for one user. The sum of t-
values calculated for each frequency is shown in Fig. 6. In this
case, the most significant frequencies were in the 8-12Hz and
18-25Hz bands.
In order to directly compare the classification accuracies,
best K features were calculated using GA and t-test. The LDA
classifier was used for the t-test selection. The results of the
classification accuracies for both methods, depending on the Fig. 7. Classification accuracy when selecting features using GA and t-tets for
number of selected features, are presented in Fig. 7. different number of features.
Differences of classification results for GA and t-test [2] D. Hu, W. Li, and X. Chen, “Feature extraction of motor imagery EEG
selection methods results from the way these methods work (the signals based on wavelet packet decomposition,” in The 2011
IEEE/ICME International Conference on Complex Medical Engineering,
wrapper method vs. the filter method) [14]. Single run of the GA 2011, pp. 694–697.
selection method took 20.3 seconds (using Matlab and Paralel
[3] E. Yom-Tov and G. F. Inbar, “Feature selection for the classification of
Computing Toolbox - 4 workers, Intel Core i7-6700K 4GHz movements from single movement-related potentials,” IEEE
CPU 16GB RAM). Transactions on Neural Systems and Rehabilitation Engineering, vol. 10,
no. 3, pp. 170–177, Sep. 2002.
In addition, the authors calculated the location of ERD/ERS
[4] I. Rejer, “Genetic Algorithms in EEG Feature Selection for the
sources using the LORETA algorithm [15]. The differences in Classification of Movements of the Left and Right Hand,” in Proceedings
brain activity for right and left hand movement imagining are of the 8th International Conference on Computer Recognition Systems
shown in Fig. 8. CORES 2013, R. Burduk, K. Jackowski, M. Kurzynski, M. Wozniak, and
A. Zolnierek, Eds. Heidelberg: Springer International Publishing, 2013,
pp. 579–589.
[5] R. Corralejo, R. Hornero, and D. Álvarez, “Feature selection using a
genetic algorithm in a motor imagery-based Brain Computer Interface,”
in 2011 Annual International Conference of the IEEE Engineering in
Medicine and Biology Society, 2011, pp. 7703–7706.
[6] M. Kołodziej, A. Majkowski, and R. Rak, “Implementation of genetic
algorithms to feature selection for the use of brain-computer interface,”
Przegląd Elektrotechniczny, vol. 87, pp. 71–73, 2011.
[7] H.-J. Hwang, K. Kwon, and C.-H. Im, “Neurofeedback-based motor
imagery training for brain-computer interface (BCI),” J. Neurosci.
Fig. 8. Differences in brain activity for right and left hand movement Methods, vol. 179, no. 1, pp. 150–156, Apr. 2009.
imagining calculated for the frequency of 10.25Hz. [8] D. J. McFarland, L. M. McCane, S. V. David, and J. R. Wolpaw, “Spatial
filter selection for EEG-based communication,” Electroencephalography
The designated areas indicate the greatest differences in the Clinical Neurophysiology, vol. 103, no. 3, pp. 386–394, Sep. 1997.
areas of sensorimotor cortex. This is in line with the theory of [9] S. Osowski, Metody i narzędzia eksploracji danych. Legionowo: BTC,
ERD/ERS potentials [16]. 2013.
[10] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification. Wiley-
VI. CONCLUSIONS Interscience, 2000.
[11] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution
Genetic algorithms are well suited to selection of best Programs (3rd Ed.). London, UK, UK: Springer-Verlag, 1996.
features for ERD / ERS signal classification. Selected features
[12] L. He, Y. Hu, Y. Li, and D. Li, “Channel selection by Rayleigh coefficient
for the best individual can greatly increase the effectiveness of maximization based genetic algorithm for classifying single-trial motor
BCI. During the experiments it turned out that selection made imagery EEG,” Neurocomputing, vol. 121, pp. 423–433, Dec. 2013.
using GA gives more accurate results than selection made with [13] I.-H. Lee, G. H. Lushington, and M. Visvanathan, “A filter-based feature
a popular t-test. Selection using GA shows that the best features selection approach for identifying potential biomarkers for lung cancer,”
are in the frequency range of 8-12Hz. Journal of Clinical Bioinformatics, vol. 1, p. 11, 2011.
[14] G. Chandrashekar and F. Sahin, “A survey on feature selection methods,”
ACKNOWLEDGMENT Computers & Electrical Engineering, vol. 40, no. 1, pp. 16–28, Jan. 2014.
The research was partially funded by the National Science [15] R. D. Pascual-Marqui, “Standardized low-resolution brain
electromagnetic tomography (sLORETA): technical details,” Methods
Center, (grant no. 2011/03/N/HS6/02798). Find Exp Clin Pharmacol, vol. 24, pp. 5–12, 2002.
[16] G. Pfurtscheller, C. Brunner, A. Schlögl, and F. H. Lopes da Silva, “Mu
REFERENCES
rhythm (de)synchronization and EEG single-trial classification of
[1] J. R. Wolpaw et al., “Brain-computer interface technology: a review of different motor imagery tasks,” NeuroImage, vol. 31, no. 1, pp. 153–159,
the first international meeting,” IEEE Trans Rehabil Eng, vol. 8, no. 2, May 2006.
pp. 164–173, Jun. 2000.

Вам также может понравиться