9th Annual Conference of the International FES Society September 2004 – Bournemouth, UK
EEG-based Brain-Computer Interface For Hand Grasp Control: Feature Extraction by Using ICA Erfanian A, Erfani A Dept. Biomed. Eng., Faculty of Electrical Eng., Iran University of Science and Technology Tehran, IRAN Email:
[email protected]
Abstract This work provides a natural basis for developing an efficient EEG-based braincomputer interface (BCI) for natural control of prosthetic hand grasp. The system is based on single-feature obtained by independent component analysis (ICA) of single-channel EEG. ICA is a useful technique that allows blind separation of sources, linearly mixed, assuming only the statistical independence of these sources. This suggests the possibility of using ICA to separate different independent brain activities during motor imagery into separate components. The tasks to be discriminated are the imagination of hand grasping and opening and the resting state. The results indicate that the proposed scheme improves the classification accuracy of the EEG patterns during motor imagery.
1
Introduction
Recently, many attempts have been done to use the electroencephalogram (EEG) as a new communication channel between human brain and computer. This new communication channel is called EEG-based brain-computer interface (BCI). To date, different types of BCI were suggested by different research groups. A survey and compilation of the progress in this field is contained in Mason et al [1]. To classify the EEG patterns, feature vectors must be created. The classification performance is profoundly affected by the choice of feature set. Even when the features presented contain enough information about the output class, they may not predict the output correctly because the dimension of feature space may be so large that it may require numerous instances to determine the relationship. It was reported that the performance of classifier systems deteriorates as new irrelevant features are added [2]. In this work, we used independent component analysis (ICA) to extract new features for
classifying the EEG patterns associated with the resting state and the imagined hand movements. ICA is a statistical technique which allows blind separation of sources, linearly mixed at the sensors, assuming only the statistical independence of these sources [3]. Such method was used to separate neural activity from muscle and blink artifacts in spontaneous EEG data [4]. It was verified that the ICA can separate artificial, stimulus-locked, responselocked, and non-event related background EEG activities into separate components [5]. This suggests the possibility of using ICA to separate different independent brain activities during motor imagery into separate components. For purpose of feature extraction, the key question is which components carry more information about the task. In [2], a method was proposed for ICA to select a number of components that carry information about the task and a number of components that do not. It was shown that the proposed algorithm reduces the dimension of feature space while improving classification performance. In this work, we employed this method for feature selection during classification of EEG patterns associated with resting state and motor imagery.
2
ICA-Based Feature Extraction
ICA is a statistical technique in which observed random data are linearly transformed into components that are maximally independent from each other [3]. The output of ICA is a set of maximally independent components which are linear combinations of observed data. These components do not carry any information about the class labels. In [2], a feature extraction method was proposed for binary classification problem by incorporating standard ICA algorithm (ICA-FX). The structure of ICA-FX is shown in Fig. 1. Here, the original observed data x(k ) = [x1 (k ), x2 (k ),..., xN (k ) ] is fully connected to
9th Annual Conference of the International FES Society September 2004 – Bournemouth, UK u = [u1 , u2 ,..., u N
]T , class label c is connected to T u a = [u1 , u 2 ,..., u M ] and u N +1 = c . The weight matrix is w1, N +1 ⎤ ⎡ w1,1 K w1, N ⎢ ⎥ M M M ⎢ ⎥ ⎢ wM ,1 K wM , N wM , N +1 ⎥ ⎢ ⎥ W = ⎢ wM +1,1 K wM +1, N 0 ⎥ ⎢ ⎥ M M M ⎢ ⎥ ⎢ w N ,1 K w N , N ⎥ 0 ⎢ ⎥ 0 ⎢⎣0 K 0 K ⎥⎦
any warning tone. Data were recorded for 5 s during each trial experiment and each trial was repeated 50 times for each task. One of the major problems in developing realtime BCI is the eye blink artifacts. We used adaptive noise canceller filter using neural network for real-time removing the eye blinks interference from the EEG signals [7].
4
Now, the aim is to separate the observed space x into two linear subspace: one that is spanned by that contains maximal f a = [ f1 ,..., f M ] information about the class label c and the other spanned by f b = [ f M ,..., f N ] that is independent of c as much as possible. In [2], a learning rule was derived for W using a similar approach for deriving the learning rule for ICA. The learning rule was obtained as
Independent Component Analysis of EEG During Motor Imagery
Fig. 2 (a) shows the brain potentials obtained by ensemble averaging of 50 trials artifact-free EEG during hand movement imagination for one experiment day. The event-related potentials and DC-potential shifts associated with motor imagery are quit evident in these figures. The subsequent DC-shifts after the beginning of imagination are related to creating and maintaining mental imagery.
W t +1 = W t + µ1 [ I N − φ (u ) f T ]W T v t +1 = v at − µ 2φ (u a )c.
I N is an N × N identity matrix, and µ1 and µ2 are learning rates that can be set differently. -W 1,N+1c x1
u1
• • •
• • •
xM
uM
xM+1
uM+1
• • •
• • •
xN
uN
c
uN+1
+
f1
• • • +
fM
-W M,N+1c
Figure 1: The structure of feature extraction based on ICA (ICA-FX) [2].
3
(a)
Experiments
The EEG data of healthy right-handed volunteer subjects and a below elbow amputee were recorded at a sampling rate of 256 by Ag/AgCl scalp electrodes. The eye blinks were recorded by placing an electrode on the forehead above the left brow line. Depending on the cue visual stimuli which is appeared on the monitor of computer, the subject imagines the hand grasping or opening. If the visual stimuli is not appeared, the subject does not perform a specific task. During each trial experiment, one task was performed without
(b) Figure 2: Ensemble average of 50 trials artifactfree EEG (a) and ICA components (b) during hand movement imagination for one experiment session.
To apply ICA, subepochs of artifact-free EEG (from 0 to 2000 ms after visual stimulus) obtained during 50 trials were concatenated. After training, unmixing weighting matrix was frozen and applied to each single-trial 7channel EEG. The ensemble averaging of ICA components over 50 trials during motor imagery is shown in Fig. 2 (b). It is observed that ICA could perfectly detect and separate different brain activities during motor imagery. The event-related potential associated with
9th Annual Conference of the International FES Society September 2004 – Bournemouth, UK
motor imagery is quite evident in ICA component 5. ICA component 6 reveals the slow EEG responses. This component also reveals the visual evoked-potential associated with visual cue. ICA components 4 and 7 account for gamma activity which increases during imagination.
5
ICA-Based EEG Classification
A 1-s time interval starting 0.20 s after cue presentation was used for feature extraction and classification. The AR model of order 17 was estimated from the 1-s interval of singlechannel during each trial of experiment. Then AR parameters was applied to ICA-FX. The value of m was set to 1. Therefore, one component contains maximal information about the class label c and other components are independent of class c as much as possible. Based on the value of f1, the input patterns can be classified. In this work, we employed ICA-based classifier and compared the performance with a neural network classifier. The multilayer perceptron (MLP) with back-propagation learning rule was used. The MLP network considered in this
output nodes. The classifiers were trained with data obtained during 50% of the experimental trials and were validated with data obtained during the subsequent trials. The learning process is stopped when it is apparent that the generalization performance has peaked. To assess the robustness of the proposed scheme in EEG classification, two different data sets were created for training and evaluating the classifier. For each of the two data sets obtained during each experiment day, a classifier was trained and evaluated, then the results were averaged.
6
Results
Table 1 summarizes the results of singlechannel EEG classification for different subjects. Here, the average of classification accuracy over different day experiments is reported on each subject. We observed that single-feature which is computed by ICA improved the EEG classification accuracy compared to multi-feature neural network classifier. At the best case, an average accuracy as high as 82%, 75%, 82%, 75%, 81%, and 78% is achieved for the subjects AE, SN, ME,
Table 1: Single-channel EEG Classification accuracies using neural network (MLP) and ICA-FX
F3 Subject
MLP
ICAFX
Cz
T5
MLP
ICAFX
Pz
MLP
ICAFX
MLP
ICAFX
F4
Fz
C3
MLP
ICAFX
MLP
ICAFX
MLP
ICAFX
AE (normal)
70.5
73.5
74.7
78.1
71.3
78.8
75.6
82.2
68.3
71.8
67.7
74.2
70
77.7
SN (normal)
64.6
72.5
61.3
71.1
66.8
72.5
65
72.1
65.3
69
62.6
68.5
65.3
75.3
ME (normal)
73.8
74.5
75.3
82.1
74
76.3
79.3
78.5
72.1
74.8
70.8
68.3
71.6
75.6
EA (normal)
61.6
68.5
63.8
65.8
66.1
73.6
65.6
72.3
72
74.8
69.5
75
70.6
74.6
ST (normal)
75.5
77.7
75.2
80
78.8
80.7
76.6
76.6
75
78.6
75.5
79.1
79.2
77.5
SM(amputee
57.2
59.2
66.7
77.7
63.5
64
60
67.7
65.7
66.7
63
65.5
66.2
63.7
study consists of two hidden layers each containing hyperbolic tangent units and two References [1]
[2]
[3]
[4] [5]
Mason, A general framework for brain-computer interface design, IEEE Trans. Neural Systems and Rehab Eng.,11(1), 2003. Kwak, Feature extraction based on ICA for binary classification problems, IEEE Trans. knowledge and data eng, 15(6):1374-1388, 2003. Hyvarinen, Fast and robust fixed-point algorithms for ICA, IEEE Trans. on Neural Network, 10(1): 629-634, 1999. T Jung, Removing EEG artifacts by blind source separation, Psychophysiology, 163-178, 2000. Jung, Analysis and visualization of single-trial event-related potentials,” Human Brain Mapping, 14: 166-185, 2001.
EA, ST, and SM, respectively by using ICAFX. [6]
[7]
Bell, An Information-maximization approach to blind separation and blind deconvolution, Neural Computation, 7(6), 1995. Erfanian, Real-time eye blink suppression using neural adaptive filters for EEG-based brain computer interface, 24th Ann Int. Conf. IEEE/EMBS, 2002.