Function Classification of Eeg Signal for Human Computer Interface

This paper proposes a modified algorithm for function classification of left and right hand imagery movements obtained from EEG signal. The Electroencephalogram (EEG) is the signal acquired from human brain to monitor and identify human actions to different stimuli. The data was obtained from BCI competition III (b) 2003, acquired by Graz University of Technology. The EEG recorded was sampled with 125 Hz and was filtered between 0.5 and 30Hz.

The features were extracted using Discrete Wavelet Transform(DWT). To obtain precise detail information, the EEG signal was processed with dimensionality reduction techniques used as (i) Singular Value Decomposition & (ii) LDA. The Support Vector Machines (SVM) was used for optimal classification of each motor movement. The result for binary class SVM was at the accuracy level of 100% . The results established the accuracy of Singular Value Decomposition as the best tool to identify the imagery movements.

Introduction

Feature extraction and function classification has always been the challenging task in EEG signal. The EEG signal gives the detail information about electrical activity in the brain. It provides an alternate form of communication for people with disability. Our work had focussed on reducing the complexity of the information and on the other hand, retaining the vital information fromC3 and C4 electrode placement. C3 and C4 is the integral part of delivering the sensorimetric information from the brain. The EEG is obtained from 10-20 International Standard electrode placements [11] on the surface of the skull. The positions C3 and C4 are the regions to provide theta rythms. In our proposed work, the left and right hand motor imagery movements had been classified.

The extensive research on Feature extraction and function classification had been presented with great success. But still, complexity handling remains the major issue in classification of EEG signal. Xiao-Dong ZHANG, et.al [2] had presented the algorithm for controlling the prosthetic hand . The EEG signal was analyzed based on multiple complicated hand movements. The author concluded that the claasification obtained from Support Vector Machines was much better as compared to ANN. Andrews S. et al [21] had presented the singular value decomposition (SVD) for reduction of noise and data dimension. The experimental results gave a very low false accept rate (FAR) and false reject rate (FRR) and a near negligible equal error rate (EER) of 2.91%.

Sachin Garg et al [22] demonstrated the use of wavelet transform for feature extraction of EEG signal. The author asserted that after extracting the coefficients, it was remarkably easier to calculate the statistical parameters of the EEG signal. The another author, Ashwini Nakate et al [24] had also advocated the use of discrete wavelet transform technique to decompose the EEG signal. Priyanka Khatwani et al [25] presents DWT technique for denoising the EEG signal data. Rajesh Singla et.al [26] presented the motor imagery movement of wrist, wrist rotation clockwise/anticlockwise elbow and ankle backward/forward. It had been advocated that DWT technique was best suited for feature extraction of EEG signal.

Abdulhamit Subasi et.al [27] presented the comparison between different techniques used to manipulate the EEG signal data. Principal Component Analysis (PCA),Independent Component Analysis (ICA) and Linear Discriminate Analysis (LDA) was used to reduce the dimensionality of signal. Siuly et. Al. [28] proposed the statistical algorithmto correctly classify the function of EEG signal. Thanh et. Al [29] had analyzed EEG signals using Type-2 fuzzy logic method. The results had showed the low computation cost with good accuracy.

A.B.M. Hossain et al [30] had proposed the Probabilistic Neural Network algorithm for the optimal function classification of EEG signal. The author had claimed the accuracy rate of approximately 99.7 %. M.A. Hassan et al [31] work was focused on modification of back propagation neural network for EEG signal. The classification rate was in the range of 97 to 100 % accuracy. It had also concluded that that time domain features extracted from EEG were more reliable for function classification.

The paper is organized as follows (Figure 1.1): Section II is the Data Acquisition. It discusses the detailed information of database of Left and Right Hand Imagery movements. The section III is focussed on feature extraction using DWT. The section IV is dealing with dimensionality reduction techniques. Further, section V discusses the identification of motor movements. Lastly, section VI validate our proposed algorithm. Whereas Swction VII concludes the work.

Data Acquisition

The database was obtained from Graz Univ. of Technology (BCI Competition III(b),2003). The signals, left and right hand imagery movement were pre-processed to eliminate artifacts from various noises (biosignals/ external). Three electrodes (C3, Cz, and C4) were placed to record the EEG data with a sampling frequency of 125 Hz. The bandpass-filtered were used with frequency range between 0.5Hz and 30 Hz, and a notch filter at 50Hz was also used to remove the artifacts.

The dataset was recorded from a normal subject (female, 25y). The subject was given no information about the recording. The comfortable chair with arm set was provided. The task was to acquire imagined left/ right hand movements. The experiment consists of 7 runs with 40 trials each. Each trial is of 9s length. After initial rest of 2 sec., the recording had been initiated for the respective motor movements. The trials were then selected for random training and testing to classify the imagery movements. In our proposed work, C3 and C4(placement of electrodes) were considered for further analysis.[16]

The figure 1.3 is the flow diagram of feature extraction using DWT. The three motor movements were dispensed to refine the coefficients using Symlet at the decomposition level of ‘3’ [38], so no useful information must not diminished. Using Symlet, the features extracted were near symmetric and have the least asymmetry. The associated scaling filters are near linear-phase filters.

Dimensionality Reduction Technique

For further analysis, the dimensionality reduction, Singular Value Decomposition (SVD) and Linear Discriminant Analysis (LDA) was implemented to increase the computational efficiency of the proposed algorithm. It was useful to remove any unrelated and redundant features from the coefficients extracted from DWT. The SVD theorem is given by:

Xnxp= UnxnSnxpVTpxp (1)

The column shows the three files and corresponding to three different imagery movements in succession. Each row, gave the output of an individual subject at trial. The sample of extracted features from Symlet as given below:-

Similarly, Linear Discriminant Analysis was employed. The mean of different classes were calculated to define the measure of separation between the respective imagery movements. To define the maximal difference between imagery movements, the Fisher Discriminant Analysis was implemented. It provides the linear function for maximal projection (Table 1.2).The table illustrates for three different EEG signal viz. left, right and at rest. The total of three subjects were taken for training.

Function Classification using SVM

In order to realize the motor imagery movements of left and right hand, both Binary and Multi Class SVM was implemented. The figure 1.3(b) defined the organization of SVM. To achieve optimization for hyperplane, the Gaussian kernel function was used. The function is best suited random variability in the EEG signal.

The features that could not be classified became the support vectors and hence, the efficiency of the classifier increases. The total of three subjects were took for training the support vectors and the the two subjects were implemented for testing. As shown in Table 1.3 (a) to (e), the two different movements were tested with individual subjects. The RST was investigated with the other two imagery movements i.e. RM and LM. Similarly, LM was investigated with RM. The rate of correct class in training set was 100% for the each case as discussed below. The training was done with ten files and testing is done for five files different other than the files that were used for training.

However, if all the three imagery motor movements were considered, then the rate of classification is approx. 97% (SVD-Multi Class SVM).The method proposed by Hassan et al had classified right and left hand imagery movement to assert the accuracy of 100%. The work carried by Riheen at al showed the accuracy of 97%, followed by other studies.

Conclusion

In the proposed algorithm, the more robust and computationally efficient algorithm had been implemented for function classification of EEG signal. We projected the motor imagery of left and right hand movement using LIBSVM Support Vector Machines as a classifier tool. Using the coefficients obtained from wavelet, the result was implemented for dimensionality reduction ( LDA and SVD). Our result obtained, were approximately accurate in the range of 92% to 100%.

The highest classification accuracy of 100 % is obtained from SVD as a dimensionality reduction tool. However, the work may be extended by accumulating more imagery movements. The subjects for training and testing may further be added to testify for larger database. We had compared our work for LDA and SVD. More techniques may be explored to strengthen the accuracy of our proposed method. Furthermore, Multi class SVM based on decision tree can also be used to increase the computation efficiency of identification of imagery movements using EEG signal.