Pattern recognition of motor imagery EEG using wavelet transform Pattern recognition of motor imagery EEG using wavelet transform

This paper presents a novel effective method for ABSTRACT feature extraction of motor imaginary. We combine the discrete wavelet transform (DWT) with Brain-computer interface (BCI) provides new autoregressive model (AR) to extract more useful communication and control channels that do information for non-stationary EEG signals. Apply-not depend on the brain's normal output of ing this method to analyze the Graz dataset for BCI peripheral nerves and muscles. In this paper, competition 2003, we achieved the classification we report on results of developing a single accuracy of 90.0%. trial online motor imagery feature extraction method for BCI. The wavelet coefficients and


INTRODUCTION
was displayed as cue.At the same time the subject Left and right hand movement imagery can modify was asked to move a bar into the direction of the cue. the neuronal activity in the primary sensorimotor The feedback was based on AAR parameters of chanareas, leading to the changes of the mu rhythm and nel #1 (C3) and #3 (C4), the AAR parameters were beta rhythm.BCI requires effective online processcombined with a discriminant analysis into one outing method to classify these EEG signals in order to put parameter.construct a system enabling severely physically dis- The recording was made using a G.tec amplifier abled patients to communication with their surroundand a Ag/AgCl electrodes.Three bipolar EEG chanings [1-4].

Procedure
The flow chart of processing single-trial motor imagery EEG is shown as in .First, the time window was used to filter the data in temporal domain in order to get the segment that contained the most obvious difference between the two motor imagery tasks.Then EEG signals were decomposed into the frequency sub-bands using DWT and a set for statistical features was extracted from the sub-bands to represent the distribution of wavelet coefficients according to the characteristics of motor imagery EEG signals.Also the sixth-order AR coefficients of segmentation EEG signals were estimated using Burg's algorithm.Next, the combination features of wavelet coefnels (anterior '+', posterior '-') were measured over ficients and the AR coefficients were used as an input C3, Cz and C4 [ ].The EEG was sampled vector.Finally linear discriminant analysis (LDA) with 128Hz, it was filtered between 0.5 and 30Hz.based on mahalanobis distance was utilized to clas-Similar experiments are described in [6].
sify computed features into different categories that The experiment consists of 7 runs with 40 trials represent the left or right hand movement imagery.each.All runs were conducted on the same day with several minutes break durrng experiment.One half of 2.4.Feature extraction using discrete wavelet the datasets are provided for training; others are for transforms evaluating the performance of the system.
Classic Fourier transform has succeeded in stationary signals processing.However, EEG signal con-

Feature consideration
tains non-stationary or transitory characteristics.Central brain oscillations in the mu rhythm in the Thus it is not suitable to directly apply Fourier transrange of 7-12Hz and beta above 13Hz bands are form to such signals.The wavelet transform decomstrongly related to sensorimotor tasks.Sensory stimposes a signal into a set of functions obtained by ulation, motor behavior, mental imagery can change shifting and dilating one single function called the functional connectivity cortex which results in an mother wavelet [10 11].Continuous wavelet transamplitude suppression or in an amplitude enhanceform is given by ment .This phenomenon was also called eventrelated desynchronization (ERD) and event-related synchronization (ERS) [7 8].Left and right hand movement imagery is typically accompanied with ERD in the mu and beta rhythms and has the characteristic of contralateral dominance.
Where (t) is the mother wavelet, is the scale The power spectrums on C3 and C4 of the training parameter and is the shift parameter.In principle set are shown in .It indicates that the power the CWT produced an infinite number of coefficients, spectrums mainly distribute in the range of 8-13Hz thus it provides a redundant representation of the sigand 19-24Hz.In addition, the power of mu and beta nal.rhythms evoked by right hand movement imagery is The DWT provides a highly efficient wavelet replower than that of left hand movement imagery for resentation that can be implemented with a simple channel C3, and it is contrary for channel C4 which is recursive filter scheme and the original signal reconconsistent with the principle of contralateral domistruction can be obtained by an inverse filter.The pro-

Statistical features wavelet Coefficients
Coefficients of autoregressive model EEG temporal filter linear discriminant analysis cedure of multi-resolution decomposition of a signal trum and too high tends to introduce spurious peaks.
x[n] is schematically shown in .
Here order six was used based on the suggestions [9].The number of levels of decomposition is chosen Then the Burg's method was used to estimate the on the basis of the dominant frequency components AR coefficients.This method is more accurate and of the signal.According to the motor imagery EEG yields better resolution without the problem of specsignals itself, we chose the level of 4 and the wavelet tral 'leakage' as compared to other methods such as of Daubechies order 10.As a result, the EEG signal is Levison-Durbin as it uses the data points directly.In decomposed into the details D1-D3 and approxima-addition, the Burg's method can minimize both fortion A3.The ranges of different frequency band are ward and backward error.shown in .
Next the AR coefficients were computed and we The extracted wavelet coefficients show the distri-got six coefficients for each channel, giving a total of bution of the motor imagery signal in time and fre-12 AR coefficients features for each EEG segment for quency.It can be seen from the table that the compo-a motor imagery task.nent D3 decomposition is within the mu rhythm, D2 is within the beta rhythm.Statistics over the set of 2.6.Linear discriminant analysis (LDA) As to the LDA method, assume that each data ele-(3) Average power of the wavelet coefficients.
ment s has m features.Then, an element s is one i i These features represent the frequency distribupoint in a dimensional feature space.The number of tion and the amount of changes in frequency distribuexamples is n , each example is assigned to one of two tion.Thus 12 statistical features of wavelet coefficlasses C={0,1}; Then, S is a matrix of size n×m, and cients are obtained for two channels.
C is a vector of size n.N .And N are the number of 0 1 elements for class 0 and 1, respectively.

Feature extraction using autoregressive
The mean of each class c is the mean over all s c i model with i being all elements with in class c .The total EEG signal can be considered as the output of a linear mean of the data is filter driven by a white noise.This filter, referred to as AR, is a linear combination of the previous output itself.A zero-mean, stationary autoregressive process of order p is given by Where p is the model order, x(n) is the signal at the sampled point n, a (i) is the AR coefficients and (n) p is a zero-mean white noise.In application, the values of the a (i) have to be estimated from the finite samp ples of data x(1),x(2),x(3),…,x(N).
The first important things involved in using AR model is determining the optimal AR model order since too low a model order tends to smooth the spec-

Figure 4 .
Figure 4. Flow chart of the data processing.

Frequency
wavelet coefficients were computed so as to reduce LDA is one of the most effective linear classification the total dimension of the feature vectors.The statismethods for brain-computer interface, and it requires tical features of each sub-band are as follows: fewer examples for obtaining a reliable classifier out-(1) Mean of the absolute values of the coefficients.put [12].(2) Standard deviation of the coefficients.

Figure 5 .
Figure 5. Decomposition of DWT; h[n] is the high-pass filter; g[n] is the low-pass filter.

Table 1 .
Frequencies correspond to different levels of deposition for daubechies order 10 wavelet with a sample rate 128HZ.