A WAVELET-BASED APPROACH FOR THE EXTRACTION OF EVENT RELATED POTENTIALS FROM EEG M. Fatourechi 1,2 , S.G. Mason 2 , G.E. Birch 1,2 and R.K. Ward 1 1 Dept. of Electrical & Computer Engineering, University of British Columbia, Vancouver, BC, Canada 2 Neil Squire Foundation, Vancouver, BC, Canada ABSTRACT Event Related Potentials (ERPs) are of interest to many researchers seeking knowledge about the functions of the brain. ERPs are low-frequency events that are usually obscured in single trial analysis. To visualize these signals; most of the reliable solutions at the present time use the ensemble averages of many single trials. In this paper, a wavelet-based method called Statistical Coefficient Selection (SCS) is used for the extraction of ERPs from EEG signals. Unlike other wavelet- based denoising methods, the current method does not focus on the wavelet coefficients of the signal itself. Instead, it selects the coefficients based on the statistical study of trials from training data set. Simulation results show the superiority of the proposed SCS method in extracting ERPs in comparison with other filtering approaches. 1. INTRODUCTION Event Related Potentials (ERPs) are parts of the EEG signal that are time-locked to a sensory, motor, or cognitive process and therefore provide an electrophysiological window onto brain function during cognition. They have a characteristic pattern that is more or less reproducible under similar experimental conditions [1]. The origin of an ERP might be an external stimulator (for example a flash light) or it can be initiated internally such as a result of making a movement. In the literature, two significant applications make use of ERPs: diagnosing neurological disorders [2] and development of brain- computer-interface (BCI) systems [3]. For both applications, many methods that extract ERPs from the background EEG have been explored. The main problem of ERP extraction is that the amplitude of an ERP is much smaller than that of the background EEG. This makes its detection very hard in single trial analysis. Instead of extraction of ERPs from single trials many methods have focused on the extraction of ERPs from ensemble averages of several single trials (i.e. data segments including the pre- and post-stimulus activity are averaged). Since ERPs are time locked to the stimulus, it is assumed that their contribution during the averaging process will add up while the ongoing EEG and unrelated components are attenuated. This will result in higher Signal to Noise Ratio (SNR). Since ERPs are non-stationary, time-invariant approaches such as Fourier Transform are not likely to give acceptable results. On the other hand, the joint time–frequency resolution obtained by the wavelet transform makes it a good candidate for the extraction of the details as well as the approximations of time-varying, non-stationary signals [4]. For the effective extraction of ERPs from the background EEG using wavelet transform, we need a strategy which 1) chooses the coefficients associated with the ERP and, 2) considers the fact that ERPs vary significantly from time to time [5]. Several methods have been proposed for extracting ERPs from the background EEG with various success. Many researchers use level-dependant thresholding schemes by defining a criterion for the selection of the threshold of each level [6-7]. For example, in [6], the authors apply a level dependant threshold based on the median absolute deviation of wavelet coefficients in each level. In [7], the authors report good noise reduction in ERPs simply by discarding three upper level bands. Many researchers manually select the wavelet coefficients assumed to be associated with an ERP [8-9]. For example in [8], the authors select Visual Evoked Potentials (VEPs) based on a single wavelet coefficient in the delta band of the EEG. Also in some approaches the wavelet coefficients are selected based on a similarity criterion between the coefficients associated with the waveform and the coefficients associated with the grand ensemble average of all the test waveforms [10]. The main problem with the methods that select coefficients manually is their vulnerability to the human error. These methods also do not take the time varying property of ERPs into consideration. On the other hand, methods based on the threshold selection or a measure of similarity with the grand ensemble averages cannot filter the coefficients associated with the background EEG effectively, because many of these coefficients lie in the same frequency spectrum of ERPs. Therefore, the proposed selection method should not only consider the energies of the coefficients which are attributed to ERPs, but it should also consider their variations throughout the time. This paper proposes a new scheme, which selects the individual coefficients associated with an ERP automatically. The proposed method attempts to overcome the vulnerabilities of the previously mentioned methods. To be more specific, in this method, the wavelet coefficients that are sought have high amplitude values (the ones with high energy in the case of orthogonal wavelets) and low amplitude variance over many trials. In other words, the current method does not focus on the wavelet coefficients of the signal themselves. Instead, it selects the coefficients based on the statistical study of training data set. The organization of the paper is as follows: the focus of Section 2 is on the wavelet analysis. In Section 3, the proposed II - 737 0-7803-8484-9/04/$20.00 ©2004 IEEE ICASSP 2004 ➠ ➡