Motor Imagery EEG Signal Classification Scheme Based on Autoregressive Reflection Coefficients

In brain-computer interface (BCI) applications, classification of electroencephalogram (EEG) data for different motor imagery (MI) tasks is a major concern. In this paper, an efficient MI task classification scheme is proposed based on autoregressive (AR) modeling of the EEG signal. From given EEG recording, after some basic preprocessing operations, the processed EEG data of each channel is windowed into several frames and AR parameters are extracted using least-square Yule-Walker method. Considering the reflection coefficients from the autoregressive modeling, a set of features is extracted from the average of the coefficients of the specified frames. In order to reduce the dimension of the proposed feature matrix, principal component analysis (PCA) is employed. For the purpose of classification, train and test sets are formed based on leave one out cross validation and then linear discriminant analysis (LDA) based classifier is used. Simulation is carried out on publicly available MI dataset IVa of BCI Competition-III and a very satisfactory performance is obtained in classifying the MI data in two classes, namely right hand and right foot MI tasks. Proposed classification scheme not only offers significant reduction in feature dimensionality but also provides satisfactory classification accuracy.