Enhancing LDA-based discrimination of left and right hand motor imagery: Outperforming the winner of BCI Competition II

Due to the potential applications of Brain-Computer Interfaces (BCI), like producing rehabilitation systems for disabled people, many researches have been aimed at minimizing the error of BCI systems. In this paper, we used left and right hand motor imagery EEG data provided by Graz University of Technology for the BCI Competition II. We attempted to achieve a better misclassification rate while selecting less features compared with various former reported researches on this dataset. We used linear discriminant analysis (LDA) as the classifier due to its low computational cost and previously reported promising results. Furthermore, we investigated what features have major impacts on local or global minimization of the misclassification rate. Also, we briefly assessed the effect of changing window length on the misclassification rate. In this paper first, a set of various statistical, spectral, wavelet-based, connectivity, and chaotic features was extracted from EEG data. Subsequently, an LDA-based wrapper Sequential Forward Selection (SFS) scheme was used for selecting optimum subset of features for each data window. Finally, data windows were classified by LDA. We achieved less misclassification rate using less features compared with previous LDA-based researches and the winner of BCI competition II on the same dataset. Also, the absolute mean of the third-level wavelet detail coefficients (related to μ-band) and the skewness were the two features that together yielded the best local discrimination results.