ThinICA-CSP algorithm for discrimination of multiclass motor imagery movements

Abstract

This paper presents a ThinICA-CSP1 algorithm for discrimination of multiclass motor imagery (MI) movements for Brain Computer Interfacing (BCI) applications. This algorithm performs a joint approximate diagonalization of the second and higher order statistics of the observations with the aim of identifying the relevant independent components of the EEG signals and their corresponding spatial filters. In order to speed up the convergence, the algorithm is initialized from the multiclass Common Spatial Pattern (CSP) filter matrix. This helps the ICA algorithm to find the closest solution to the problem. The algorithm was tested on BCI competition IV dataset 2a and the obtained performance was compared with two existing methods. An improvement in classification performance is observed using the ThinICA-CSP algorithm.