Information based source number estimation for probabilistic common spatial pattern in motor imagery BCI system

Common Spatial Pattern (CSP) is one of the popular and effective methods for discriminating two class electroencephalogram (EEG) measurements. Its probabilistic counterpart by resolving the problem of overfitting as the main limitation of CSP attracted much attention, especially in the motor imaginary based brain computer interface (BCI) applications. Since the computational efficiency is a paramount issue in real-time EEG classification, in this paper, assuming additive isotropic noise, maximum a posteriori (MAP)-based iterative updating algorithm is applied. However, the performance of this algorithm depends on the model size which must be predetermined. To this end, three information based source number estimations including Akaike Information Criterion (AIC), Minimum Description Length (MDL) and Bayesian Information Criteria (BIC) were used. The experimental results on a publicly available Ilia dataset from BCI competition III demonstrate higher classification accuracy compared to CSP and existing Tikhonov regularized CSP (TR-CSP) models. In addition, a significant decrease in run-time was achieved using the proposed method.