Subject specific parameter selection for the EEG classifier using common spatial patterns

Common Spatial Patterns (CSP) is a popular feature extraction method for single-trial multichannel EEG classification in BCI applications. CSP generates spatial filters that optimize the distinction between two classes. The performance of CSP is highly dependent on how many filters are selected. However, in most BCI applications the number of filters is fixed. In this work we demonstrate that the optimal number of filters is highly dependent on subject, task, and task timing. We show that optimization of these parameters on per subject, task, and timing basis improves classification accuracy.