Spectral feature extraction from EEG based motor imagery using common spatial patterns

Noninvasive MI-BCI (Motor imagery Brain computer interface) allows people to communicate and control external devices through EEG signals. Feature extraction is an important bloc to obtain a reliable classification accuracy of motor imagery tasks. Common spatial patterns (CSP) is a frequently used algorithm for EEG feature extraction, but its performance relies on the subject-specific frequency band. This paper shows the powerful effect of CSP in discrimination between two classes of motor imagery (left and right hand). Using projected training data on CSP this study demonstrates that subject-specific frequency bands can easily be determined. The experimental results obtained using two public EEG datasets (BCI competition IV dataset 2a and 2b) demonstrate that the subject-specific frequency bands extracted in offline analysis phase using CSP help improve the classification performance of MI-BCI.