High accuracy decoding of motor imagery directions from EEG-based brain computer interface using filter bank spatially regularised common spatial pattern method

One of the important requirements of a practical Brain Computer Interface (BCI) system is the ability to establish multiple control commands corresponding to different kinematics of motor imagery. Most of the previous BCIbased motor imagery studies in literature focus on classifying left vs. right hand motor imagery from electroencephalogram (EEG) signals. Very few studies have reported decoding imagined hand movement kinematics from EEG-based BCI. The present study decodes the left vs. right directional information from the Motor Imagery (MI) of dominant hand movement using EEG-based BCI. The proposed method employs common spatial pattern (CSP) and its variants as features to decode imagined (motor imagery) bidirectional hand movements. The direction discriminability of these features are enhanced using the regularisation technique. Spatial regularisation based on electrode positions is also incorporated for comparison. The regularisation methods are applied on overlapped frequency bands and the results are compared. The classifications of extracted features are done using 5-fold cross-validation and Linear Discriminant Analysis (LDA). The study on 15 healthy subjects shows that filter bank-based spatially regularized CSP method (FBSRCSP) offers the highest average classification accuracy of 90% for decoding bidirectional motor imagery of hand movement. This is a substantial improvement in classification accuracy by 26.3% compared to the best method reported in literature on the same dataset.