Hierarchical Spectral-Temporal Feature Learning for Motor Task Recognition in Brain Computer Interfaces

Feature extraction is an indispensable step in classification of electroencephalography (EEG) signals for motor task recognition. This paper proposes a novel method for motor imagery brain-computer interface (BCI) EEG signals classi¬fication based on spectral-temporal common spatial patterns feature extraction and deep convolutional neural network. Unlike other existing methods, the proposed method is built upon both frequency and time domain features of the EEG signals. These features are used as input to the proposed convolutional neural network for a higher level feature extraction and classification. The proposed network is comprised of convolutional and fully- connected layers to hierarchically learn the saliency in the signal. Several experiments are conducted on standard datasets taken from BCI competitions to assess the performance of the proposed method and to compare it with those of the other existing methods. The results exhibit a significant improvement in classification accuracy for a number of subjects when using the proposed method.