Efficient Channel Selection Approach for Motor Imaginary Classification based on Convolutional Neural Network

Brain Computer Interface (BCI) may be the only way to communicate and control for disabled people. Someone's intention can be decoded from their brainwaves during motor imagery action. This can be used to help them control their environment without making any physical movement. To decode someone's intention from brainwaves during motor imagery activities, machine learning models trained on features extracted from the acquired EEG signals have been used. Although the technique has been successful, it has encountered several limitations and difficulties especially during feature extraction. Moreover, many current BCI systems rely on a large number of channels (e.g. 64) to capture spatial information which are necessary during training a machine learning model. In this study, Convolutional Neural Network (CNN) is used to decode five motor imagery intentions from EEG signals obtained from four subjects using 64 channels EEG device. A CNN model trained on raw EEG data managed to achieve a mean classification accuracy of 99.7%. Channel selection based on learned weights extracted from a trained CNN model has been performed with subsequent models trained on only two selected channels with higher weights attained a high accuracy (average of 98%) among three participants out of four.