Enhancing the performance of a deep convolutional neural network model for motor imagery classification using EEG channel-wise attention module

The classification of motor imagery-electroencephalography (MI-EEG) is a growing research field in brain-computer interface, which allows people with motor disabilities to communicate with the outside world through assistive devices. Although deep learning-based models have revolutionized MI-EEG decoding, dealing with the MI-EEG signals remains challenging due to the signals being non-stationary, containing noisy signals, and having a low signal-to-noise-ratio. This study proposes to employ a novel EEG channel-wise attention module (ECWAM) in a deep convolutional neural network (deep CNN) to enhance the accuracy of MI-EEG decoding. The proposed method calculates the channel score for each mu band EEG channel and amplifies the prominent EEG channels based on their channel scores. The proposed method is evaluated on 54 subjects, binary class MI dataset from the Korea University EEG dataset. Additionally, the proposed method is compared with the conventional channel-wise attention module mentioned in the literature. The results for the hold-out analysis outcomes suggest that the proposed deep CNN with ECWAM has statistically improved the average classification accuracy of the baseline deep CNN model from 63.96% to 68.98%, with<i>p</i>-value = 0.02 for the subject-specific MI classification. Further, the scalp map of the EEG channel ranking obtained by the proposed method and the conventional channel-wise attention module mentioned in the literature is also compared. The results of the comparison show that the proposed method yields a higher channel ranking in the brain's motor cortex region, which is the primary contributing area for MI activity.