Motor imagery classification using a novel CNN in EEG-BCI with common average reference and sliding window techniques

A popular research area in electroencephalography (EEG) is a brain-computer interface (BCI), which involves the classification of MI tasks. In this work, we proposed a novel convolutional neural network (CNN)-based method to recognize the motor imagery (MI) activities of left and right hand movements in the EEG-based BCI system. The proposed work focused on improving performance using a common average reference (CAR) filter and a novel CNN model. CAR is used to increase the signal to noise ratio of the EEG signal. We utilize a novel CNN model for EEG-based MI classification, incorporating both local and global features. Initially, we applied the CAR filter on collected EEG data. In the subsequent phase, we utilize the sliding window technique to generate a number of time segments and address the issue of overfitting. Next, we applied short-time Fourier transform (STFT) on the time segments and it provides spectrogram images. Additionally, we extracted the mu and beta bands from spectrogram images and concatenated these bands. Furthermore, the proposed model receives the concatenated images as input. The novel CNN model comprises convolution, a pooling layer, skip connection, concatenation, dropout, and a fully connected layer. The proposed methodology was evaluated on EEG data from the BCI competition IV dataset-2b, with 80% used for training and 20% for testing. Compared to the state-of-the-art models, our novel methodology achieved much better performance with an accuracy of 91.75% and also provided insightful visualizations through gradient weight class activation maps (Grad-CAM) analysis.