SincMSNet: A Sinc filter convolutional neural network for EEG motor imagery classification.

Motor imagery (MI) is a commonly employed experimental paradigm in brain-computer interfaces (BCIs). Nevertheless, the decoding of MI-EEG using convolutional neural networks (CNNs) is still deemed challenging due to the variability of individuals and the non-stationarity of EEG signals. 
Approach: We propose an end-to-end convolutional neural network (CNN) called SincMSNet for MI decoding. SincMSNet utilizes the Sinc filter to extract subject-specific frequency band information, and mixed-depth convolution to extract multi-scale temporal information for each band. Spatial convolutional blocks are then used to extract spatial features, while the temporal log-variance block is used to acquire classification features.
Main results: We assessed SincMSNet on two MI datasets and compared it to several state-of-the-art MI decoding methods. Our results demonstrate that SincMSNet surpasses the benchmark methods, achieving an average accuracy of 80.70% and 71.50% in the four-class and two-class of hold-out classification, respectively. Furthermore, the acquired filter sets exhibit the network's capability to provide higher relevance to individual features. 
Significance: SincMSNet is a promising method to enhance the performance of MI-EEG decoding, and is available for use through the source code at https://github.com/Want2Vanish/SincMSNet.