Motor Imagery EEG Decoding Based on Multi-Branch Separable Temporal Convolutional Network

Extensive researchers have done on classifying of motor imagery (MI), and the current findings indicate that it is possible to achieve relatively accurate classification decoding of the two-class of MI. However, with the increase in EEG-based MI classification tasks, the four-class motor imagery classification decoding task still requires further research. To solve this problem, we design a hybrid network (SMT) that integrates initial spatiotemporal convolution, multi-branch separable convolution (MSC), and temporal network convolution (TCN) to decode EEG signals. Specifically, spatiotemporal convolution module is used to capture both spatio-temporal features. We design a MSC to capture temporal features at multiple scales. High-level temporal features are captured by the TCN. The SMT was evaluated on the public benchmark BCI Comparison IV 2a and 2b datasets, the results showed that the classification accuracies for the 2a and 2b datasets were 82.18% and 87.67% respectively, with kappa values of 0.761 and 0.738. The results indicate that the SMT surpasses the performance of most existing methods.