SCTrans: Motor Imagery EEG Classification Method based on CNN-Transformer Structure

As Brain-Computer Interfaces (BCI) systems have been developed to transform human intentions into control commands using EEG data, Motor Imagery (MI) BCI is becoming more and more important in BCI paradigms. However, the non-stationarity and substantial inter-subject variability of EEG data make classification difficult. To resolve the mentioned constraints of MI-EEG classification, we proposed SCTrans, a novel model combining CNN and Transformer networks that can be used for MI-EEG classification problems. Shallow and deep features for EEG data extraction using the CNN module and Transformer module respectively. Experiments show that the classification accuracy and F1-score of SCTrans are significantly better than the SOTA models in both subject-independent and subject-dependent manners.