CNN-based Classification of Left and Right Hand Motor Imagery

Brain-Computer Interface is one of the most promising subjects of study in the field of medicine today, and its most important theoretical basis is Motor Imagery (MI) classification. Due to the complex and disturbing components of the Electroen-cephalography (EEG) signal, the accuracy of MI classification has been limited. The proposed study is aimed at the left and right motor imagery classification. The Common Spatial Pattern (CSP) features of multiple frequency bands were extracted to achieve higher accuracy, and Convolutional Neural Network (CNN) was introduced as a classifier to achieve feature redevelopment and classification. The research utilized a relatively commonly used CSP-linear discriminant analysis (LDA) classifier as the baseline. On BCI Competition IV 2b datasets, it finally achieved an average accuracy of 77.00±5.22% for single subject and 69% for cross-subject, which outweigh the baseline methods by 22% and 19% respectively. The experiment also found the best net parameters and data length in this specific work based on a large number of trials.