An Improved Schema of Brain-computer Interface Based on Motor Imagery with Eye Movement

Abstract

Abstract Motor imagery (MI) can induce electroencephalogram (EEG) and realize human-computer interaction, but this kind of interaction has poor robustness and low stability. To solve these problems, we improved MI paradigms with eye movement and proposed convolutional neural network classification models based on attention mechanism. We conducted a comparative study to evaluate the performance of MI with different eye movement patterns, i.e. smooth pursuit MI (PMI), saccade MI (SMI) and pure MI. The differences between Squeeze-Excitation (SE) module and Convolutional Block Attention Module (CBAM) module were also explored. The results of power spectral density (PSD) showed that PMI paradigm induced the most significant event-related desynchronization (ERD) phenomenon and the average classification accuracy for PMI signals was also the best in the three paradigms. The combined EEGNet and SE framework achieved an average classification accuracy of 90.77%, which performed better than the model without attention module. PMI can optimize attention allocation of subjects, assist in the construction of motion thinking, and improve the quality of MI signals. EEGNet with SE module showed improved classification performance.