A Spatial Feature-Based Adaptive Technique to Match Models for Motor Imagery EEG Signals

In recent years, deep learning has been widely used in the field of motor imagery brain-computer interfaces, facilitating great improvements in classification accuracy in this field. However, due to the specificity of EEG data, the same model has uneven learning effects on different subject features. The difference in learning effects means that training with only one model results in signals that are not correctly identified in some subjects, which is not conducive to the diffusion of motor imagery EEG signal classification techniques. During the study, we found that each subject data has a suitable model for it, and assigning a suitable model to the data for training can better ensure a better classification accuracy for each subject. In this paper, we propose a technique for automatic matching of motion imagery EEG signals with recognition models. This is a technique for extracting spatial features of motor imagery EEG signals based on graph convolution. Because the spatial features are highly stable and all time steps are included in the extraction of spatial features, the features extracted are macroscopic in nature. Based on the big data screening, we selected four excellent motor imagery EEG signal classification models and put the data of subjects who performed well on these models into the graph convolution model for supervised learning. The completed learning graph convolution model can be well trained to assign appropriate classical models of EEG signals to different subjects based on spatial features. After using this method, we achieved an average classification accuracy of 88$\%$ on the BCI Competition IV dataset 2a.