Cross-Frequency Graph Neural Networks for EEG Based Motor Imagery

Nowadays, the application of deep learning has significantly advanced motor imagery research, which is one of the major topic of brain-computer interfaces (BCIs). Traditional approaches primarily focus on extracting temporal features or leveraging spatial topological information from multi-channel EEG signals. However, these methods often overlook cross-frequency spatial interactions, which are crucial for capturing comprehensive neural dynamics. To address the issues mentions above, we propose a new model, Cross-Frequency Graph Neural Network (CF-GNN), which constructs graphs across different frequency bands to model both spatial and frequency-domain relationships. By applying a sliding window mechanism along the frequency dimension, our method adaptively captures cross-frequency spatial dependencies. Experimental results demonstrate that CF-GNN achieves classification accuracies of 80.24% on a four-class motor imagery dataset, gaining competitive results comparing to the baseline methods. The experiment results show the effectiveness of graph learning in capturing cross-frequency spatial interactions, offering a promising direction for enhancing the motor imagery classification accuracy.