Graph Attention Network for Motor Imagery Classification

Motor Imagery (MI) classification is a crucial task in brain-computer interfaces, enabling the translation of neural activity into actionable commands for various applications. Despite growing interest in EEG-based identification, the field faces challenges such as temporal variability and complex spatial dependencies in electroencephalogram (EEG) data. This paper presents a novel method, referred to as GAMI, which integrates spatial convolution filters and leverages the complex structure of a graph formed by connected electrodes. Specifically, our method uses spatial convolution filters to improve the representation of EEG signals and potentially reduce noise during data acquisition. These refined features are then processed by Graph Attention Network (GAT) layers, which leverage the graph structure of EEG electrodes and their connections to capture complex spatial dependencies. The combination of convolutional filters and GAT layers provides a robust solution for extracting meaningful features from MI-related EEG data. Additionally, our approach integrates data alignment techniques to employ transfer learning, addressing the challenge of limited data for individual subjects. Intensive experimental results on two benchmark datasets demonstrate the superiority of the GAMI model over traditional methods, showcasing improved accuracy and robustness in MI classification tasks.