ASTGNN: adaptive spatio-temporal graph neural network for motor imagery recognition

Motor imagery electroencephalogram recognition is a key area in brain-computer interfaces, with applications in human-computer interaction, rehabilitation, and virtual reality. Traditional methods often overlook the brain's topological characteristics and the correlations between EEG channels, resulting in suboptimal decoding. To address this, we propose an adaptive spatio-temporal graph neural network (ASTGNN), which constructs a brain graph topology and adaptively learns the topological adjacency matrix, exploring both common and individual electrode connections. The spatial features are extracted using a graph convolutional network, while a gated position-aware self-attention mechanism captures movement information and global dependencies, enhancing temporal feature extraction. Experiments show that ASTGNN significantly improves recognition, achieving an average accuracy of 87.36%.