Multi-scale temporal electroencephalography signal classification using a dynamic graph attention network framework

Electroencephalography (EEG)-based decoding is critical for brain–computer interface (BCI) applications, especially in motor imagery (MI) tasks. However, traditional convolutional neural networks often struggle in capturing long-range temporal dependencies due to limited receptive fields. To overcome this, a hybrid framework that combines temporal convolutional networks (TCNs) with graph attention networks (GATs) is proposed. This framework effectively models multi-scale temporal dynamics and inter-temporal dependencies in EEG signals. Herein, the TCN module is used for the extraction of rich short- and long-term temporal features. This is followed by dynamic graph construction, where each node represents a time step, and edge weights are adaptively learned. Then, the GAT module propagates temporal information by emphasizing the most relevant connections. In addition, the temporal slicing and shuffling data augmentation strategy is introduced to improve generalization. Evaluations on three benchmark motor imagery electroencephalography datasets, namely, BCIC-IV-2a, BCIC-IV-2b, and High Gamma Dataset (HGD), are carried out. The evaluations are performed using standard evaluation metrics, namely, classification accuracy and Cohen’s kappa coefficient across subjects. Accordingly, the experimental outcomes reveal that the proposed framework consistently outperforms state-of-the-art baselines across the three employed datasets. The findings highlight the importance of multi-scale temporal feature fusion using graph attention mechanisms for effective motor imagery EEG signal decoding.