A Multi-Band Self-Attention Network for Motor Imagery Classification

Brain-computer interface (BCI) systems create a novel communication method between humans and machines by translating human thoughts into actionable commands to control external devices. Motor imagery (MI) electroencephalogram (EEG) signals have significant applicability in various medical and non-medical industries, including stroke rehabilitation, wheelchair control, and drone operation. However, the practical application of EEG remains limited by the decoding performance and generalization ability of MI signalsThis study introduces a multi-branch self-attention network for motor imagery (MI) signal classification. Each branch independently processes EEG signals decomposed into distinct frequency bands through convolutional neural networks (CNNs) and multi-head self-attention (MHA) mechanisms, enabling the extraction of both fundamental and discriminative spatial-temporal features. To further capture dynamic temporal dependencies, long short-term memory (LSTM) networks are integrated. We systematically evaluate three signal decomposition ensemble empirical mode decomposition (EEMD), wavelet packet decomposition (WPD), and brain rhythm-based decomposition-to optimize feature representation. Extensive experiments on the BCI Competition IV 2a dataset demonstrate state-of-the-art performance, with subject-dependent and subject-independent accuracies of 84.04% and 71.67%, respectively. Comparative analyses against benchmark models (EEGNet, EEGTCNet, ShallowConvNet, etc.) validate the superiority of our approach in classification accuracy and generalization capabilityClinical relevance- This study investigates the methods for decoding motor imagery EEG signals and establishes the positive role of each module in classification. The improvement in accuracy can lead to better outcomes in medical applications such as controlling prosthetics, wheelchairs, and stroke rehabilitation.