Spatial-Temporal Mamba Network for EEG-based Motor Imagery Classification

Abstract

Motor imagery (MI) classification is key for brain-computer interfaces (BCIs). Until recent years, numerous models had been proposed, ranging from classical algorithms like Common Spatial Pattern (CSP) to deep learning models such as convolutional neural networks (CNNs) and transformers. However, these models have shown limitations in areas such as generalizability, contextuality and scalability when it comes to effectively extracting the complex spatial-temporal information inherent in electroencephalography (EEG) signals. To address these limitations, we introduce Spatial-Temporal Mamba Network (STMambaNet), an innovative model leveraging the Mamba state space architecture, which excels in processing extended sequences with linear scalability. By incorporating spatial and temporal Mamba encoders, STMambaNet effectively captures the intricate dynamics in both space and time, significantly enhancing the decoding performance of EEG signals for MI classification. Experimental results on BCI Competition IV 2a and 2b datasets demonstrate STMambaNet's superiority over existing models, establishing it as a powerful tool for advancing MI-based BCIs and improving real-world BCI systems.