Brain computer interface (BCI) systems translate noisy electroencephalogram (EEG) signals into control commands, with motor imagery (MI) classification being a central research focus. However, MI EEG decoding remains challenging due to the non-stationary nature of brain signals and their complex spatiotemporal patterns. This paper presents Mamba2-CNN, a hybrid architecture that integrates convolutional neural networks (CNNs) with the Mamba2 module based on the Structured State Space Duality (SSD) framework. The CNN component captures local spatial–temporal EEG patterns, while the Mamba2 backbone models long-range temporal dependencies with linear computational complexity, addressing the limitations of purely convolutional models. Comprehensive evaluations on BCI Competition IV-2a, IV-2b, and SEED, demonstrate that Mamba2-CNN achieves consistently competitive or state-of-the-art performance across both MI and emotion recognition tasks. Additional ablation and efficiency analyses indicate that the proposed model maintains a favorable balance between accuracy, latency, and memory usage, highlighting its practical relevance for real-time and resource-constrained BCI applications.
Xia et al. (Thu,) studied this question.