Interpretable dual-layer state space net for capturing fine-grained temporal features in EEG-based motor imagery BCIs

Motor Imagery (MI) is a cognitive process central to Brain-Computer Interface (BCI)-driven motor rehabilitation. However, existing high-performance MI decoding models are often ”black boxes” that achieve high accuracy at the cost of sacrificing the fine-grained temporal resolution crucial for neuroscientific discovery. This paper introduces an interpretable deep learning framework centered on DLSSNet, to capture and interpret the fine-grained brain activities of MI. Its fully convolutional architecture preserves complete temporal information, while novel Task-Relevance Evaluation (TRE) and Attention-Based Embedding (ABE) modules effectively extract discriminative information from these high-resolution features. Validated on two public benchmark datasets (BCI2a and HGD), DLSSNet significantly outperformed state-of-the-art models with average accuracies of 78.51% and 95.49%, respectively, while fully preserving the fine-grained feature sequences for subsequent analysis. More importantly, our interpretability protocol reveals that DLSSNet learns a set of prototypical neural states that are functionally linked to specific MI tasks. The visualized topographies of these states are highly consistent with known Sensorimotor Rhythm (SMR) patterns, and we further demonstrate that these patterns are stable across multiple subjects. In conclusion, by bridging the gap between high classification accuracy and neurophysiological plausibility, this paper provides a powerful new framework for integrating high-performance BCI decoding with data-driven neuroscientific discovery.