SleepMamba: Dual-axis selective state-space model for multimodal sleep staging

This is the authors' abstract. We don't add key points for this paper.

Effective sleep staging from raw polysomnography remains challenging due to heterogeneous modalities and complex temporal dynamics. Existing methods often rely on handcrafted time-frequency transforms or adopt complex architectures, lacking structured mechanisms to model multi-resolution features and efficiently integrate heterogeneous modalities. We propose SleepMamba, a lightweight and scalable architecture that unifies multiscale representation learning with dual-axis state-space modeling to capture both temporal and modal dependencies. The architecture includes: (i) a multiscale convolutional encoder that extracts resolution-aware features directly from raw PSG inputs, (ii) a dual-axis state-space module that concurrently models intra- and inter-modal interactions within each epoch, and (iii) a bidirectional temporal Mamba that captures inter-epoch dependencies to ensure sequence-level coherence. Experiments on SleepEDF-78 and SHHS1 demonstrate that SleepMamba achieves competitive performance with improved computational efficiency, offering a unified and structured modeling framework for multimodal sleep analysis.