Omni-Sleep foundation model improves sleep staging with physiological hierarchy

Sleep analysis has long been constrained by models that treat multimodal biosignals—EEG, EOG, EMG, ECG, respiration—as flat inputs, ignoring their physiological organization. Omni-Sleep introduces a topology-constrained approach by explicitly modeling the central nervous system (CNS) and autonomic nervous system (ANS) as separate subsystems. Through three objectives—intra-system consistency (capturing shared factors within CNS or ANS signals), inter-system synchronization (aligning trajectories between brain and body), and latent masked temporal modeling (learning long-range sleep dynamics)—the model builds a rich, structured representation of sleep physiology. It was pre-trained on over 100,000 hours of multi-center polysomnography data, making it one of the largest sleep AI models to date.

Evaluated across multiple datasets and modality-ablation settings, Omni-Sleep consistently outperforms strong foundation-model baselines on sleep staging (classifying N1, N2, N3, REM) and multi-disease classification (e.g., sleep apnea, insomnia). Key advantages include markedly improved label efficiency—achieving high accuracy with far fewer annotated examples—and robust generalization to unseen datasets. The model also tolerates missing channels (e.g., missing ECG), a critical feature for real-world clinical settings where sensor dropouts are common. By demonstrating that explicit physiological hierarchy boosts representation learning, Omni-Sleep points toward a new paradigm for sleep diagnostics and personalized health monitoring. Code is publicly available.