Omni-Sleep: A Foundation Model for Sleep Analysis Using CNS-ANS Dynamics

A new sleep foundation model, Omni-Sleep, has been developed to improve the analysis of sleep physiology. Unlike previous models that treat biosignals uniformly, Omni-Sleep incorporates the distinct yet coordinated dynamics of the central nervous system (CNS) and autonomic nervous system (ANS) as a physiological prior. This approach allows for more structured representation learning. The model was trained on an extensive dataset of over 100,000 hours of multimodal polysomnography data, including EEG, EOG, EMG, ECG, and respiration. Omni-Sleep employs three key objectives: ensuring consistency within CNS and ANS subsystems, synchronizing trajectories between these systems to model brain-body dynamics, and using masked temporal modeling for long-horizon sleep patterns. Evaluations show Omni-Sleep surpasses current foundation model baselines in tasks like sleep staging and multi-disease classification. It demonstrates improved label efficiency, better generalization across different datasets, and enhanced robustness to missing data modalities, underscoring the benefits of integrating physiological hierarchy into sleep representation learning.