On July 4, 2026, researchers Zhoujie Hou and colleagues introduced Omni-Sleep, an innovative sleep foundation model that leverages hierarchical contrastive learning of the central nervous system (CNS) and autonomic nervous system (ANS) dynamics. This model aims to improve sleep representation learning by utilizing multimodal polysomnography signals including EEG, EOG, EMG, ECG, and respiration.
Understanding the Omni-Sleep Model
Current sleep foundation models often combine diverse biosignals without regard for their physiological organization. Omni-Sleep addresses this limitation by employing a CNS/ANS partition as a physiological prior, facilitating topology-constrained representation learning. The model's learning process is guided by three main objectives:
- Intra-system consistency: This objective captures shared subsystem-level factors within neural and cardio-respiratory signals.
- Inter-system synchronization: This aligns subsystem trajectories to effectively model brain-body dynamics.
- Latent-space masked temporal modeling: This captures long-horizon sleep dynamics.
Evaluation and Performance
Pre-trained on over 100,000 hours of multi-center multimodal PSG data, Omni-Sleep was evaluated for its effectiveness in sleep staging and multi-disease classification. The results demonstrated that Omni-Sleep outperforms several strong foundation-model baselines, offering:
