Omni-Sleep: A Sleep Foundation Model via Hierarchical Contrastive Learning of CNS--ANS Dynamic
Omni-Sleep introduces a sleep foundation model that leverages the Central Nervous System (CNS) and Autonomic Nervous System (ANS) partition as a physiological prior for topology-constrained representation learning. The model utilizes three distinct training objectives: intra-system consistency, inter-system synchronization, and latent-space masked temporal modeling to capture complex brain-body dynamics. Pre-trained on over 100,000 hours of multi-center multimodal Polysomnography (PSG) data, dem
Analysis
TL;DR
- Omni-Sleep introduces a sleep foundation model that leverages the Central Nervous System (CNS) and Autonomic Nervous System (ANS) partition as a physiological prior for topology-constrained representation learning.
- The model utilizes three distinct training objectives: intra-system consistency, inter-system synchronization, and latent-space masked temporal modeling to capture complex brain-body dynamics.
- Pre-trained on over 100,000 hours of multi-center multimodal Polysomnography (PSG) data, demonstrating superior performance in sleep staging and multi-disease classification tasks.
- Omni-Sleep exhibits improved label efficiency, robust cross-dataset generalization, and resilience to missing modalities compared to existing topology-agnostic foundation models.
Why It Matters
This research addresses a critical gap in current sleep AI models by integrating physiological structure into representation learning, moving beyond simple signal fusion. For AI practitioners and medical researchers, it demonstrates how embedding domain-specific biological hierarchies can significantly enhance model robustness and generalizability in healthcare applications.
Technical Details
- Physiological Prior: The architecture explicitly partitions biosignals (EEG, EOG, EMG, ECG, respiration) into CNS and ANS groups, enforcing topology constraints rather than treating all inputs as homogeneous.
- Three-Objective Learning:
- Intra-system consistency: Captures shared factors within neural or cardio-respiratory subsystems.
- Inter-system synchronization: Aligns trajectories between CNS and ANS to model brain-body interactions.
- Latent-space masked temporal modeling: Learns long-horizon dependencies in sleep dynamics.
- Dataset Scale: Trained on a massive corpus of over 100,000 hours of multimodal PSG data collected from multiple centers.
- Evaluation Metrics: Benchmarked against strong foundation-model baselines on sleep staging and multi-disease classification, with specific tests on modality ablation to prove robustness.
Industry Insight
- Integrating biological priors into foundation models is a viable strategy for improving reliability in clinical AI, suggesting that future medical models should prioritize anatomical or physiological structure over generic fusion techniques.
- The demonstrated robustness to missing modalities offers practical value for real-world deployment, where incomplete sensor data is common, reducing the need for strict hardware standardization across different clinical settings.
- The availability of code and the large-scale pre-training dataset may accelerate research in multimodal health monitoring, encouraging the development of similar hierarchical approaches in other physiological domains.
Disclaimer: The above content is generated by AI and is for reference only.