| name | omni-sleep-foundation |
| description | Omni-Sleep sleep foundation model methodology using CNS/ANS hierarchical contrastive learning for topology-constrained multimodal PSG representation learning. Use when working with sleep staging, affective BCI, polysomnography analysis, CNS-ANS dynamics, multimodal biosignal foundation models, or physiological hierarchy in representation learning. |
| metadata | {"arxiv_id":"2607.07720","published":"2026-07-04","authors":"Zhoujie Hou, Song Wang, Kexin Lou, Mo Wang, Chen Wei, Quanying Liu","tags":["sleep-foundation-model","CNS-ANS","polysomnography","contrastive-learning","EEG","multimodal"]} |
Omni-Sleep: Sleep Foundation Model via Hierarchical Contrastive Learning
Overview
Omni-Sleep is a sleep foundation model that uses the CNS/ANS (Central/Autonomic Nervous System) partition as a physiological prior for topology-constrained representation learning. Pre-trained on 100,000+ hours of multi-center multimodal PSG data, it addresses the gap where existing sleep models fuse heterogeneous biosignals in a topology-agnostic manner.
Core Architecture
Three Learning Objectives
- Intra-System Consistency: Captures shared subsystem-level factors within neural (EEG, EOG, EMG) and cardio-respiratory (ECG, respiration) signals separately
- Inter-System Synchronization: Aligns CNS and ANS subsystem trajectories to model brain-body dynamics cross-correlations
- Latent-Space Masked Temporal Modeling: Captures long-horizon sleep dynamics through masked token prediction in the latent space
Physiological Prior Structure
CNS Branch: EEG + EOG + EMG → Neural subsystem encoder
ANS Branch: ECG + Respiration → Cardio-respiratory encoder
↓
Cross-system synchronization module
↓
Unified sleep representation → Sleep staging / Disease classification
Key Innovations
- Topology-constrained fusion: Uses known physiological organization (CNS vs ANS) rather than treating all modalities equally
- 100K+ hours pre-training: Multi-center multimodal PSG dataset, significantly larger than prior work
- Modality robustness: Graceful degradation when modalities are missing during inference
- Cross-dataset generalization: Strong transfer across different sleep labs and recording protocols
Application Domains
- Sleep stage classification (AASM standard)
- Multi-disease sleep disorder classification (apnea, insomnia, narcolepsy, etc.)
- Brain-computer interface for sleep-state monitoring
- Multimodal biosignal representation learning
- Physiological hierarchy in AI models
Activation Keywords
omni-sleep, sleep foundation model, CNS-ANS dynamics, polysomnography, PSG, sleep staging, brain-body dynamics, physiological hierarchy, topology-constrained learning, multimodal biosignal, affective computing, sleep physiology
Pitfalls
- Modality availability: Requires at least one CNS and one ANS modality for full inter-system synchronization objective
- Data requirements: 100K+ hours pre-training is a significant compute investment; fine-tuning on smaller datasets is more practical
- Code availability: Repository at https://github.com/AutoBrain-sleep/OmniSleep — check for latest release before implementation
- Generalization limits: While showing cross-dataset generalization, the model may still require domain adaptation for very different recording setups
- Not a replacement for clinical judgment: Foundation model outputs should be validated against clinical ground truth