Andrii Bidochko

• Updated: July 13, 2026
• 6 min read

Omni-Sleep: A Sleep Foundation Model via Hierarchical Contrastive Learning of CNS–ANS Dynamic

Direct Answer

Omni‑Sleep is a new sleep‑foundation model that learns a unified representation of central‑nervous‑system (CNS) and autonomic‑nervous‑system (ANS) signals through hierarchical contrastive learning. By embedding the physiological hierarchy of brain‑body dynamics into its architecture, the model delivers more accurate sleep staging and disease classification while remaining robust to missing modalities.

Background: Why This Problem Is Hard

Polysomnography (PSG) captures a rich tapestry of biosignals—EEG, EOG, EMG, ECG, and respiratory traces—that together describe the sleep state. Translating this multimodal data into reliable clinical insights faces three intertwined challenges:

• Physiological heterogeneity: CNS signals (EEG, EOG) reflect cortical activity, whereas ANS signals (ECG, respiration) encode cardio‑respiratory regulation. Their sampling rates, noise profiles, and temporal dynamics differ dramatically.
• Signal interdependence: Sleep stages emerge from coordinated brain‑body interactions. Ignoring this synchronization leads to fragmented representations that miss critical biomarkers.
• Data scarcity and modality loss: Real‑world PSG recordings often suffer from sensor dropout, motion artifacts, or incomplete channel sets, making models that assume a fixed input topology brittle.

Most existing sleep foundation models treat all channels as a flat bag of inputs, applying generic fusion layers that lack any physiological prior. Consequently, they struggle with cross‑dataset generalization, require large labeled corpora, and degrade sharply when a modality is missing.

What the Researchers Propose

Omni‑Sleep introduces a three‑pronged learning framework that respects the CNS/ANS hierarchy:

1. Intra‑system consistency: Within each subsystem (CNS or ANS), the model aligns embeddings from different modalities (e.g., EEG vs. EOG) to capture shared latent factors such as arousal level or muscle tone.
2. Inter‑system synchronization: A contrastive objective forces the CNS and ANS embeddings to evolve in lockstep, mirroring the physiological coupling that defines sleep architecture.
3. Latent‑space masked temporal modeling: By masking random time windows in the joint latent space and predicting them, the model learns long‑range dependencies that span entire sleep cycles.

These objectives are combined in a single encoder‑decoder backbone that processes each modality through a modality‑specific front‑end, then merges them under the CNS/ANS partition before applying the contrastive losses.

How It Works in Practice

The operational pipeline can be broken down into four stages:

1. Modality‑Specific Front‑Ends

Each raw signal passes through a lightweight convolutional encoder tuned to its sampling frequency and noise characteristics. The encoders output modality‑level embeddings while preserving temporal resolution.

2. System‑Level Aggregation

Embeddings from CNS channels (EEG, EOG, EMG) are pooled into a CNS vector; ANS channels (ECG, respiration) form an ANS vector. This explicit partition enforces the physiological prior.

3. Hierarchical Contrastive Learning

• Intra‑system contrast: Positive pairs are drawn from different CNS modalities recorded simultaneously; negatives are shuffled across time or subjects.
• Inter‑system contrast: The CNS and ANS vectors from the same epoch are treated as a positive pair, while vectors from mismatched epochs serve as negatives.

These contrastive steps are computed with a temperature‑scaled dot‑product loss, encouraging the model to discover shared latent factors without explicit labels.

4. Masked Temporal Decoder

After the contrastive stage, a transformer‑style decoder receives a partially masked latent sequence and learns to reconstruct the missing windows. This step injects long‑range temporal awareness, enabling the model to capture the progression of sleep cycles.

Because the architecture treats the CNS/ANS split as a hard constraint, it can gracefully handle missing channels: if a modality is absent, its front‑end is simply omitted, and the contrastive objectives adapt to the available signals.

Evaluation & Results

Omni‑Sleep was pre‑trained on more than 100,000 hours of multi‑center PSG recordings, spanning diverse demographics and equipment vendors. The authors evaluated the model on two downstream tasks:

Sleep Staging

• Benchmark datasets included Sleep‑EDF, MASS, and a proprietary clinical cohort.
• Omni‑Sleep achieved a macro‑averaged F1 score 4–6 % higher than the best publicly available foundation models, even when up to two modalities were removed.
• Label‑efficiency experiments showed that fine‑tuning with just 10 % of the usual annotated data retained 92 % of full‑data performance.

Multi‑Disease Classification

• Tasks covered apnea detection, REM behavior disorder, and narcolepsy classification.
• The model’s latent space clustered disease phenotypes more distinctly than baseline models, leading to a 7 % boost in AUROC on a held‑out test set.

Across all experiments, Omni‑Sleep demonstrated superior cross‑dataset generalization, confirming that the physiological hierarchy acts as a regularizer that mitigates overfitting to site‑specific quirks.

Why This Matters for AI Systems and Agents

For AI practitioners building health‑focused agents, Omni‑Sleep offers three concrete advantages:

• Robust multimodal fusion: The CNS/ANS partition provides a reusable blueprint for any system that must combine heterogeneous biosignals, from wearable health monitors to ICU dashboards.
• Label‑efficient fine‑tuning: Agents can adapt the foundation model to niche clinical endpoints with minimal annotation effort, accelerating time‑to‑value.
• Resilience to sensor loss: In real‑world deployments, devices frequently drop connections. Omni‑Sleep’s topology‑aware design ensures graceful degradation rather than catastrophic failure.

These properties align directly with the needs of UBOS platform overview, where modular AI pipelines must remain operational despite fluctuating data streams. Moreover, the model’s latent representations can be indexed in a vector database such as Chroma DB integration, enabling rapid similarity search for patient cohort analysis.

Developers can also embed Omni‑Sleep into conversational health assistants. By coupling the model with OpenAI ChatGPT integration, agents can translate raw PSG embeddings into natural‑language explanations for clinicians or patients, bridging the gap between signal processing and actionable insight.

For organizations seeking to automate sleep‑study workflows, the Workflow automation studio can orchestrate data ingestion, model inference, and report generation in a single no‑code pipeline, dramatically reducing operational overhead.

Read more about how UBOS is integrating cutting‑edge models like Omni‑Sleep in our UBOS blog post on Omni‑Sleep.

What Comes Next

While Omni‑Sleep marks a significant step forward, several avenues remain open for exploration:

• Extension to home‑based wearables: Adapting the model to lower‑fidelity signals (e.g., PPG, actigraphy) could democratize sleep monitoring beyond the sleep lab.
• Real‑time inference: Optimizing the encoder for edge devices would enable on‑device staging, reducing latency for closed‑loop interventions.
• Multi‑task pre‑training: Incorporating related tasks such as circadian rhythm prediction or stress detection could enrich the latent space further.
• Explainability layers: Adding attention visualizations that map latent factors back to physiological events would increase clinician trust.

Addressing these challenges will require collaboration across signal processing, clinical neuroscience, and systems engineering. UBOS is already exploring partnerships through its UBOS partner program, inviting startups and research labs to co‑develop next‑generation health agents.

For enterprises looking to embed such capabilities at scale, the Enterprise AI platform by UBOS offers managed infrastructure, compliance tooling, and dedicated support to accelerate deployment.

Finally, the open‑source community can contribute to the project via the official repository linked in the Omni‑Sleep paper. Collaborative benchmarking and dataset expansion will be key to validating the model across diverse populations.

Ready to experiment with hierarchical multimodal models in your own AI pipelines? Explore the full suite of UBOS tools and start building tomorrow’s health agents today.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.