A new sleep AI model bets that respecting physiological structure is worth more than throwing all biosignals into one undifferentiated pile.
Researchers introduced Omni-Sleep, a foundation model for sleep analysis pre-trained on more than 100,000 hours of polysomnography data from multiple clinical centers. Where prior models blend EEG, EOG, EMG, ECG, and respiratory signals into a single feature space, Omni-Sleep splits them along an anatomical boundary: central nervous system (CNS) signals on one side, autonomic nervous system (ANS) signals on the other. The model trains three objectives in parallel: consistency within each subsystem, synchronization between the two, and long-range temporal patterns across a full night of sleep. That structure also improves robustness when sensors go missing, which happens regularly outside controlled settings.
Sleep staging and disease classification from biosignals are clinically important tasks that still rely heavily on expert manual review. A model that performs well across clinical sites and patient populations without retraining could meaningfully shorten the wait time for sleep disorder diagnosis. Omni-Sleep's design also suggests a broader principle: in medical AI, building around biological reality rather than treating all signals as interchangeable tends to pay off.
The paper reports outperforming strong foundation-model baselines across datasets and modality-ablation settings. All evaluations stay within research-grade recordings, though, which leaves open the harder question: does the physiological hierarchy still help when sensor dropout is just Tuesday in an actual sleep clinic, not a deliberate experimental condition?