Sleep AI Finds 5 Risk Groups, More Than Doubling Mortality Risk Over AHI
Updated
Updated · Nature.com · Aug 3
Sleep AI Finds 5 Risk Groups, More Than Doubling Mortality Risk Over AHI
3 articles · Updated · Nature.com · Aug 3
Summary
More than 10,000 clinical sleep recordings linked to electronic medical records let researchers train a foundation model that split patients into five risk groups with sharply different mortality, cardiovascular and neurological outcomes.
The highest-risk group, RG5, showed 2.38 times the mortality risk of the lowest group after adjustment, while conventional apnea-hypopnea index severity categories showed little or no predictive value.
The model learned from full-night polysomnography signals—not just breathing events—using multimodal data such as EEG, ECG, oxygen saturation and airflow, with sensitivity tests showing non-respiratory channels materially shaped risk assignment.
An external test in the Sleep Heart Health Study also separated higher- and lower-risk patients despite lower-resolution data, supporting the approach as a scalable path toward precision sleep medicine.