SLEEP risk stratification using a foundation model identified five groups with distinct mortality and disease trajectories
Researchers developed a transformer-based foundation model using more than 10,000 clinical polysomnography recordings linked to longitudinal electronic medical records. The model analyzed multimodal sleep physiology, including neurological, cardiac, respiratory, muscular, and oxygen saturation signals, rather than relying on individual summary measures
The resulting physiologic representations separated patients into five stable risk groups. These groups showed progressively different rates of mortality and incident cardiovascular, neurological, psychiatric, and metabolic conditions
Patients in the highest-risk group had more than twice the adjusted mortality risk of those in the lowest-risk group. After adjustment for demographics, body mass index, comorbidities, and the apnea–hypopnea index, the highest-risk group had a mortality hazard ratio of 2.38
Foundation Model Outperforms Conventional Sleep Measures
Risk increased across several clinical outcomes. The highest-risk group showed elevated hazards for major adverse cardiovascular events, heart failure, myocardial infarction, atrial fibrillation, cognitive impairment, and epilepsy. The strongest associations included epilepsy, with a hazard ratio of 2.40, and atrial fibrillation, with a hazard ratio of 2.23
Sensitivity analyses indicated that the model was not driven solely by respiratory measurements. Removing electroencephalogram or electrocardiogram inputs substantially disrupted risk-group assignments, demonstrating that neural and cardiac signals contributed meaningfully to sleep risk stratification
External Validation Supports Broader Application
The framework was validated in an independent population-based cohort containing lower-resolution sleep recordings. Despite differences in available signals and sampling rates, the model continued to identify groups with distinct risks of mortality and heart failure
The findings suggest artificial intelligence could recover clinically important information embedded within routine sleep studies that conventional metrics systematically overlook. This approach could eventually help identify patients who may benefit from cardiovascular monitoring or early neurological assessment
However, the retrospective design prevents causal conclusions. Treatment adherence and medication use were also incompletely captured, while prospective trials and validation in additional populations will be required before clinical implementation
Reference
Bilal E et al. A foundation model for sleep-based risk stratification and clinical outcomes. Nat Commun. 2026;17:7603
Featured Image: blackday on Adobe Stock
