By decoding neural, cardiac, respiratory, and other signals hidden within routine sleep studies, the model identified high-risk patients that conventional apnea measurements failed to distinguish.
Paper: A foundation model for sleep-based risk stratification and clinical outcomes. Image Credit: F01 PHOTO / Shutterstock
Researchers have developed an artificial intelligence (AI)-based model that can extract hidden physiological signals from routine sleep test data to stratify patients according to long-term health risks. The study is published in the journal Nature Communications.
Background
Sleep is a regulated biological state essential for physical and mental well-being. Disrupted sleep can substantially affect quality of life and increase the risk of cardiovascular, neurologic, and psychiatric disease.
Sleep disorders are highly prevalent worldwide. They represent a substantial global health burden. Sleep apnea is one of the most common sleep disorders and affects nearly one billion adults worldwide.
Polysomnography is the gold standard test for diagnosing sleep disorders. However, the clinical interpretation of this test is often reduced to single summary measures of limited prognostic value, such as the apnea–hypopnea index (AHI). This index captures only limited information on sleep physiology and thus cannot fully evaluate sleep integrity.
A multidisciplinary research team recently developed an AI-based model that can decode hidden sleep patterns associated with cardiovascular disease, cognitive impairment, and mortality risk.
Model design
The researchers developed an AI foundation model that can analyze full-night polysomnography results to identify hidden physiological patterns associated with long-term clinical outcomes.
Unlike traditional machine learning models, which often depend on task-specific designs and require complete retraining for new tasks, foundation models are broad, general-purpose models trained on large, diverse datasets. This methodology allows these models to adapt to new datasets or tasks with minimal additional training.
However, applying this methodology to polysomnography data is challenging, as sleep recordings are large, noisy, and highly individual-specific. These challenges highlight the need for developing innovative methods that can transform physiological signals recorded at different frequencies into uniform tokens for machine learning models.
Here, researchers trained their foundation model using a unique resource of 10,000 high-resolution polysomnography studies from the Cleveland Clinic Sleep Signals, Testing, and Reports Linked to Patient Traits (STARLIT) Registry, which were linked to electronic medical records. After quality control, 9,608 studies from 9,297 patients were retained for clustering analyses.
Key findings and significance
By extracting and analyzing hidden physiological characteristics across neural, ocular, muscular, cardiac, oxygenation, and ventilatory signals from routine sleep recordings, the foundation model identified five embedding-derived patient groups with markedly different trajectories for mortality, major adverse cardiovascular events, atrial fibrillation, cognitive impairment, and epilepsy.
Patients categorized into the highest-risk group showed more than double the mortality risk compared to those in the lowest-risk category. In contrast, the traditional AHI severity categories failed to predict mortality.
For decades, the AHI has been widely used to define sleep-disordered breathing severity. Yet, the index failed to detect any association with mortality in the study cohort, indicating insufficient ability to stratify clinically meaningful risks.
The AI-derived risk groups, on the other hand, showed strong, graded associations with clinical outcomes, even after adjusting for demographics, comorbidities, and the AHI itself. These findings suggest that the foundation model can capture more detailed pathophysiological signals of sleep associated with clinical outcomes beyond those captured by airway obstruction alone.
The American Thoracic Society (ATS) has prioritized developing scalable, objective measures that go beyond the AHI to predict cardiovascular and neurological outcomes. The current findings directly contribute to these priorities by providing a scalable, externally validated framework that uncovers hidden physiologic signals with clear clinical implications.
To further validate the model’s robustness, researchers applied a simplified two-group version of the stratification approach to data obtained from the Sleep Heart Health Study (SHHS). The approach reproduced associations with mortality and heart failure in both men and women.
In contrast, earlier SHHS analyses linked severe sleep-disordered breathing to heart failure only in men and found mortality associations primarily among middle-aged men with severe disease. The consistency of results across independent cohorts using different polysomnography protocols supports the approach’s potential generalizability.
Notably, SHHS participants in the high-risk group exhibited markedly short total sleep time, similar to the marked physiological disruptions observed in the highest-risk cluster in the STARLIT cohort, further supporting the biological plausibility and clinical relevance.
The major strength of the study is the integration of multimodal polysomnographic signals with longitudinal electronic medical records from a large, demographically diverse clinical cohort. The linked records provided an average total clinical observation window of 14.5 years, including medical history before and follow-up after polysomnography. This integration improved the richness of the clinical data and allowed more comprehensive risk stratification than either source alone.
The foundation model, however, was trained using technician-supervised objectives, including sleep stages, respiratory events, and oxygen desaturations. A possible influence of these objectives on the learned embedding cannot be fully ignored. Future research exploring self-supervised and disease-targeted objectives is therefore needed to reduce dependence on technician-defined labels.
Other limitations include unavailable objective data on positive airway pressure adherence, incomplete medication records, possible residual confounding, and reliance on electronic health record diagnostic codes. The retrospective design also prevents causal conclusions, and further external validation and prospective clinical trials will be required before the model can be implemented in healthcare settings. Such studies will also need to determine whether the approach can accurately guide risk assessment and improve outcomes for individual patients.
