Summary: Researchers have developed an artificial intelligence foundation model that extracts hidden physiological signals from routine overnight polysomnography to predict long-term health risks, identifying patient groups with markedly different outcomes that standard sleep metrics miss.
Key Facts
- Two-fold mortality risk difference: The AI model stratified patients into five risk groups. Individuals in the highest-risk group experienced roughly double the five-year mortality risk compared with those in the lowest-risk group.
- Outperforms the Apnea–Hypopnea Index (AHI): Unlike the traditional AHI, which summarizes breathing events per hour, the model captures latent physiological features spanning brain, lung, muscle, and cardiac signals that provide stronger prognostic value.
- Balanced performance across sexes: Where AHI has historically correlated more strongly with male presentations, the AI model showed similar predictive accuracy for cardiovascular, neurological, and mortality outcomes in both men and women.
- Unlocks underused data in routine sleep tests: Routine in-lab polysomnograms—an estimated 1 to 4 million performed annually in the U.S.—contain rich, underexploited physiologic information that this model can transform into actionable risk stratification.
Source: Cleveland Clinic
A new artificial intelligence model can mine routine sleep-study recordings to reveal long-term health risks, according to research published in Nature Communications. The interdisciplinary team that developed the model uncovered latent sleep physiology patterns linked to cardiovascular disease, cognitive decline, and mortality—signals not captured by standard summary measures.
The study analyzed sleep recordings from the Cleveland Clinic STARLIT registry and validated findings in an independent nationwide cohort. Using these clinical polysomnograms, the AI model learned complex patterns across multiple physiologic channels and identified five discrete patient subtypes with sharply different long-term trajectories. Patients in the highest-risk subtype had about twice the five-year mortality risk compared with those in the lowest-risk subtype—an important distinction that the apnea–hypopnea index (AHI) did not reveal.
Polysomnography captures continuous signals from the brain, heart, lungs, and muscles. Historically, clinicians have relied on a handful of summary measurements—most notably the AHI—to assess sleep apnea severity. Those summaries discard nuanced, time-varying information that can reflect cumulative cardiovascular, neurologic, and metabolic stress. By contrast, the AI foundation model uses full-spectrum physiologic data to extract prognostic biomarkers and latent features not evident to human reviewers or traditional metrics.
“For decades we have reduced an entire night’s physiology to a small number of summary scores,” said Reena Mehra, M.D., professor of medicine at the University of Washington and the study’s senior clinical author. “AI lets us move beyond those summaries and learn from the full richness of sleep physiology.”
The research was conducted by a multidisciplinary team of sleep clinicians, data scientists, neuroscientists and AI experts assembled through the Discovery Accelerator, a long-term collaboration between Cleveland Clinic and IBM focused on applying advanced computation to life-science discovery. Using more than 10,000 clinical sleep recordings linked to electronic medical records, the model learned representations of sleep physiology that generalized to other datasets, including the independent Sleep Heart Health Study.
Jeffrey Rogers, Ph.D., corresponding author and adjunct professor of neurosurgery at Yale School of Medicine, emphasized that modern AI methods can recover clinically meaningful information embedded within routine tests. “These models reveal patient groups with distinct long-term risks that conventional metrics systematically miss,” he said.
Beyond risk stratification, the approach offers new avenues to study how sleep physiology relates to health outcomes. Extracted features may help researchers and clinicians identify patients at elevated risk for cardiovascular and neurological disease earlier, enabling more targeted prevention and personalized treatment plans.
“Sleep is foundational to health,” said Matheus Lima Diniz Araujo, Ph.D., a sleep researcher at Cleveland Clinic. “Nearly 70 million Americans live with chronic disorders of sleep and wakefulness. This work suggests routine sleep testing can add significant preventive value by revealing individualized risk signals.”
Carl Saab, Ph.D., chief scientist of the Discovery Accelerator, noted the importance of validating these findings in broader, diverse populations and expanding collaboration across clinical and technical communities to translate the model into routine care. Lead author Erhan Bilal, Ph.D., added that because everyone sleeps, sleep studies provide a powerful, universal window into physiologic health—and this foundation-model approach is a first step toward unlocking that potential.
The research team includes Erhan Bilal, Ph.D.; Matheus Lima Diniz Araujo, Ph.D.; Kristen L. Beck, Ph.D.; Catherine M. Heinzinger, D.O.; Samer Ghosn, B.S.; Nancy Foldvary-Schaefer, D.O.; Carl Y. Saab, Ph.D.; Jeffrey L. Rogers, Ph.D.; and Reena Mehra, M.D.
Key Questions Answered:
A: The AHI reduces an entire night of complex, continuous physiologic recordings into a single score of breathing pauses per hour. That simplification omits detailed heart rate variability, brainwave architecture, muscle tone patterns, and subtle oxygen desaturation dynamics that better reflect cardiovascular, neurological, and metabolic strain.
A: Because the model analyzes full-spectrum physiologic signals instead of relying only on counts of upper-airway events, it captures features that are equally predictive across sexes. This helps overcome the historical bias in which the AHI correlated more strongly with male sleep-apnea presentations.
A: Next steps include validation across diverse, international populations; expanded partnerships with technology and industry stakeholders; regulatory and clinical integration work; and embedding the algorithm into sleep-lab software so automated, interpretable risk reports can accompany standard clinical metrics.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full by the staff.
- Additional context was provided by the editorial team.
About this AI and sleep research news
Author: Alicia Reale
Source: Cleveland Clinic
Contact: Alicia Reale – Cleveland Clinic
Image credit: Neuroscience News
Original Research: Open access. “A foundation model for sleep-based risk stratification and clinical outcomes” by Erhan Bilal, Matheus Lima Diniz Araujo, Kristen L. Beck, Catherine M. Heinzinger, Samer Ghosn, Carl Y. Saab, Nancy Foldvary-Schaefer, Jeffrey L. Rogers & Reena Mehra. Nature Communications. DOI: 10.1038/s41467-026-75326-9
Abstract
A foundation model for sleep-based risk stratification and clinical outcomes
Clinical sleep studies record multiple physiologic channels, yet interpretation is often reduced to a few summary measures with limited prognostic value, such as the apnea–hypopnea index. The authors present a foundation model trained on more than 10,000 clinical sleep recordings linked to electronic health records to learn rich representations of sleep physiology.
The study shows that routine sleep physiology contains latent risk structure that conventional metrics overlook. The model identifies five patient risk groups with markedly different trajectories for mortality, cardiovascular disease, and neurological disease. The highest-risk group exhibits more than double the mortality risk of the lowest-risk group, whereas traditional AHI severity categories provide limited predictive discrimination.
The framework generalizes to independent datasets with lower-resolution signals, demonstrating that foundation models can recover clinically meaningful risk information embedded in routine sleep recordings. This approach provides a scalable path toward precision sleep medicine and earlier, more personalized preventive care.