Sleep not only reboots the body but also continuously reflects how the body is functioning under disease, aging, and stress. Clinicians have been measuring simple summary scores, in particular the Apnea-Hypopnea Index (AHI), from polysomnography (PSG) for decades for diagnosing sleep disorders like sleep apnea, but these summary scores have been unable to predict who will truly suffer long-term negative health consequences such as heart disease, cognitive decline, and pre-emptive death.
Described in this research, a new artificial intelligence (AI) foundation model has been developed that examines not only the richness of overnight PSG data to find hidden physiological signals but also categorizes subjects into different long-term risk groups. This finding that much more prognostic information lies dormant in routinely performed PSG exams demonstrates the potential for a scalable method to advance toward precision sleep medicine and early prevention.
Exploring sleep physiology of human life
The underlying premise of this investigation is that sleep physiology, including brain activity (brain waves, breathing, cardiac cycles, muscle signals, and oxygen levels measured overnight), contains a latent “latent risk structure,” which routine measurements of this physiology typically disregard. Classic polysomnography recording parameters (including brain signals from the scalp; eye movements; muscle electrical signals from facial and chin muscles; respiration flow via air entering through and leaving the nose and mouth; movements of the chest and abdomen that move as you breathe; and the amount of oxygen present in the blood), these timeseries are clinically compressed to very simple features and particularly AHI (number of events per hour breathing that stops). However, they often mean much more for things that are critical at the intersection of cerebrocardiovascular risk.
The foundation model helps bridge that gap by modeling high-level representations from an unadulterated (or nearly unadulterated) raw PSG instead of well-established features. If you want to boil it down in the simplest possible terms, essentially you have an AI that is told to “listen to” one full night’s sleep over a battery of recording channels and find any trends over that time that seem to be predictive down the line of outcomes like mortality, cardiovascular disease, or the onset of dementia, but it is entirely free of what human experts might be looking for in a night of sleep. That, you can see, becomes a larger argument for the replacement of narrowly targeted, count-based metrics in favor of comprehensive, data-driven risk stratifications based on sleep.
Experimental design
This study was conducted by a multidisciplinary team consisting of Erhan Bilal, Matheus Lima Diniz Araujo, Kristen L. Beck, Catherine M. Heinzinger, Samer Ghosn, Carl Y. Saab, Nancy Foldvary Schaefer, Jeffrey L. Rogers, and Reena Mehra, representing fields including sleep medicine, biomedical engineering, data science, and neuroscience. The research was undertaken in conjunction with the Cleveland Clinic Discovery Accelerator and the IBM corporation, while clinical sleep data utilized primarily originates from Cleveland Clinic’s STARLIT registry (Sleep Signals, Testing, and Reports Linked to Patient Traits).
“For methodologically speaking, they aggregated over 10 000 in hospital sleep recordings for patients with outcome follow‑ups in their electronic health records, for mortality and cardiovascular and neurological disease, and they fed this information, and the PSGs themselves, into a foundation model (the kind of big, general-purpose AI),” explains Maja. It’s trained on the underlying PSGs across multiple channels, not just any pre-calculated features derived from sleep PSGs, to create a general representation of an entire night of sleep that’s then used to perform this cluster of five subtypes, with each patient being assigned to a cluster with a particular physiological characterization of that sleep pattern.
Then those clusters were evaluated in terms of five‑year mortality and risk of incident cardiovascular or neurological diseases “ and in a second external dataset,” that in this case, of lower‑quality, that did enable high risk and low risk to be predicted.” The major study objectives were to:
- Establish if AI can extract clinically relevant risk information on PSG that is not encoded in standard measures of AHI.
- Discover clinically relevant risk subgroups that have distinct long-term outcomes.
- Determine whether AI extraction of information on PSG offers a sex-specific advantage for prediction.
Read More: Overcoming Aversion to Artificial Intelligence in the Workplace
Sleep physiological signals as predictors of organ systems functioning
The most notable was the revelation of five unique ‘sleep-based risk groups’ with significantly disparate risks for mortality, cardiovascular disease, and neurodegenerative disorders. The highest risk group saw a risk of mortality over 5 years that was more than double compared to the lowest risk group, a disparity that couldn’t be detected through conventional AHI categories. This means even if two patients had comparable AHIs, one may exist in a completely different risk category if their entire sleep pattern is evaluated.
The model had strong prognostic value for cardiovascular and neurologic disease, providing insight into which features beyond the upper airway within the sleep signals are relevant for longer-term system organ health. It’s critical to note that the AI approach performed with similar predictive performance in women and men, differing from AHI, which historically performs better in men for sleep apnea, and may account for sex bias in risk calculation.
Further, our methodology generalized to data collected in the Sleep Heart Health Study, implying that our learned representations were not specific to our specific hospital. Taken together, our research demonstrates that dense, multi-channel sleep data contains predictive biomarkers for survival and for chronic disease that we previously could not obtain by analyzing overall summary metrics.
Read More: How to Improve Sleep Quality with Better Sleep Hygiene
Looking through the author’s lens
Modern AI offers doctors the ability to “regain a great part of the amount of information held in a single night of sleep physiology that is often currently considered” in terms of use, explain the authors, that is, standard tests in a medical test can also detect substantially greater pathophysiological and prognostic information than is currently measured, and the foundation models, as a solution, are an option to extract this vastness.” They conclude the search for latent risk groups by defining not only a triumph for tech science but also for sleep medicine, when such classifications can lead to the earlier recognition and management of some of cardiovascular and neurological disorders.
Clinically, they say that using AHI alone to evaluate cardiovascular risk is likely a “black boxing” technique, “that squeezes into an event counting ‘head count,’ not understanding that pattern and physiology reflect cardiac workload and brain/metabolic loading.” An AI-based group approach could instead flag high long-term risks in seemingly low-risk patients, so appropriate monitoring or clinical and pharmaceutical or behavioral approaches and consultation with the patient could be taken.
The authors add that achieving a predictive accuracy equivalent across sexes also addresses a more perennial problem: equity in diagnostics in sleep medicine. The researchers are calling on future study efforts to include additional verification across varied patient populations and expect to partner with organizations on their next steps to help embed such a clinical risk assessment tool into clinical decision-making and practice, hopefully using readily integrated AI with existing sleep lab software.
Conclusion
This study provides evidence that a base AI model trained on routinely collected clinical sleep study data can stratify patients into five physiologically derived risk groups that exhibit sharply different five-year probabilities of dying, cardiovascular disease, or neurological disease. Patients in the highest-risk group also have more than twice the five-year mortality risk of patients in the lowest-risk group, a pattern masked by standard AHI metrics, illustrating the limitations of summary measures. The model works in an independent population and is roughly as effective in men and women-showing that the use of AI-powered sleep stratification appears robust and that it is perhaps also more equalizing.
References +
- Neuroscience News. (2026c, August 3). AI analyzes sleep data to predict cognitive decline and health. https://neurosciencenews.com/ai-sleep-cognitive-decline-31171/
- Neuroscience News. (2026c, August 3). AI analyzes sleep data to predict cognitive decline and health. https://neurosciencenews.com/ai-sleep-cognitive-decline-31171/
- Neuroscience News. (2026e, August 5). Neuroscience News. https://neurosciencenews.com/neuroscience-topics/machine-learning/
- Cejudo, A., Arrojo, M., Martín, C., & Almeida, A. (2026). AI and Wearables for Early Detection of Cognitive Impairment and Dementia: Systematic Review. Journal of Medical Internet Research, 28, e86262. https://doi.org/10.2196/86262
- AI sleep analysis ties brain age to a 40% dementia risk hike. (2026, July 31). https://www.avantgardenews.com/news/ai-sleep-analysis-linked-to-higher-dementia-risk-20260731 6. Del Valle, I. (2026, May 10). The Displacement of Human Sleep Architecture by AI-Driven Habits. https://www.linkedin.com/pulse/displacement-human-sleep-architecture-ai-driven-habits-del-valle-oy 9te
- Nosta, J. (2025, November 6). Sleep edits experience into memory and builds the enduring self. Psychology Today. https://www.psychologytoday.com/us/blog/the-digital-self/202511/sleep-is-the-line-ai-cannot-cross
- Brain Waves Measured During Sleep Predict Cognitive Impairment Years Before Symptoms Appear | Mass General Brigham. (2025, March 4). https://www.massgeneralbrigham.org/en/about/newsroom/press-releases/brain-waves-during-sleep-pr edict-cognitive-impairment
