A new AI model from Cleveland Clinic analyzes polysomnography data to stratify long-term disease risk, reframing sleep disruption as a critical health metric.
On September 8, 2026, Cleveland Clinic researchers published findings on a new artificial intelligence model. This transformer-based foundation model analyzes polysomnography data to stratify long-term clinical risk. The work was conducted through a 10-year research partnership between Cleveland Clinic and IBM. This joint initiative involves sleep physicians, artificial intelligence researchers, data scientists, and neuroscientists.
The project evaluates high-resolution sleep physiology to identify patterns associated with future disease and mortality. This research shifts clinical assessment away from single metrics like the apnea-hypopnea index. Instead, it demonstrates how complex physiological signals can group patients into distinct risk categories. These groupings help predict outcomes for cardiovascular, neurologic, and psychiatric events.
The Cleveland Clinic team trained the foundation model using approximately 10,000 polysomnography studies. These records were drawn from the STARLIT registry and linked to electronic medical records covering more than a decade. Each study captured roughly eight hours of continuous data on brain activity, heart rate, and breathing. The recordings also included detailed measurements of body movement and blood oxygen levels.
Rather than relying only on conventional diagnostic thresholds, the model converted sleep physiology into numerical representations. The algorithm clustered these high-dimensional embeddings into five distinct groups designated RG1 through RG5. The model reviewed the complete dataset for complex patterns that are difficult for human reviewers to spot. Cleveland Clinic noted that technologists and physicians typically interpret polysomnograms in 30-second segments.
The foundation model searched across the entire recording to establish physiological signatures. Matheus Lima Diniz Araujo, PhD, stated that the apnea-hypopnea index alone had limited prognostic utility. The model identified clinically meaningful patterns even among patients with the same average index score. This comprehensive approach aligns with a broader shift toward multimodal analysis of complete physiological recordings.
The model analyzed multiple channels of physiological data simultaneously to build its classifications. These recordings included detailed measurements of body movement, respiratory effort, and dynamic changes in blood oxygen levels. By processing all available channels collectively, the artificial intelligence system avoided the limitations of isolating a single biological variable. This comprehensive review process allowed the researchers to identify subtle physiological deteriorations that precede major clinical events.
Diagnostics World News reported that the STARLIT registry contains close to 300,000 total sleep studies. These records consist primarily of polysomnograms linked with electronic medical record information. This extensive database provides a robust foundation for identifying subtle connections between nighttime physiology and long-term health outcomes. The findings illustrate the growing role of foundation models in clinical data analysis.
For founders and operators with demanding schedules, these findings reframe how we should approach rest. Quality sleep is not merely a behavioral tool for supporting daily executive performance. It is a complex biological state that reflects long-term cardiovascular and neurological resilience. When sleep architecture is chronically fragmented, the physical toll extends far beyond subjective morning fatigue.
Persistent sleep disruption carries cumulative risk across multiple biological systems. Ignoring these physiological changes actively compromises the systems required for high-level decision making. Addressing severe sleep issues early is a critical component of any approach to healthy aging and executive longevity. The association with specific psychiatric and neurologic incident outcomes underscores the systemic nature of sleep deprivation.
Many professionals monitor their rest using wearable technology, but these devices cannot replicate clinical polysomnography. Consumer data is highly useful for tracking behavioral consistency and resting heart rate trends. However, commercial wearables are not validated to predict the specific disease outcomes analyzed by the Cleveland Clinic team. Executives experiencing persistent daytime sleepiness, witnessed breathing pauses, or unexplained cognitive changes should prioritize formal clinical evaluation.
Organizations can still monitor subjective markers of physical readiness while utilizing clinical resources for severe cases. A baseline monitoring protocol should track sleep regularity, subjective recovery scores, and daytime alertness. When an operator notices an unexplained drop in daily workload tolerance, this behavioral data can inform the ensuing clinical discussion. The goal is to build a reliable recovery structure that uses basic tracking as a baseline and professional evaluation as the escalation path.
Organizations that support high-level operators could use these concepts to build tiered risk protocols. A structured program might track basic regularity and subjective recovery as a primary baseline. When individuals display persistent disruption or specific clinical symptoms, the next step should be targeted referral. Sustained focus and cognition require treating severe sleep issues as significant medical risks rather than mere lifestyle problems.
The foundation model successfully grouped patients into specific categories based on the severity of their physiological abnormalities. The researchers characterized RG1 and RG2 as groups representing patients with minimal polysomnography abnormalities. In contrast, RG5 represented patients with multiple comorbidities and data patterns consistent with severe sleep disruption. The study reported a progressive, monotonic increase in incident cardiovascular, neurologic, and psychiatric outcomes from RG1 through RG5.
The highest clinical risks were consistently observed in the fifth group. According to summaries provided by Cleveland Clinic and the University of Washington, patients categorized in RG5 had more than twice the five-year mortality risk of those in RG1. The researchers also noted that the model was independently validated in a nationwide patient cohort. These statistics highlight the strong association between comprehensive sleep physiology and significant long-term health events.
While the study demonstrates significant correlations, the model identifies associations rather than absolute future certainties. The five risk groups represent statistical categories based on observed incidence rates, not definitive clinical diagnoses. The model does not predict exactly when an individual patient will experience a specific disease event. Furthermore, the available summaries do not establish that changing a person's sleep profile will directly reduce the predicted risks.
The publicly available summary from Cleveland Clinic lacks certain statistical details required for a complete scientific review. It does not provide outcome-specific hazard ratios, exact event counts, or comprehensive adjustment strategies. The reliance on clinical polysomnography data linked with electronic medical records also introduces the potential for referral bias. Patients undergoing formal sleep studies typically have pre-existing symptoms, which means the findings may not apply evenly to healthy adults.
The reliance on historical medical records introduces the possibility of unmeasured confounding variables. Factors such as diet, exercise habits, and baseline stress levels were not explicitly detailed as adjustment criteria in the primary public summaries. These lifestyle inputs heavily influence both sleep architecture and long-term cardiovascular risk. Without detailed adjustment strategies for these variables, the exact causal weight of sleep disruption remains difficult to isolate.
The output of this model may be difficult for clinicians to interpret mechanistically. High-dimensional embeddings identify predictive patterns mathematically, meaning they do not necessarily explain which specific physiological features caused the risk classification. The Cleveland Clinic article confirmed independent validation in a nationwide cohort, but it omitted the validation group's sample size. Finally, these results cannot be translated into a claim that executives can determine their disease risk from consumer-grade smartwatch data.
These findings present a potential route toward what Cleveland Clinic refers to as precision sleep medicine. Araujo noted that the project could expand the value of routine sleep testing by reinforcing its connection to chronic disease. Rather than stopping at a basic apnea diagnosis, a comprehensive sleep study could trigger a broader clinical evaluation. Araujo suggested that higher risk classifications could help guide referrals to specialists such as cardiologists or neurologists.
Before artificial intelligence models can be deployed routinely, the industry must address several technical challenges. Other recent research has emphasized the need for harmonized data labels and multicenter external validation. Calibration across different age brackets and comorbidity groups is strictly necessary to ensure accuracy across diverse patient populations. Explainable systems will be required so that attending physicians understand the exact reasoning behind a specific risk classification.
Future models will likely function as decision-support tools designed to complement expert medical judgment. As healthcare institutions continue to combine clinical datasets with machine learning expertise, sleep and recovery assessments will become increasingly sophisticated. For high-performing professionals, this evolution will eventually provide a more accurate picture of how rest impacts long-term physical resilience. Until these tools achieve clinical standardization, a formal medical evaluation remains the gold standard for addressing severe sleep disruption.
Stay connected for research and practical guidance on executive performance, energy, focus, sleep, recovery and longevity. Ideas built for people who want to stay sharp, capable and effective for the long run.
Build habits and systems that support clear thinking, steady energy and long term capacity throughout a demanding career.
explore the Blog
