September 1, 2026 | For decades now, doctors have been referring people suspected of having obstructive sleep apnea to a specialized medical facility for a polysomnogram (PSG). The gold-standard overnight test collects millions of data points from across the body, only a small subset of which gets prioritized for scoring the severity of the condition—notably, the apnea-hypopnea index (AHI) ranking how often their breathing slows or stops per hour of sleep.
This remains the disappointing norm at 45,000 sleep centers across the U.S., according to neurologist Nancy Foldvary-Schaefer, D.O., M.S., at the Cleveland Clinic. Epidemiological studies conducted over the last 10 years fail to adequately link the AHI to long-term health outcomes like cardiovascular disease and mortality, she says, both of which are strongly linked to severe sleep apnea. “Last year the American Thoracic Society called for studies searching for other PSG-recorded variables linked to poor health outcomes.”
As is abundantly clear to Foldvary-Schaefer, a comprehensive analysis of all available PSG data with machine learning to detect patterns invisible to the human eye could reveal “hidden gems” in sleep studies and help crack the mystery. An electromyography (EMG) sensor tracking electrical muscle activity is a required and fundamental part of a sleep study and can help recognize a sleep disorder that strongly predicts Parkinson’s disease, yet “this signal is not routinely analyzed in sleep laboratories,” she offers as an example.
The point was well made in a study, recently published in Nature Communications (DOI: 10.1038/s41467-026-75326-9), where artificial intelligence (AI) was used to identify previously undetected, long-term health risks in patients who have undergone PSG. The gems, in this case, included heart disease, cognitive decline, and, most remarkably, death—and the foundation model identified distinct risk groups that outperformed conventional AHI-based severity categories.
A key collaborator on the study was Matheus Lima Diniz Araujo, Ph.D., a computer scientist and sleep researcher at the Cleveland Clinic. A polysomnogram is a unique and powerful data source for AI models because it captures continuous, multi-channel physiological data in real time while a patient sleeps. The test provides a complete picture of the body’s systems in that it measures brain waves, eye movements, heart rate, breathing, and muscle activity with a variety of strategically placed sensors.
“Sleep, like exercise and diet, is one of the third foundational pillars of health and wellness,” Foldvary-Schaefer emphasizes. “There is a reason why we sleep and it’s not just because there’s nothing else to do at the end of the day. Sleep is restorative to every cell in the brain and the body, so it makes sense that the data being collected by these sensors can help identify early warning signs of chronic diseases.”
A polysomnogram is used to diagnose an assortment of conditions, including those causing extreme daytime sleepiness, abnormal nighttime behaviors, chronic trouble sleeping, and involuntary movements during rest. But by and large people go to sleep labs for an assessment of sleep apnea, says Foldvary-Schaefer, and “that has led to an extreme focus on just a few variables out of thousands of data points that an overnight sleep study collects.”
The combined 55-bed overnight sleep laboratory capacity across the Cleveland Clinic’s network makes it one of the largest of its kind in the U.S., and a good starting point for investigating “how these other signals might be even better predictors of long-term outcomes than the variables the field has been so focused on,” she says. Data for the study was conveniently provided by the Cleveland Clinic’s STARLIT (sleep signals, testing, and reports linked to patient traits) registry.
The registry consolidates data from close to 300,000 sleep studies of all kinds, primarily PSGs, together with electronic medical record data on those patients, Araujo says. The PSG data ranges from the unprocessed electrical signals from electroencephalogram (EEG) sensors capturing 250 data points per second to the overnight summary of all the captured metrics. “It’s basically an organized, curated database that we have been working very hard to maintain ... [to] support the research we’re doing.”
For the latest study, the pattern detection work was “100% data-driven” by an unsupervised foundation model, he reports. But it was enabled by “computing knowledge and power from IBM,” the Cleveland Clinic’s partner in a 10-year Discovery Accelerator joint research initiative together with neuroscientists and clinicians in the Cleveland Clinic’s Research Institute and Sleep Disorders Center. The computational exercise identified five patient risk groups with markedly different trajectories for mortality, cardiovascular, and neurological disease.
As time passed, patients in risk group 1 (3,357 people) and 2 (1,877 people) were much healthier than individuals in the other groups, including group 3 (2,867 people) and 4 (1,144 people), Araujo points out, and risk predictions reliably increased in tandem with undesirable outcomes. The highest risk category (group 5, with 363 people) that had the least amount of sleep was found to have more than double the mortality risk of the lowest, whereas AHI severity categories showed limited predictive value.
Sleep studies require many hours of manual review to score and interpret the hundreds of pages of data generated, says Foldvary-Schaefer. The process requires the application of multiple sensors and overnight monitoring of vital signs, breathing interruptions, and irregular body movements. “Technologists will assign a stage [e.g., light sleep, REM] and identify events [e.g., apnea, leg movement, arousal] on every 30-second epoch across the entire night and then the physician reviews each of those, so it is very labor-intensive and completely subject to the way our field [i.e., American Academy of Sleep Medicine] defines these variables.”
While her eye will see some other subtleties, Foldvary-Schaefer adds, it is “not going to be able to pull together all these variables and create risk categories like the machine learning model could.” The goal now is to replicate and validate the work, as well as examine specific neurological disease outcomes. Longer term, the hope is to be able to assign every sleep study to a level of risk that could impact longevity of millions of patients in the U.S. who undergo polysomnography every year.
Given that PGS data is exceptionally rich for AI and machine learning models, Araujo says he imagines the utility of the test expanding to conditions beyond sleep disorders. Cost implications aside, there could be potential clinical value in everyone getting a sleep test at least once in their lifetime.
Based on years of clinical work, it is estimated that about 25% of the population is at risk of sleep-disordered breathing, says Foldvary-Schaefer, including children and teenagers. Sleep apnea is closely linked to metabolism, and this connection is a major factor in the current epidemic of teen obesity. It is also known that roughly 15-30% of the U.S. adult population is at any time struggling with chronic insomnia, she continues. People who sleep the least are among those falling in the highest risk category (group 5), as calculated by AI.
Over each of the last few years, sleep centers in the country have collectively performed several million polysomnograms, Foldvary-Schaefer notes. Insurance companies aren’t fond of the tests because they’re expensive, “but, wow, what a deal that would be if we could get so much more risk assessment out of a sleep study ... [and] give patients and payers more for their money.”
At minimum, perhaps the executive physicals people are getting in middle age could include a PSG, she muses. If the work done to date can be replicated and refined, then perhaps an argument can be made to add a sleep study to the arsenal of wellness tests done at these annual exams.
The AI model, unlike the AHI, did a notably good job of predicting outcomes for both sexes and not just men, says Foldvary-Schaefer. “In clinical sleep medicine there is a significant gender disparity in certain disorders that we treat, the most important one being sleep apnea.”
The classic description of a person with sleep apnea is “a heavyset man with a big neck who snores like a freight train, and that’s what drives people into the sleep lab,” she says. But research over the last decade or two has revealed that “post-menopausal women are at no less risk of obstructive sleep apnea than men, but they look completely different—they’re thinner, they have smaller necks, they don’t snore as loudly, and they present with insomnia, unrefreshing sleep, and fatigue,” and often do not have bed partners to provide ancillary sleep information.
It is also now recognized that women can be symptomatic but have a lower, “less impressive” AHI, says Foldvary-Schaefer. She points to the influence of progesterone and estrogen receptors in the upper airway musculature that help maintain airway patency pre-menopause. When those sex hormones decline post-menopause, it is thought to drive the collapsibility of the upper airway that puts them at risk for sleep apnea.
But the sleep apnea that women experience is intensely focused in REM sleep when there is already a loss of muscle tone, she explains, so the muscles controlling the airway can get either very floppy or temporarily paralyzed. The AI model didn’t show a difference, gender-wise, when categorizing people based on risk and that’s likely because “the AHI isn’t driving stratification into these different groups, but the EEG data ... and the cardiology data is.”
Foldvary-Schaefer says she has a colleague in the sleep research field, who is now using AI to predict sleep apnea from EKG data, which is typically part of executive physicals and wellness checks for middle-aged people. Actigraphy (movement) and patient behavior data that can be captured for multiple days via wearables will likely also be a target for AI risk prediction purposes, adds Araujo. But he remains a fan of PSG because it captures data from so many different types of biosensors, records sequentially, and typically for at least eight hours.
Being a clinician, Foldvary-Schaefer’s says her “dream scenario” is for every sleep report to have a risk attached to it based on the foundation model work, which would have practical benefits for patients—including important information for longevity. “I don’t know of a single test in medicine that can predict mortality ... the sleep study is the only test that brings [so many] sensors in so beautifully together.”