Can your sleep study predict your risk of dying?
Yes, far better than the apnea score it usually gets reduced to. When an AI model read the raw signals from 10,000 overnight sleep studies, it sorted patients into five risk groups, and the highest-risk group had more than double the risk of dying of the lowest. The apnea score those same patients were given predicted nothing about survival.
If you have had an overnight sleep study, you probably left with one number: the apnea-hypopnea index, or AHI. It counts how many times per hour your breathing stops or gets shallow. That single count decides whether you are labeled mild, moderate, or severe, and whether you get a CPAP machine.
A sleep study records far more than breathing. It captures brain waves, eye movements, muscle tone, heart rhythm, chest and belly effort, airflow, snoring, and blood oxygen all night long. Almost all of that gets thrown away once the AHI is calculated. This study asked what is hiding in the part we discard.
What the data show
Researchers at the Cleveland Clinic trained a transformer model, the same family of AI that powers chatbots, on the raw signals from 10,000 in-lab sleep studies. Clustering the results produced five patient groups. Risk climbed steadily from group 1 to group 5.
The highest-risk group had about 171% higher risk of death from any cause than the lowest-risk group, after adjusting for age and gender (hazard ratio 2.71). The true effect is very likely somewhere between 93% and 281% higher (95% CI 1.93 to 3.81), and there is far less than a 1 in a million chance this is coincidence.
The gap held up when the researchers put the apnea score itself into the model. Cluster membership still predicted death, with about 138% higher risk in the top group (hazard ratio 2.38). In matched comparisons, 80.6% of the lowest-risk group stayed free of major heart events versus 67.3% of the highest-risk group. That is a difference of about 133 people per 1,000. For cognitive impairment, the numbers were 84.3% versus 75.2%, a difference of about 91 per 1,000.
The highest-risk group also carried more new disease across the board: about 123% higher risk of atrial fibrillation (hazard ratio 2.23), 140% higher risk of epilepsy (2.40), 93% higher risk of cognitive impairment (1.93), and 84% higher risk of heart attack (1.84).
Dr. Kumar’s Take
The most damning number in this paper is the one that did nothing. Mild, moderate, and severe apnea categories showed no significant link to death at all. In one analysis that excluded patients on CPAP, severe apnea carried a hazard ratio of 1.05, meaning essentially no measurable difference from having a normal score.
I have watched patients be reassured by a low AHI and watched others panic over a high one. This paper suggests both reactions can be wrong. The high-risk group included people from every apnea severity band, including people whose breathing looked fine. The number that determines who gets treated is not the number that tracks who gets sick.
How the study was done
The cohort came from the Cleveland Clinic STARLIT-10K registry: 10,000 in-lab sleep studies from 9,661 patients, linked to their medical records. After quality control, 9,608 studies from 9,297 patients were used. The median observation window ran 15.1 years, with a mean of about 14.5 years, so outcomes were tracked over a long stretch of real clinical care.
The model was trained to do three ordinary jobs first: score sleep stages, spot breathing events, and detect oxygen drops. Those tasks were a means to an end. They forced the model to learn what sleep physiology actually looks like, and the internal representation it built was then used to group patients.
Does it hold up outside one hospital?
The team applied the same pipeline, unchanged, to the Sleep Heart Health Study, a community-based cohort of more than 6,000 people recorded with fewer channels and lower resolution. The pattern repeated. The highest-risk group had about 58% higher risk of death (hazard ratio 1.58, very likely between 7% and 132% higher), and about 113% higher risk of new heart failure (2.13). That is roughly a 2 in 100 chance of coincidence for each.
This matters because the original apnea-based analyses of that same cohort only found a mortality signal in men aged 40 to 70 with severe disease, and none in women. The AI-derived groups worked without those restrictions.
What this does not tell you yet
This is an observational study. It shows that certain sleep patterns travel with worse outcomes, not that fixing those patterns changes anything. Nobody has run a trial where high-risk patients get a different treatment and live longer.
The main cohort was people referred to a sleep lab, which is not the general population. The model’s code is not public, and the Cleveland Clinic data is not shareable, so independent teams cannot rebuild this exactly. The outside validation is real and reassuring, but it is one cohort.
Practical Takeaways
- If you have had a sleep study, ask for the full report rather than just the AHI number, since the recording contains brain, heart, and oxygen data that the summary score leaves out.
- A normal or mild apnea score is not a clean bill of health for your heart and brain, because in this study high-risk patients were spread across every apnea severity category.
- Nothing here changes today’s treatment rules, so if you have been prescribed CPAP for diagnosed sleep apnea, keep using it while this approach is still being tested.
- Track the standard risk factors you can act on now, including blood pressure, weight, blood sugar, and total sleep time, since the highest-risk patients in the validation cohort slept notably short nights.
Related Studies and Research
- Insomnia plus sleep apnea raises heart disease risk nearly 4-fold
- Bright evening light linked to higher risk of eye disease
- Ultra-processed foods linked to 47% higher heart disease risk
- Nearly 9 in 10 US heart disease deaths are preventable
FAQs
Can I find out which risk group I am in?
Not yet. This model lives in a research setting and has not been approved or released as a clinical tool, so no sleep lab can currently assign you a group. The researchers did find a simpler shortcut worth watching: a measure they call spectral sleep fragmentation, which tracks how often sleep flips between stages in under ten minutes, predicted the simpler two-group split with more than 90% accuracy on its own. That kind of measure could be calculated from data labs already collect, which is the most likely path from this paper to your next appointment.
Does this mean the apnea score is useless?
It means the apnea score answers a narrower question than most people think it does. AHI describes how often your airway collapses, and it remains the basis for diagnosis and insurance coverage under current guidelines. What this study challenges is the habit of reading it as a general health risk score. When the researchers tried to reproduce their five groups using conventional sleep measures, including AHI, arousal index, sleep stages, and oxygen levels, accuracy fell apart, meaning the risk signal is not a repackaged version of numbers we already report.
Why would brain waves matter for heart risk?
The researchers tested this directly by blanking out individual signal types and rerunning the analysis. When they zeroed out the brain wave channels, patients got reshuffled across risk groups, and the same happened when they zeroed out the heart rhythm channel. That tells us the risk signal is not just breathing in disguise. Sleep involves the nervous system and the cardiovascular system working together all night, and disruption in one shows up in the other, which is likely why a measure built only from breathing events misses so much.
Bottom Line
A routine overnight sleep study carries far more information about your future health than the single apnea score you are handed. When an AI model read the raw recordings from 10,000 patients, it found five distinct groups whose risk of death, heart disease, and neurologic disease separated sharply, with more than a doubling of mortality risk in the top group, while the conventional apnea categories predicted nothing. The finding repeated in an independent community cohort. This is not a test you can order today, but it is strong evidence that the number sleep medicine has leaned on for decades is measuring the wrong thing.
