A single night of sleep, recorded in a standard sleep test, can support predictions for 130 conditions, according to a Stanford-led study in Nature Medicine. During the test, sensors track brain activity, heart rhythm, breathing, oxygen levels and muscle movement. The test is costly because it needs specialist staff, controlled conditions and expert interpretation. Yet in routine practice its data is used mainly to diagnose sleep apnea, even though the test can pick up a broader range of sleep disorders. Much of what it records therefore goes unused.
The study’s AI model, SleepFM, puts that unused data to work. It is a single model, trained once and reused for many predictions, and it learned from 585,000 hours of sleep recordings from about 65,000 people. For each night, it takes in readings from several body systems together, rather than one at a time.
SleepFM then matches those sleep patterns against what later appears in patients’ medical records. It learns those links directly from large archives of recordings, rather than relying on a small set of hand-picked measures. The conditions it covers include dementia, heart failure, chronic kidney disease, stroke and death from any cause, along with a wider set of heart and metabolic conditions. All of that comes from a test hospitals already pay to run.
Those predictions only have value if they change the care a patient receives. A prediction that leaves the patient’s path through care untouched adds work and risk without improving outcomes. One that sends patients down clearer routes, such as faster follow-up for higher-risk patients, can shorten the time to treatment and make better use of scarce specialists. That is where cost, speed and quality of service improve together.
The study is also clear about its limits. The work depends on full recordings from clinical sleep tests rather than consumer wearables, so its home is sleep labs and hospitals. The people in it had been referred for sleep studies and may have been sicker than the general population, which limits how far the results can be applied. The study itself says wider use depends on testing in new groups and settings, including healthier people.
