Now loading...
Researchers at Stanford Medicine have developed an artificial intelligence system that analyzes a single night’s sleep data to forecast the likelihood of over 100 medical conditions, positioning sleep as a vital indicator of future health rather than mere rest.
The innovation, known as SleepFM, draws on vast amounts of polysomnography recordings, which track brain waves, breathing patterns, muscle movements, heart rhythms and eye activity during sleep studies. Trained on about 600,000 hours of such data from around 65,000 individuals, the model identifies subtle patterns that signal potential health risks years in advance.
Emmanuel Mignot, a professor of sleep medicine at Stanford and one of the study’s co-lead authors, highlighted the wealth of information captured in these overnight sessions. For years, sleep laboratories have gathered detailed physiological profiles, but much of that detail went unexplored beyond diagnosing issues like sleep apnea.
SleepFM functions as a foundational AI tool, similar to those that process language, but here it deciphers the intricacies of sleep by dividing recordings into five-second intervals and examining how various bodily signals interact. A novel aspect involved simulating missing data streams, compelling the model to learn the interconnections between organ systems and physiological processes.
Initial evaluations showed the system matching or surpassing existing methods for classifying sleep phases and gauging apnea intensity. More significantly, when integrated with extended patient health histories covering up to 25 years, SleepFM demonstrated reliable projections for 130 specific ailments out of over 1,000 examined categories.
The predictions proved most robust for cardiovascular issues, various cancers, complications during pregnancy and psychiatric conditions. It excelled particularly in anticipating Parkinson’s disease, dementia, myocardial infarction, hypertension-related heart problems, and malignancies such as those affecting the breast and prostate, often achieving over 80 percent accuracy in determining the sequence of onset among participants.
James Zou, an associate professor at Stanford and the other co-lead author, expressed surprise at the model’s consistency across varied scenarios. The key revelations stemmed not from isolated metrics but from discrepancies among signals, such as a relaxed brain paired with an elevated heart rate, which could indicate systemic strain.
These findings align with perspectives in aging research, where health decline often arises from desynchronized bodily functions rather than isolated failures. While SleepFM is not intended for direct diagnosis or to supplant clinical expertise, it suggests sleep tracking could evolve into a routine method for spotting early warnings, especially alongside devices like fitness wearables.
The team plans further work to refine forecasts with supplementary information and to unpack the model’s internal logic using tailored analytical approaches. This approach underscores a shift in viewing sleep as an ongoing chronicle of physiological harmony and potential vulnerabilities.
For more details, see the study published in Nature Medicine and related coverage from Stanford Medicine.
