Stanford researchers have published SleepFM, a new AI foundation model designed to extract long-term health risk signals from a single overnight sleep recording. The work suggests that sleep—captured through brain waves, heart activity, breathing, and muscle signals—may contain early warning patterns for a wide range of conditions, including neurological and cardiovascular risks.

According to the paper, SleepFM was trained on roughly 600,000 hours of sleep data from about 65,000 participants, learning cross-signal relationships rather than treating each sensor stream in isolation. A key idea is that when physiological signals become “out of sync” (for example, deep-sleep brainwave patterns paired with unusually elevated heart activity), the model can flag the mismatch as a potential marker associated with future disease risk.

The team also reports linking decades of Stanford Sleep Clinic records to sleep measurements, enabling evaluation across a very large set of disease categories. In reported results, the model showed strong predictive performance for several headline conditions (including Parkinson’s and dementia) and for broader outcomes such as elevated risk of heart events and overall mortality risk. The authors frame the system as a research step toward earlier detection rather than a direct clinical diagnosis tool.

Why this matters: we spend roughly a third of our lives asleep, yet clinical sleep data is often used mainly for diagnosing sleep disorders. If models like SleepFM continue to validate across diverse populations and settings, sleep recordings—potentially including future wearable-grade sensors—could become a richer layer for preventive health monitoring. The paper underscores both the promise and the caution: translating these findings into real-world healthcare would require rigorous clinical validation, careful bias evaluation, and clear guidance on how such predictions should be used alongside medical professionals.