SleepFM was trained on nearly 600,000 hours of polysomnography from approximately 65,000 participants. A single night of sleep data predicts over 100 future diseases — Parkinson's (C-index 0.89), prostate cancer (0.89), breast cancer (0.87), dementia (0.85).
The model uses brain, heart, respiratory, and movement signals recorded during sleep. Nothing else. No blood tests, no imaging, no family history, no symptoms. One night of sleep contains enough physiological signal to predict diseases that will manifest years later.
The technical innovation is leave-one-out contrastive learning, which harmonizes multiple physiological data streams into a shared representation. But the conceptual finding is more important than the method: sleep is a diagnostic window. During sleep, the body reveals subclinical signatures of disease that are masked during waking hours.
The masking makes biological sense. Waking physiology is dominated by voluntary control — you regulate your breathing, posture, heart rate, and movement in response to the environment. During sleep, voluntary control is suspended. The autonomous systems run without oversight. And autonomous systems running without oversight reveal their degradation more honestly than supervised systems being actively corrected.
The analogy is a factory inspection. When the manager is watching, the workers compensate for equipment flaws. When the manager leaves, the flaws become visible. Sleep is the manager leaving. The subclinical Parkinson's tremor that's suppressed by waking motor control surfaces as a sleep movement pattern. The cardiovascular strain that's compensated by waking baroreflexes appears as a nocturnal heart rate anomaly.
The through-claim: the most informative measurement of a system may come when the system is not actively performing. Waking physiology is a controlled performance. Sleeping physiology is the rehearsal where the understudy doesn't know the audience is watching — and the understudy's mistakes reveal what the lead performer has been hiding.