Stanford researchers built a model that diagnoses 130 diseases from a single night of sleep data. The accuracy numbers are unsettling.
A team at Stanford Medicine trained an AI on 585,000 hours of sleep recordings from 65,000 people. The model, called SleepFM, can now predict whether you'll develop cancer, Parkinson's disease, or dementia — from one night of sleep.
Published in Nature Medicine in January 2026, it's the first foundation model built on polysomnography data. And its accuracy isn't in the "promising early results" range. It's in the "better than your doctor" range.
What the numbers say
Polysomnography records everything: brain waves, heart rhythm, respiratory rate, eye movements, muscle activity, blood oxygen. Hospitals use it to diagnose sleep apnea. SleepFM looked at the same data and found something else entirely.
The model scanned over 1,000 disease categories and found 130 conditions it could predict with a C-index above 0.75 — the threshold considered clinically useful. The standouts:
- Prostate cancer: 0.89
- Parkinson's disease: 0.89
- Breast cancer: 0.87
- Dementia: 0.85
- Hypertensive heart disease: 0.84
- Death: 0.84
- Heart attack: 0.81
For context, a C-index of 0.5 is a coin flip. Medical screening tools typically score between 0.7 and 0.8. SleepFM is beating established clinical benchmarks for conditions that have nothing obvious to do with sleep.
How it works
SleepFM is a foundation model — the same architecture behind GPT and Claude, adapted for physiological signals instead of text. It was trained on recordings from roughly 35,000 patients, ages 2 to 96, collected at Stanford's sleep clinic between 1999 and 2024. Those recordings were matched to up to 25 years of follow-up medical records.
The model breaks sleep into five-second segments and analyzes multiple signal channels simultaneously — brain activity against heart rhythm against respiratory flow. Co-senior author James Zou noted that the most predictive information came from "contrasting the different channels." The disease signal isn't in any one measurement. It's in the relationship between them.
That's the part no human clinician could replicate. A sleep technologist reads an EEG. A cardiologist reads a heart rhythm. SleepFM reads both at once and finds patterns that exist only in the cross-channel interference.
The gap between lab and life
Here's where it gets complicated. Polysomnography requires a clinical sleep lab. Sensors glued to your scalp. Wires across your chest. An overnight stay in a hospital bed. It's expensive, invasive, and available to a small fraction of the population.
Consumer wearables — Apple Watch, Oura Ring, Whoop — track sleep too. But their accuracy drops off fast. Oura achieves 76-80% sensitivity for sleep staging. Apple Watch overestimates total sleep time by nearly 40 minutes on average and has a wake-detection specificity of just 47%. None of them capture brain waves. None of them measure respiratory airflow. None of them produce the multi-channel data SleepFM needs.
The model that can predict cancer with 87% accuracy requires equipment that costs thousands of dollars and a night in a hospital. The device on your wrist that tracks sleep for free can't tell if you're awake.
What this actually means
SleepFM doesn't need to reach your Apple Watch to matter. It needs to reach the 39 million Americans who already get polysomnography each year for suspected sleep apnea. Those patients already generate the data. The diagnosis is already happening. SleepFM would just read the same recording and flag 130 additional conditions the sleep doctor wasn't looking for.
That's the real application: opportunistic screening. You went to the hospital for snoring. You leave with an early Parkinson's warning.
The model was funded by NIH and the Chan-Zuckerberg Biohub. Lead researchers include Rahul Thapa and Magnus Ruud Kjaer, with co-senior authors Emmanuel Mignot (Stanford's Craig Reynolds Professor in Sleep Medicine) and James Zou (biomedical data science). It was validated across multiple international cohorts, including data from the Technical University of Denmark and Harvard Medical School.
The question no one's asking
If a single night of sleep contains enough information to predict cancer five years out, what does that say about every night of sleep data that Apple, Google, Samsung, and Amazon are already collecting?
Those companies own billions of hours of sleep recordings. Lower fidelity than polysomnography, but improving every hardware generation. The FDA cleared an AI algorithm in 2024 to screen for sleep apnea using consumer pulse oximetry alone. The gap between clinical-grade and consumer-grade is closing.
SleepFM proves the signal exists. The question is who gets to read it — your doctor or your insurer.\n\n---\n\n*If you work with AI tools daily, check out my AI prompt engineering packs — battle-tested prompts for developers, writers, and builders.*
Top comments (0)