An AI read one night of sleep and forecast 130 diseases. Read the fine print.

A polysomnogram is one of the densest physiological records medicine routinely collects. Over seven or eight hours it captures brain activity, cardiac rhythm, muscle tone, eye movement, and breathing, simultaneously and continuously. Then, in ordinary clinical practice, nearly all of it is thrown away — compressed into a sleep-stage summary and an apnea-hypopnea index, and the file is archived.

On January 6, 2026, Nature Medicine published a paper arguing that the discarded portion contains a great deal of information about a person's future [s1].

What was built

Researchers at Stanford trained a model called SleepFM on roughly 585,000 hours of polysomnography from about 65,000 participants — 59,572 recordings drawn from several cohorts [s1]. The largest was 35,052 recordings from the Stanford Sleep Medicine Center spanning 1999 to 2024, which is what makes the study unusual: some of those patients have been followed for as long as 25 years, so their sleep recordings can be linked to what actually happened to them afterward [s2].

Rather than being trained to score sleep stages, SleepFM was trained with a contrastive learning approach across the different signal types in a polysomnogram — EEG, ECG, EMG, and respiratory channels — to produce a general representation of a night of sleep [s1]. "SleepFM is essentially learning the language of sleep," said James Zou, PhD, Associate Professor of Biomedical Data Science at Stanford [s2].

What it predicted

The team screened more than 1,000 disease categories and reports that 130 could be predicted from a single night's recording at a C-index of at least 0.75, after Bonferroni correction (P<0.01) [s1]. C-index is a measure of how well a model ranks who develops a condition sooner; 0.5 is a coin flip, 1.0 is perfect.

The strongest results, as reported by Stanford: Parkinson's disease at 0.89, prostate cancer at 0.89, breast cancer at 0.87, dementia at 0.85, hypertensive heart disease at 0.84, all-cause mortality at 0.84, and myocardial infarction at 0.81 [s2]. The paper reports heart failure at 0.80, chronic kidney disease at 0.79, stroke at 0.78, and atrial fibrillation at 0.78 [s1].

On conventional sleep tasks the model was competitive rather than dominant: mean F1 scores of 0.70 to 0.78 for sleep staging across cohorts, and accuracy of 0.69 for classifying apnea severity and 0.87 for classifying apnea presence [s1].

That Parkinson's disease sits at the top is not surprising to sleep clinicians — REM sleep behavior disorder is a well-established early feature of neurodegenerative disease. That prostate and breast cancer sit alongside it is much harder to explain, and the paper does not claim to explain it.

Why this is not a clinical test

The paper's limitations section is unusually forthright, and it is the part of the study that should govern how the headline numbers are read.

The population is not the general population. The authors write that the dataset "consists primarily of patients referred for sleep studies due to suspected sleep disorders or other medical conditions requiring overnight monitoring" [s1]. These are people sick enough, or symptomatic enough, to have been sent to a sleep lab. A model that ranks risk well within that group has not been shown to do so among asymptomatic people, which is exactly the population a screening test would target.

External validation was narrow. The Sleep Heart Health Study cohort was excluded from pretraining and used as a held-out test, where the model performed well — stroke at a C-index of 0.82, congestive heart failure at 0.85, cardiovascular mortality at 0.88 [s1]. But only 6 of the 1,041 disease conditions could be evaluated there, because of limited diagnostic overlap with the Stanford cohort, which the authors say "prevented a comprehensive evaluation of generalization" [s1]. The 130-disease figure has not been externally replicated.

Performance decayed over time. The authors report "some degradation in temporal test sets, highlighting the challenge of maintaining predictive accuracy over time" [s1]. A model trained on recordings from one era does not automatically hold up on recordings from a later one — equipment, scoring conventions, and patient mix all drift.

Nobody can say why it works. Interpreting the model's predictions is "inherently challenging due to the complexity of the learned features," the authors write [s1]. Stanford's own summary notes that the predictions do not come with English-language explanations and that interpretation techniques are still in development [s2]. For risk prediction this is not merely an academic concern: a doctor who cannot say what drove a prediction cannot act on it, and cannot tell whether the model has learned biology or learned an artifact of who gets referred for a sleep study.

Sleep staging still lags in places. On one external dataset the model's staging F1 fell to 0.55, behind specialized staging models [s1].

And it requires a sleep lab. Every number above comes from full polysomnography. The authors note that integrating wearable data remains unexplored [s2]. Nothing here transfers to a consumer sleep tracker.

Where it fits

This lands in a year when sleep data is being pushed toward prediction from several directions at once. The American Academy of Sleep Medicine's Q1 2026 trend report groups SleepFM alongside Mayo Clinic work on AI-enabled ECG analysis to detect obstructive sleep apnea, and FDA plans to deploy agentic AI in premarket review and post-market surveillance [s3].

The reasonable reading of SleepFM is not that a night of sleep can now tell you whether you will get dementia. It is that polysomnography contains far more predictive signal than the two or three numbers currently extracted from it, and that a large model can find some of it. The authors themselves position the model as something that "can complement existing risk assessment tools," and say integration with electronic health records, omics, and imaging is needed before deployment [s1].

What to watch

Whether the 130-disease result replicates in a population-based cohort rather than a sleep-clinic-referred one. Whether interpretability methods advance far enough to say what the model is seeing. And whether anyone attempts a prospective study — the current work is retrospective, linking archived recordings to diagnoses that have already occurred.

This article describes a research finding. SleepFM is not an approved or cleared clinical test, and nothing in it constitutes medical advice.

Sources

• [s1] Thapa, Kjaer et al., A multimodal sleep foundation model for disease prediction, Nature Medicine, 2026-01-06; 32(2):752–762.
• [s2] AI model predicts disease risk while you sleep, Stanford Medicine, 2026-01-06.
• [s3] Key trends in sleep medicine: Q1 2026, American Academy of Sleep Medicine, 2026.