What SleepFM can learn about future health from a night in a sleep laboratory

SleepFM’s study published online in Nature Medicine in January 2026 tested later disease risks in 5,019 held-out Stanford participants. For 130 conditions, both the concordance index and the area under the receiver operating characteristic curve (AUROC) reached at least 0.75. These measurements concern discrimination of subsequent risk, not overnight diagnosis.[1]

The recording is much richer than a nightly sleep score

Polysomnography, or PSG, records several physiological processes together. The NHLBI describes measurements including brain activity, breathing, oxygen levels and heart activity. A sleep laboratory can therefore observe relationships across signals over a night, not just the duration or a consumer device’s estimated sleep stages.[2]

Imagine reducing a concert to a single loudness number. That number could be useful, but it would not retain when different instruments entered or how their rhythms related. This is only an analogy, yet it captures the attraction of analysing several biological time series together: the relationships may carry information lost in a summary.

Earlier SleepFM research explored learning joint representations of brain, cardiac and respiratory signals. Representation learning compresses a complicated recording into features that another model can use for a particular task. The 2024 work was an earlier research stage, not the same disease-prediction experiment reported in 2026.[4]

Pretraining used approximately 65,000 participants and 585,000 recording hours. The disease model also used age and sex. Its later diagnostic labels came from health-record codes; the separate AUROC measure used a six-year horizon. The pretraining corpus and held-out test are different populations.[1]

A score of 0.80 does not mean an 80% personal risk

For a simplified concordance example, imagine ten patient pairs for which the earlier event can be compared. The model assigns a higher risk to the person with the earlier event in eight pairs. Eight divided by ten is 0.80. Actual survival analysis must handle incomplete follow-up and ties; this toy calculation isolates the ordering idea.

Nothing in that calculation says that either person has an 80% chance of becoming ill. A ranking can be useful for deciding which people merit closer investigation while still leaving another question unanswered: are the numerical probabilities well calibrated? That distinction matters before a score is turned into an appointment, a scan or a reassurance.

For external testing, the researchers adapted SleepFM using 3,291 Sleep Heart Health Study participants and evaluated it on a separate 2,000 participants. The external experiment covered six cardiovascular outcomes. It therefore tests transfer to another cohort for those outcomes, while the 130-condition result comes from the Stanford evaluation.[1]

Another study asks whether familiar categories hide different risks

In August, Bilal and colleagues reported a separate foundation model using clinical sleep recordings linked to health records. It identified five patient groups with different later health trajectories. Cleveland Clinic’s account describes approximately twice the five-year mortality risk in the highest-risk group compared with the lowest, a separation that conventional apnea–hypopnea index categories did not capture.[3][5]

The apnea–hypopnea index counts breathing interruptions per hour. The August study’s broader grouping asks a different question: do patients whose nights contain similar combinations of physiological behaviour also have similar later outcomes? It reported transfer to the Sleep Heart Health Study, extending the investigation beyond its original clinical data.[3]

Together, these studies suggest that existing recordings contain additional prognostic information. Whether using that information improves clinical outcomes requires a further test of the decisions and care it produces. Neither study establishes a benefit from arranging sleep testing solely for broad disease prediction.[3]

The intriguing part is not that sleep contains a hidden diagnosis of everything. It is that hours of coordinated physiology may remain useful after the original sleep question has been answered. These studies make that possibility measurable, with defined cohorts and endpoints rather than a mysterious all-purpose sleep score.

Report an error or suggest a correction