Can sleep data predict future disease?

A night of sleep contains a great deal of information

A sleep record may first bring to mind how many hours you slept.

Detailed sleep tests in medical settings collect much more information.

Brain waves, heart activity, breathing, blood oxygen and muscle movements: sensors on the body record these signals throughout the night. This test is called polysomnography, or PSG (NHLBI, “Sleep Studies”).

These tests investigate sleep problems such as obstructive sleep apnoea.

The record captures how the body behaved during sleep as well as its duration. Researchers are exploring whether this rich information can contribute more to understanding health.

AI that learned from about 65,000 people’s sleep

The SleepFM study was published in Nature Medicine in January 2026.

The researchers used more than 585,000 hours of sleep-test data from approximately 65,000 participants for learning. The model handles multiple physiological signals and combines age and sex with the recordings for disease-risk prediction.

The study reported predictive performance for later dementia, myocardial infarction and other outcomes from one night’s recording. For dementia, a metric called the C-index was 0.85 (Thapa et al., 2026).

What does 0.85 tell us about how well the model predicts?

0.85 is not your probability of developing disease

The C-index measures agreement between the order of observed events and the order of risks assigned by the AI.

Broadly, among two people who can be compared, it asks whether the person whose event actually occurred earlier was also ranked at higher risk.

A result of 0.85 therefore does not mean you have an 85% chance of dementia, or that the model correctly diagnosed 85 of 100 people.

Ranking risks well and accurately estimating a percentage over a specified number of years also need separate assessment. Checking whether predicted probabilities agree with observed event rates is called calibration (Riley et al., 2024).

You do not need to memorise the terminology.

When a study mentions accuracy, ask what was predicted and how the figure was calculated. That small step can change how you read the result.

A recorded diagnosis date is not the date a disease began

This research links sleep-test recordings with diagnoses and other information subsequently recorded in electronic health records (Thapa et al., 2026).

It helps to distinguish the date a diagnosis was recorded from the date disease began within the body.

Changes may progress gradually before a later test or consultation leads to a diagnosis. Predicting a recorded diagnosis does not establish the moment the disease started.

Likewise, predicting disease from a sleep feature does not prove that the feature causes the disease.

Conclusions such as “increasing deep sleep prevents this disease” would require further, different research.

Whose data were used also matters

The original SleepFM paper explains that much of the data came from people tested for suspected sleep problems or related indications. Some diseases were evaluated in another population, but not every outcome could be checked in the same way (Thapa et al., 2026).

Performance in people undergoing testing does not necessarily apply unchanged to a broad population going about daily life.

This principle matters across medical AI.

Does the model help people different from those used for learning? Do results hold in another hospital or time period? Evaluation should resemble the conditions of intended use (Collins et al., 2024).

Both the size of a dataset and its relevance to the intended users matter.

From sleep-laboratory learning to wrist sensors

This research is moving further.

SleepFM-2, released in September 2026 and revised on 2 October, uses more than two million hours of sleep recordings.

Such AI learns features from many signals, then uses them for tasks including sleep-stage classification and disease-risk prediction. Because it provides a basis for different analyses, it is called a foundation model.

The follow-up explores sleep staging from head-worn EEG and wrist pulse signals, as well as disease-risk prediction from wrist movement recordings. The data and analysis methods differ by task (Thapa et al., SleepFM-2, 2026).

As of 19 October 2026, this work is a preprint that has not undergone peer review. Its findings need to be read alongside future validation.

Can features learned from comprehensive sleep tests also work with more accessible sensors? That is an interesting question raised by this follow-up.

Can Apple Watch sleep records do the same thing?

You may now feel curious about giving your own sleep graph to an AI.

But displayed minutes of deep sleep or a sleep score are not the same as the physiological signal recordings used in research.

Apple’s 2025 technical documentation describes estimating Apple Watch sleep stages from accelerometer signals. The watch does not directly measure brain waves (Apple, “Estimating Sleep Stages from Apple Watch”).

Even research using wrist movement signals does not necessarily input only the sleep durations or scores visible on a screen.

Which sensor collected which signals, and how were they analysed? Those differences need checking when connecting research findings to everyday devices.

What needs testing beyond successful prediction?

For a technology to support health, its value beyond predictive performance matters.

What tests or consultations follow a high-risk result? Do those decisions bring necessary care sooner? Can unnecessary testing and burdens be avoided?

Medical prediction models also need evaluation of this practical usefulness (Riley et al., 2024).

Sleep data may contain health clues that have not yet been fully used. AI is gradually expanding that possibility.

This morning’s sleep graph is not a forecast announcing future disease.

Still, carefully studying the body each night might one day help bring appropriate care to people who need it. SleepFM is one line of research pointing toward that future.