A study published in Nature Medicine in January 2026 revealed that sleep is not just a moment of rest. Sleep is also a powerful biological signal that can predict future disease risks. Researchers showed that by developing a multi-mode artificial intelligence model called SleepFM, they could predict the risk of more than 130 diseases with high accuracy from sleep recordings of a single night (https://www.nature.com/articles/s41591-025-04133-4).

The study is based on a massive dataset comprising over 585,000 hours of polysomnography (PSG) recordings from approximately 65,000 individuals. PSG is a gold-standard sleep study method that simultaneously records multiple physiological signals, including brainwaves (EEG), heart rate (ECG), respiration, and muscle activity. SleepFM is designed as a foundation model that learns the “language” of sleep by analyzing these different signals together.

One of the most striking aspects of the study is its high predictive accuracy for cardiovascular diseases. The model can predict outcomes such as heart failure, stroke, myocardial infarction, and death due to cardiovascular causes with significant accuracy.

In particular, external validation performed on an independent dataset yielded an AUROC value of 0.88 for cardiovascular death. This value indicates a very strong discriminatory power in clinical predictive models. Similarly, high accuracy values were obtained for stroke and heart failure. It is stated that the combined evaluation of ECG signals and respiratory parameters plays a crucial role in the model’s success.

These findings suggest that physiological signals recorded during sleep can detect cardiovascular risks early on, even before they manifest clinically.

Today, sleep tests in clinical practice are mostly used to diagnose problems such as sleep apnea, insomnia, or excessive daytime sleepiness. However, this study reveals that sleep data has a much broader potential:

These findings mean the following in daily practice:

• A patient’s sleep record for one night can provide predictions about their future cardiovascular disease risk. This can enable early intervention, especially in asymptomatic individuals.
• High-risk patients can be individually referred for more intensive lifestyle interventions, close monitoring, or further investigations.
• In family medicine and cardiology practice, artificial intelligence models integrated into electronic patient records can provide physicians with objective risk scores.
• It can be more rationally determined who should undergo further investigation or which patients should be monitored more frequently.

On the other hand, the study population in this research mainly consists of patients who applied to sleep clinics; therefore, direct generalization to the general population may be limited. In addition, the fact that the decision-making mechanism of the artificial intelligence model is not fully explainable remains a subject of debate regarding clinical acceptance. Therefore, the results should be seen as tools to support physician assessment, not as a replacement for it.

The study is based on retrospectively matching sleep lab data with electronic patient records. However, it is unclear whether all patients who underwent PSG were followed up in the same health system over the long term. The lack of a central national EHR infrastructure in the US means that some diagnoses may have been recorded in different institutions and not reflected in the dataset. This situation may pose a risk of loss to follow-up, particularly in long-term disease predictions, and could impact model performance.

Ultimately, despite the limitations of this research, the researchers emphasize that models like SleepFM can be integrated with sleep data from wearable devices in the future. As smartwatches and home sleep sensors become more prevalent, non-invasive and continuous health monitoring may become a possibility in the near future.