New AI Model Predicts Future Disease Risk Using Just One Night of Sleep
A poor night’s sleep is usually blamed for grogginess, low focus, or a bad mood the next day. But new research suggests it may also quietly reveal something far more serious: your long-term health risks. Scientists at Stanford Medicine and their collaborators have developed a powerful artificial intelligence model that can analyze data from a single night of sleep and predict the likelihood of developing more than 100 different diseases, sometimes years or even decades in advance.
The model, called SleepFM, represents one of the most ambitious efforts so far to unlock the hidden health signals embedded in sleep data. It goes far beyond identifying sleep disorders and instead treats sleep as a window into the body’s overall physiological health.
Why Sleep Data Is a Hidden Gold Mine
Sleep is one of the few times when the body can be monitored continuously for many hours under controlled conditions. In clinical sleep studies, patients undergo polysomnography, a comprehensive overnight assessment that records a wide range of physiological signals at the same time. These include brain activity, heart rhythms, breathing patterns, eye movements, muscle activity, leg movements, and airflow, among others.
Despite the richness of this data, only a small portion is typically used in routine sleep medicine. Most analyses focus on identifying sleep stages or diagnosing conditions like sleep apnea. According to the researchers, this leaves a vast amount of potentially valuable information unexplored.
SleepFM was designed to change that by analyzing sleep data in its entirety and learning how different physiological systems interact during sleep.
How SleepFM Was Built and Trained
SleepFM is what researchers call a foundation model, a type of AI system trained on massive amounts of raw data so it can later be adapted to many different tasks. Large language models are foundation models trained on text. SleepFM, in contrast, was trained on sleep physiology.
The researchers used an enormous dataset comprising approximately 585,000 to 600,000 hours of polysomnography recordings collected from around 65,000 patients. These recordings came from sleep clinics and span a wide age range, from children as young as 2 to adults up to 96 years old.
To make this data usable for AI training, the sleep recordings were broken into five-second segments, similar to how language models treat words or tokens in text. Over time, the model learned recurring patterns and relationships across multiple physiological signals.
A key technical innovation behind SleepFM is a training method known as leave-one-out contrastive learning. During training, the model is deliberately given incomplete information by hiding one type of signal, such as brain activity or heart rhythm, and is challenged to reconstruct it using the remaining signals. This forces the AI to understand how different systems in the body relate to one another during sleep.
Testing the Model on Traditional Sleep Tasks
Before tackling disease prediction, the researchers first evaluated SleepFM on standard sleep medicine tasks. These included classifying sleep stages and assessing the severity of sleep apnea.
In these benchmarks, SleepFM performed as well as or better than existing state-of-the-art models currently used in sleep research and clinical settings. This confirmed that the model could accurately interpret sleep physiology before being applied to more ambitious goals.
Linking Sleep to Long-Term Health Outcomes
The most striking aspect of the research came when the team paired sleep data with long-term electronic health records. The Stanford Sleep Medicine Center, founded in 1970 by William Dement, has accumulated decades of patient data, making it uniquely suited for this type of analysis.
For around 35,000 patients, researchers were able to link polysomnography data recorded between 1999 and 2024 with up to 25 years of follow-up health records. In total, the model examined over 1,000 disease categories documented in these records.
Out of these, 130 diseases could be predicted with meaningful accuracy using sleep data alone.
Diseases SleepFM Can Predict Particularly Well
SleepFM showed especially strong performance for several major disease categories, achieving C-index values above 0.8, a level generally considered strong in medical prediction models. The C-index measures how well a model can correctly rank which individuals are more likely to experience a disease sooner than others.
Some of the most notable predictions included:
- Parkinson’s disease, with a C-index of 0.89
- Prostate cancer, also around 0.89
- Breast cancer, approximately 0.87
- Dementia, around 0.85
- Hypertensive heart disease, about 0.84
- All-cause mortality, roughly 0.84
- Heart attack, around 0.81
The model also performed well for circulatory disorders, mental health conditions, and pregnancy-related complications. Even for diseases with slightly lower accuracy scores, the researchers noted that models with C-indices around 0.7 are already considered clinically useful in many areas of medicine.
What the Model Is Actually Detecting
One of the most intriguing findings is that no single signal appears to be responsible for these predictions. While heart-related signals contribute more to cardiovascular predictions and brain signals matter more for neurological conditions, the highest accuracy comes from combining all data streams.
In particular, the model seems sensitive to situations where different systems are out of sync. For example, a brain pattern suggesting deep sleep combined with heart signals indicating heightened alertness may be a subtle warning sign of underlying disease risk.
Although SleepFM does not explain its predictions in human language, the team is developing interpretation techniques to better understand what features the model relies on for specific diseases.
Why Sleep Is Such a Powerful Health Indicator
Sleep is increasingly recognized as a cornerstone of long-term health. During sleep, the body regulates hormones, repairs tissues, consolidates memory, and balances the autonomic nervous system. Disruptions in these processes often occur long before clinical symptoms appear.
This makes sleep an ideal signal for early disease detection. Unlike blood tests or imaging scans, sleep monitoring captures the dynamic interaction of multiple systems over many hours, providing a uniquely holistic view of health.
Limitations and What Comes Next
Despite its promise, SleepFM is not yet ready for clinical deployment. The researchers emphasize that further validation is needed, especially across more diverse populations and healthcare settings.
Future work may involve integrating data from wearable devices, which could make this type of analysis more accessible outside specialized sleep labs. There is also ongoing research into improving the model’s interpretability so clinicians can better understand and trust its predictions.
A Shift Toward Preventive Medicine
SleepFM highlights a broader shift in healthcare toward predictive and preventive medicine. Instead of waiting for symptoms to appear, models like this could one day help identify high-risk individuals early and guide lifestyle changes or medical monitoring long before disease develops.
For now, the study serves as a powerful reminder that sleep is not just a passive state of rest. It is a rich physiological process that may hold early clues about the future of our health.
Research paper:
A multimodal sleep foundation model for disease prediction, Nature Medicine (2026)
https://doi.org/10.1038/s41591-025-04133-4
