A bad night of sleep usually means feeling tired the next day. But it could also be a warning sign for serious health problems that may develop years later.
Scientists at Stanford Medicine have built a new artificial intelligence (AI) tool that can look at one night’s sleep and predict a person’s risk for over 100 diseases.
This new model is called SleepFM. It was trained using nearly 600,000 hours of sleep data from 65,000 people. The data comes from a test called polysomnography.
This test is the gold standard in sleep studies. It records many body signals during sleep, such as brain waves, heart rate, breathing, and eye and leg movements.
SleepFM uses AI to study this huge amount of sleep data and find patterns that humans may not notice. The researchers say that even though we collect many signals during a sleep study, only a small part is normally used. Now, thanks to AI, they can learn much more from this data.
SleepFM is similar to language models like ChatGPT, but instead of learning from words and sentences, it learns from short five-second pieces of sleep data. Each piece is like a “word” in the language of sleep. The model can study different signals, understand how they relate to each other, and figure out patterns linked to health.
To build this model, the team created a new training method called “leave-one-out contrastive learning.” This means they removed one signal and asked the model to guess it using the others. This helped the model learn how different signals work together.
After training, the team tested SleepFM in a few ways. First, they used it to do normal sleep study tasks like checking sleep stages and identifying sleep apnea. The model did as well or better than today’s best tools.
Then, the researchers tried something harder: using one night of sleep data to predict future disease. They had long-term health records for many of the people in their sleep studies—up to 25 years of follow-up in some cases.
The results were impressive. SleepFM could predict 130 diseases with good accuracy. It worked especially well for conditions like Parkinson’s disease, dementia, heart problems, cancer, pregnancy complications, and mental illnesses.
The model had a C-index score above 0.8 for many of these diseases. This score shows how well the model can tell who is more likely to get sick first. A score of 0.8 means the model is right 80% of the time.
For example, the model was very accurate at predicting prostate cancer (score 0.89), Parkinson’s disease (0.89), and breast cancer (0.87). It also predicted risk of death (0.84) and heart attacks (0.81) quite well.
The team is now working on making the model even better. They want to include data from wearable devices and find ways to understand how the model makes its predictions. Right now, the AI doesn’t explain things in human language, but the researchers are building tools to help interpret what the model is “seeing” in the data.
Interestingly, the model gets the best results when it looks at all the sleep signals together. For example, it might notice when a person’s brain shows they are asleep but their heart looks awake—and this mismatch might mean future health trouble.
This research was led by Dr. Emmanuel Mignot and Dr. James Zou. Other scientists from Denmark, Harvard, and hospitals in Europe also helped. The study shows that sleep is not just for rest—it could be a powerful way to detect disease early, long before symptoms begin.
If you care about sleep health, please read studies about foods that help people sleep better, and Keto diet could improve cognitive function in people with sleep loss.
For more health information, please see recent studies about the natural supplements for sound sleep, and how your diet can improve sleep quality.
The study is published in Nature Medicine.
Copyright © 2025 Knowridge Science Report. All rights reserved.
