SleepFM is an AI system developed at Stanford that analyzes sleep clinic data to predict the risk of developing over 130 diseases, such as Parkinson’s and heart failure, years before symptoms appear. By training on hundreds of thousands of hours of sleep recordings linked to medical records, SleepFM has achieved high accuracy in disease prediction, though it currently works only in clinical sleep lab settings.
SleepFM is an advanced AI system developed to predict future diseases by analyzing sleep data. Traditionally, sleep clinics collect extensive physiological data from patients overnight, including brain waves, heart rate, breathing, and muscle activity. For decades, this data has primarily been used to diagnose sleep apnea, leaving a wealth of other information largely untapped. Researchers at Stanford recognized the potential value in this unused data and set out to find new ways to utilize it for broader health predictions.
To build SleepFM, the team trained a foundation model on 585,000 hours of sleep recordings from 65,000 individuals. The data included a comprehensive range of physiological signals captured during sleep. By linking these sleep recordings to long-term medical records, researchers could identify which patients developed diseases or died years after their sleep studies. This allowed the AI to learn associations between specific sleep patterns and future health outcomes.
The model employed a technique called “leave one out contrastive learning,” which helps the AI recognize and match patterns across different body signals from the same night. This approach enabled SleepFM to detect subtle relationships between various physiological signals—such as how certain brain wave patterns might correspond with specific heart rhythms. As a result, the AI began to uncover predictive patterns that were previously invisible to human doctors.
SleepFM demonstrated remarkable accuracy in predicting a wide range of diseases. For example, it could identify early signs of Parkinson’s disease from REM sleep signals years before symptoms appeared, and it could predict the risk of stroke based on heart rhythms during deep sleep up to five years in advance. Overall, the model could make 130 different disease predictions from a single night of sleep data, achieving high accuracy rates: 87% for dementia, 93% for Parkinson’s, 80% for heart failure, and 84% for predicting death within six years. These results significantly outperformed traditional clinical models.
However, there are important limitations to consider. The current version of SleepFM only works in sleep labs equipped with full clinical setups, and most of the study participants were already ill when their data was collected. The next step for researchers is to test the model’s effectiveness on healthy individuals and in more accessible settings. Despite these challenges, SleepFM represents a promising example of how AI can unlock valuable insights from existing medical data to improve disease prediction and prevention.