Stanford's SleepFM AI Predicts 130+ Diseases From One Night's Sleep Data
Doctors traditionally examine symptoms, conduct check-ups, and review medical histories to evaluate disease risks. Now, Stanford Medicine researchers present a groundbreaking alternative. Their new artificial intelligence model, SleepFM, can forecast a person's likelihood of developing more than 100 health conditions using physiological data from just one night of sleep.
How SleepFM Works
The foundational AI model underwent training on an extensive dataset. Researchers utilized nearly 600,000 hours of sleep information gathered from 65,000 participants. This collection includes diverse sleep signals such as brain activity, heart rhythms, breathing patterns, leg movements, and eye motions. Various sensors captured these physiological readings overnight.
Dr. James Zou, an associate professor of biomedical data science and study co-author, highlighted the significance. "From an AI perspective, sleep is relatively understudied," he noted. "There's a lot of other AI work that's looking at pathology or cardiology, but relatively little looking at sleep, despite sleep being such an important part of life."
Another co-author, Dr. Emmanuel Mignot, emphasized the data richness of sleep studies. "We record an amazing number of signals when we study sleep," he explained. "It's a kind of general physiology that we study for eight hours in a subject who's completely captive."
Development and Training Process
To assemble the training data, researchers relied on a large cohort of 35,000 patients aged two to 96 years. These individuals had their polysomnography data recorded at Stanford's sleep clinic between 1999 and 2024. Approximately 585,000 hours of this sleep data paired with patients' electronic health records.
The team then split the sleep information into five-second increments. This approach mirrors how large language models use words for training. After initial training, researchers further fine-tuned SleepFM to perform various diagnostic tasks.
Impressive Performance Results
During testing, researchers evaluated SleepFM's capabilities on standard sleep analysis tasks. These included classifying different sleep stages and diagnosing sleep apnea severity. The model then tackled the more ambitious objective of predicting future disease risk from sleep data.
Analysis of over 1,000 disease types revealed SleepFM could predict 130 conditions with reasonable accuracy. For predicting various cancers, pregnancy complications, circulatory conditions, and mental disorders, SleepFM achieved an average C-index higher than 0.8.
The concordance index (C-index) measures an AI model's ability to predict which of any two individuals in a group will experience an event first. Scores above 0.8 are particularly significant since models with around 0.7 have proven useful in clinical settings.
SleepFM demonstrated remarkable performance for specific conditions:
- Parkinson's disease: C-index 0.89
- Dementia: 0.85
- Hypertensive heart disease: 0.84
- Heart attack: 0.81
- Prostate cancer: 0.89
- Breast cancer: 0.87
- Death: 0.84
Future Directions and Considerations
Researchers plan to enhance SleepFM's prediction accuracy by incorporating wearable device data into the training dataset. They also recognize the need for further investigation into how the model interprets information.
"It doesn't explain that to us in English," Zou acknowledged regarding the model's decision-making process. "But we have developed different interpretation techniques to figure out what the model is looking at when it's making a specific disease prediction."
This development arrives as AI companies increasingly enter healthcare with specialized tools and services. Recent days have seen OpenAI launch 'ChatGPT for Health' and Anthropic announce 'Claude for Healthcare.' However, the proliferation of AI-powered healthcare tools raises privacy concerns and highlights risks of 'hallucinations' that could generate inaccurate medical information.
SleepFM represents a significant advancement in sleep-focused AI research. By analyzing comprehensive physiological data from a single night's sleep, this model opens new possibilities for early disease detection and preventive healthcare strategies.
