Researchers from Stanford Medicine and collaborating institutions have developed an artificial intelligence (AI) model capable of estimating a person’s future disease risk using data from a single night of sleep. The model, called SleepFM, was reported by Stanford Medicine in January 2026 and formally described in Nature Medicine.1
The study highlights how sleep data, when analyzed using advanced machine-learning techniques, may serve as an early indicator of long-term health outcomes.
SleepFM is a foundation AI model trained to analyze full-night polysomnography recordings. Polysomnography is a clinical sleep test that measures brain waves, heart rhythm, breathing patterns, muscle activity, eye movements, and blood oxygen levels throughout sleep.1
The model learns complex physiological patterns from these signals and links them to future health events documented in electronic medical records.
The researchers trained SleepFM using more than 585,000 hours of sleep recordings from approximately 65,000 individuals. Data were collected from multiple sleep centers, primarily in the United States. Participants ranged in age from 2 to 96 years, allowing the model to capture sleep patterns across the lifespan.1
Health outcomes were tracked over several years using linked clinical records.
The findings were published in Nature Medicine in January 2026.2 Researchers emphasize that sleep reflects multiple body systems working simultaneously, making it a valuable but underused source of health information. By analyzing these signals together, SleepFM identifies patterns associated with disease development long before symptoms appear.1
SleepFM divides sleep data into short five-second segments and processes them using a transformer-based architecture similar to models used in natural language processing. The model was first validated on standard sleep tasks, such as identifying sleep stages and detecting sleep apnea.1
After validation, researchers assessed whether sleep features could predict future disease. The model estimated risk for 130 medical conditions, including cardiovascular disease, neurodegenerative disorders, certain cancers, and all-cause mortality. Many predictions achieved a concordance index above 0.8, indicating strong predictive performance.1,2
Conditions most strongly associated with sleep-based predictions included hypertensive heart disease, myocardial infarction, Parkinson’s disease, dementia, and several malignancies. The study suggests that subtle changes in sleep physiology may reflect early disease processes years before diagnosis.1,2
Researchers stress that SleepFM is not yet approved for clinical decision-making. Further validation is required across diverse populations and real-world healthcare settings. While future applications may include simplified sleep monitoring or wearable integration, these uses remain under investigation.1,3
Stanford Medicine. “AI Model Uses One Night of Sleep to Predict Disease Risk,” Stanford Medicine News, January 6, 2026, https://med.stanford.edu/news/all-news/2026/01/ai-model-one-night-sleep-predict-disease-risk.html.
X. Li et al., “Foundation Models for Sleep Analysis and Disease Prediction,” Nature Medicine 31, no. 1 (2026): 45–54, https://doi.org/10.1038/s41591-025-04133-4.
Longevity Technology. “AI Model Reads Disease Risk in a Single Night’s Sleep,” January 2026, https://longevity.technology/news/ai-model-reads-disease-risk-in-a-single-nights-sleep/.
MSM