SleepFM, a multimodal foundation model developed with data from over 585,000 hours of polysomnography, predicts 130 medical conditions, including dementia and cardiovascular diseases, from a single sleep night. Stanford University researchers utilized a novel contrastive learning approach for training, achieving significant accuracy validated across diverse cohorts. With performance measured via concordance indices for various conditions, SleepFM's predictive capabilities exemplify advancements in noninvasive health monitoring through sleep technologies.
1. SleepFM predicts 130 medical conditions from a single night of sleep data. 2. Achieved C-Index scores of 0.84 for all-cause mortality and 0.91 for Alzheimer disease. 3. Trained on over 585,000 hours of polysomnography recordings. 4. Utilizes a contrastive learning approach and integrates multiple signal modalities. 5. Demonstrated strong performance across various cohorts and generalization with external validation.
Listen Tab content
