SleepFM: New Multimodal Foundation Model Predicts 130 Diseases from a Single Night of Sleep
Key Takeaways
- ▸SleepFM predicts 130 diseases from a single night of sleep with high accuracy, including major conditions like dementia, heart disease, and all-cause mortality
- ▸Trained on 585,000+ hours of PSG data from 65,000 participants, the model demonstrates strong generalization and transfer learning capabilities
- ▸The contrastive learning approach accommodates multiple PSG configurations, addressing critical standardization and generalizability challenges in sleep analysis
Summary
Researchers have developed SleepFM, a groundbreaking multimodal foundation model trained on over 585,000 hours of polysomnography (PSG) recordings from approximately 65,000 participants. Using contrastive learning, the model creates latent sleep representations that capture the physiological and temporal structure of sleep, enabling accurate prediction of future disease risk. From just one night of sleep data, SleepFM can predict 130 different conditions with high accuracy, including life-threatening diseases like all-cause mortality (C-Index 0.84), dementia (0.85), myocardial infarction (0.81), heart failure (0.80), and stroke (0.78).
The model addresses long-standing challenges in sleep analysis by accommodating multiple PSG configurations and demonstrating strong generalizability across different cohorts. Beyond disease prediction, SleepFM performs competitively with specialized models on traditional sleep analysis tasks, achieving mean F1 scores of 0.70–0.78 for sleep staging and high accuracy for sleep apnea classification. This work demonstrates that foundation models can effectively learn the "language of sleep" from multimodal recordings, potentially enabling scalable, label-efficient clinical analysis at a population level.
- The foundation model performs competitively with specialized sleep-staging models while enabling scalable, label-efficient clinical applications
Editorial Opinion
SleepFM represents a significant advancement in applying foundation models to healthcare, leveraging the untapped diagnostic potential in routine sleep recordings. The ability to predict 130 diseases from a single night of polysomnography could fundamentally transform sleep medicine and preventive care, moving beyond traditional sleep disorders to enable early detection of serious systemic conditions. This work exemplifies how large-scale, multimodal biomedical data combined with modern deep learning can unlock actionable clinical insights that were previously inaccessible.
