A multimodal sleep foundation model for disease prediction.

Area of Science:

• Biomedical Informatics
• Sleep Medicine
• Artificial Intelligence

Background:

• Sleep is crucial for health, but its link to disease is complex and polysomnography (PSG) data is underutilized.
• Existing PSG analysis methods face challenges in standardization, generalizability, and multimodal data integration.

Purpose of the Study:

• To develop a multimodal sleep foundation model (SleepFM) to overcome PSG analysis limitations.
• To enable accurate prediction of future disease risk using sleep data.

Main Methods:

• Developed SleepFM using a novel contrastive learning approach for multimodal PSG data.
• Trained on over 585,000 hours of PSG recordings from ~65,000 participants.
• Utilized latent sleep representations for disease prediction and transfer learning.

Main Results:

• SleepFM accurately predicts 130 conditions (C-Index ≥ 0.75), including mortality, dementia, myocardial infarction, and chronic kidney disease.
• Demonstrated strong transfer learning on an independent dataset.
• Achieved competitive performance with specialized models for sleep staging and sleep apnea classification.

Conclusions:

• Foundation models can effectively learn from multimodal sleep recordings.
• SleepFM enables scalable, label-efficient sleep analysis and disease prediction.
• This approach enhances our understanding of sleep's impact on physical and mental health.