A multimodal sleep foundation model for disease prediction (www.nature.com)

Researchers have announced the development of SleepFM, a groundbreaking multimodal sleep foundation model designed for disease prediction. This model leverages a vast dataset of over 585,000 hours of polysomnography (PSG) recordings from approximately 65,000 participants, using a novel contrastive learning approach. SleepFM demonstrates its capability to predict 130 different medical conditions from a single night of sleep, achieving a C-Index of over 0.75 for various severe health issues, including dementia, heart failure, and all-cause mortality. This advancement highlights the potential of integrating sophisticated machine learning models with multimodal physiological data, significantly enhancing our understanding of sleep’s role in health management. The significance of SleepFM lies in its robust ability to generalize across diverse PSG configurations and cohorts, overcoming limitations of previous approaches that focused on isolated outcomes with smaller datasets. Its channel-agnostic design allows it to process varying recording modalities seamlessly, ensuring scalability and accuracy in predicting disease risk. By harnessing self-supervised learning techniques, SleepFM not only effectively analyzes complex sleep data but also assists in tackling broader health conditions beyond sleep disorders, thus underscoring the transformative implications for AI and machine learning in biomedical applications. The model’s strong performance on standard sleep analysis tasks further validates its utility as a comprehensive tool in clinical settings.