Sleep polysomnography predicts 130 health conditions, including CVD (www.empirical.health)

A groundbreaking study from Stanford introduced SleepFM, a multimodal foundation model that predicts 130 health conditions, including cardiovascular diseases, from just one night of polysomnography (PSG) recordings from 65,000 patients. Achieving significant accuracy levels—such as 84% for all-cause mortality and 85% for dementia—SleepFM outperforms traditional risk assessment tools, highlighting its potential in clinical settings. The model employs a unique pretraining method known as leave-one-out contrastive learning, which enhances its resilience to data variability and missing channels, thereby allowing it to integrate multiple physiological signals efficiently. The implications for the AI/ML community are substantial. SleepFM's architecture utilizes both convolutional neural networks and transformers to analyze diverse signal modalities (brain activity, EKG, respiratory, and EMG) concurrently, capitalizing on the unique data structure of PSG that reflects a controlled physiological state. This approach marks a significant advancement in predictive health technologies, suggesting that a single night of sleep can yield predictive insights comparable to years of health history. As AI continues to evolve in healthcare, solutions like SleepFM could reshape preventive and diagnostic strategies, emphasizing the importance of integrating multimodal data for enhanced predictive modelling.