
Sleep foundation model uncovers five risk groups beyond apnea severity
A transformer-based sleep foundation model trained on 10,000 high‑resolution PSG studies linked to EMRs learns rich sleep-physiology embeddings that identify five prognostic risk groups with markedly different mortality, cardiovascular, and neurologic outcomes, outperforming conventional AHI-based severity. External validation in the Sleep Heart Health Study confirms generalizability, suggesting embedding-based risk stratification could enable precision sleep medicine and targeted monitoring, while acknowledging limitations and the need for prospective trials.



