Abstract
Clinical diagnosis of sleep disorders, which are recognized contributors to morbidity and mortality, often relies on polysomnography (PSG) data. However, the vast physiologic data collected during PSG is underutilized, presenting a key opportunity to enhance characterization of sleep dysfunction and predict clinical outcomes. We introduce a sleep foundation model that uniquely integrates PSG time-series signals and electronic medical record data. Using a diverse dataset (n=10,000; mean observation period 14.5±7.1 years), our transformer-based model generates data-driven representations of latent physiological patterns. When clustered, we identified subpopulations with differential health trajectories. The highest risk-group exhibited strong correlations with all-cause mortality (unadjusted hazard ratio [HR] 4.83, 95% confidence interval [CI] 3.60-6.50, p<0.001) as well as cardiovascular outcomes and neurological outcomes, even after accounting for traditional measures. External validation in a National Sleep Research Resource cohort confirmed findings. We created a novel, clinically applicable framework leveraging information-dense PSG data to inform risk stratification and predict health outcomes beyond traditional methods.