Sleep EEG foundation models reveal within-stage microstructure that improves health screening beyond traditional stages

Sleep is a rich longitudinal biosignal reflecting integrated brain and systemic physiology, yet polysomnography reduces to a lossy clinical interface of coarse, human-defined stages. We asked whether self-supervised foundation models learn sleep EEG structure beyond traditional staging and encode enriched health information. Using 11,261 overnight recordings, we trained transformers on unlabeled sleep data and probed representations across diagnostic, demographic and functional outcomes. Self-supervised models outperformed matched from-scratch models and exceeded five-stage–supervised pretraining for several endpoints (BMI, age, mood and cognition), while remaining comparable for apnea severity and sex. In nested controls, EEG-derived self-supervised model scores retained incremental predictive value beyond demographic covariates and stage-derived sleep-report summaries, indicating that gains are not explained by model architecture or coarse sleep architecture alone. Embedding analyses show that the models recover the stage scaffold without labels yet resolve within-stage microstructure—especially within N2—that improves health prediction and supports scalable EEG-only digital biomarkers.