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

Sleep physiology provides rich longitudinal biosignals reflecting integrated brain and systemic physiology, yet polysomnography is commonly compressed into 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. Compared with architecture-matched transformers trained from random initialization on each downstream task, SSL pretraining improved performance across several outcomes. Compared with five-stage-supervised pretraining, EEG-only advantages were clearest for BMI and age, while differences for AHI, sex, and functional outcomes were smaller, nominal, or not reliable. In nested controls, EEG-derived self-supervised model scores retained incremental value beyond covariates, stage summaries, spectral summaries, and a matched five-stage representation. Embedding analyses show that models recover the stage scaffold without labels while preserving a higher-resolution, stage-anchored structure that carries task-specific health information beyond the five-stage interface.