Foundation Models Reveal Untapped Health Information in Human Polysomnographic Sleep Data

Traditional sleep assessment methods rely on visual scoring of polysomnography (PSG), categorizing sleep into discrete stages (Wake, N1, N2, N3, REM) based on standardized guidelines developed by Rechtschaffen and Kales and later refined by the American Academy of Sleep Medicine (AASM). While undeniably useful, these methods are limited by subjectivity, arbitrary thresholds, and coarse temporal resolution, potentially overlooking important physiological information embedded within sleep EEG data. Recent advances in artificial intelligence, particularly Foundation Models (FM) trained through self-supervised learning (SSL), provide promising opportunities to address these limitations by autonomously uncovering rich, data-driven representations of sleep structure. In this study, we applied a SSL approach to large-scale PSG datasets, enabling the identification of latent sleep states and transitional dynamics not fully represented by traditional scoring methods. We found that these SSL-derived representations substantially enhanced predictive accuracy for critical health metrics, including demographics, clinical sleep metrics, mood, and cognitive performance. Comparisons demonstrated that this predictive advantage was not merely due to neural network complexity, as models trained directly from scratch without SSL pretraining significantly underperformed, nor was it due to having internalized simply any representation of sleep, as models pretrained to internalize traditional sleep stages could not match the SSL-pretrained model’s predictive accuracy. Taken together, these results indicate that FMs can identify structure from sleep data that encodes more than traditional sleep stage structure, uncovering previously untapped health information. These findings underscore the considerable potential of Foundation Models and SSL techniques to extend, complement, and enrich traditional sleep analyses, offering deeper insights into sleep’s complex relationship with overall health and cognitive function.