Learning Transferable Prognostic Sleep Representations from Polysomnography for Disease and Mortality Prediction
Examiner: Mats Granath
Supervisor: Elias Stenhede (Akershus University Hospital)
Abstract
Sleep is a fundamental process deeply connected to overall health, with evidence pointing to a bidirectional relationship between impaired sleep and adverse outcomes spanning metabolic, neurocognitive, and cardiovascular disease. Polysomnography (PSG) is the gold standard diagnostic test for sleep disorders, capturing rich multimodal physiological information across the night. While deep learning on PSG has advanced rapidly, most work remains focused on sleep disorder detection, and direct prediction of long-term clinical outcomes remains comparatively underexplored. Polysomnography signals carry prognostic information that could support early risk stratification and timely intervention for systemic disease given appropriate processing. This thesis develops a spectrogram-based self-supervised foundation model that learns to separate fine-grained modality-intrinsic from cross-modal features in PSG signals. For downstream prediction, the learned representations are processed by a bidirectional Mamba model. Per-task risk scores are produced under a Cox time-to-event objective. The model was developed on 12,376 overnight recordings from 9,110 patients obtained from the Stanford Sleep Bench dataset.
The model was evaluated on a held-out test set of 2,625 recordings from 1,907 patients, predicting 13 clinical outcomes. These include cardiovascular, metabolic, and neurodegenerative disease, as well as all-cause mortality. Disease-level AUROC ranges from 0.68 to 0.81. All-cause mortality is predicted with an AUROC of 0.83. Calibration analysis shows that risk scores are consistently ranked correctly, though slightly overestimated in absolute terms. The results show that PSG contains predictive signal for long-term outcomes well beyond the sleep scoring paradigm. Explicit separation of modality-intrinsic from cross-modal structure is shown to be a viable design choice for foundation models in this domain.