0359 Sleep-JEPA: Joint Embedding Representations of Sleep Study Data Accurately Estimate Sleep Features and Long-Term Health Outcomes

Abstract Introduction Sleep disorders and deprivation disrupt daily life and are linked to 7 of the 15 leading causes of death in the United States. Early detection, clinical management, and lifestyle changes are key for reducing risk. Sleep studies vast stream of sensor data along with cohort datasets' long-term health outcome labels could be linked with advanced machine learning methods. Specifically, the transformer model, combined with self-supervised training techniques are well-suited to learn multichannel signal representations and estimate risk for various outcomes including CV disease, stroke, and myocardial infarction. Methods We utilized the Human Sleep Project (n=24,986), Mt. Sinai Polysomnography (n=8,547), MrOS (n=3,925), SHHS (n= 8,444), Wisconsin Sleep Cohort (n=2,544), APPLES (n=1089), Mignot Nature Communications (n=1355), and MESA (n=2,055) sleep datasets to train a foundational transformer using a joint embedding predictive architecture (JEPA). The model learned representations of full-night (6-12 hour) sleep studies across seven channels: a single EEG (C4-M1 or C3-M2), left EOG, chin EMG, lead II ECG, SpO2, and abdomen and thoracic respiratory rates. Following self-supervised training, these representations were input into a non-linear probing classification model with a discrete hazard loss function to predict sleep stages, objective daytime sleepiness based on MSLT, incident hypertension, stroke, myocardial infarction, and CV mortality using the SHHS and MrOS datasets. Results Sleep-JEPA’s frozen representations estimated an average area under the receiver operating characteristics curve of 0.96 across 5-stage sleep (average precision [AP] of 0.79), 0.72 on objective daytime sleepiness (MSLT < 8 min, AP of 0.35, ~2.5x prevalence), 0.77 for CV mortality at 15 years (Concordance index [C-index]: 0.78), 0.74 for stroke at 15 years (C-index: 0.70), 0.83 for myocardial infarction at 15 years (C-index: 0.80), and 0.82 for incident hypertension at 10 years (C-index: 0.70) indicating good discriminative performance. Conclusion A foundational transformer trained via the JEPA procedure can learn relevant representations of full-night, multichannel PSG data and these representations can estimate sleep features and long-term health outcomes. Future work will include additional outcomes including OSA, diabetes, and cognitive decline, baseline comparisons, and explore explainable attributions. Support (if any) NIH R01HL175992