SleepJEPA: Learning the latent world of sleep with at-home sleep data to estimate disease risk

Sleep disturbances lead to cardiovascular (CV), metabolic, and neurological diseases. While in-lab polysomnography (PSG) is the gold standard for measuring sleep disturbances, at-home PSG (hPSG) are more cost-effective, less resource intensive, and have been extensively used in large-scale studies. Further, hPSG devices that record EEG, EOG, and EMG are developing rapidly and collect similar data compared to in-lab PSG . However, the link between hPSG measurements and future disease risk is not well understood. We present SleepJEPA, a foundational sleep study representation model trained via a joint embedding predictive architecture that learns full night, multichannel sleep representations using hPSGs in the latent space, uncovering high-dimensional information that more precisely informs future health outcomes than standard clinical scoring. SleepJEPA was trained, validated, and tested with 422,035 hours of sleep signal data from 55,518 sleep studies. It accurately estimates 1- to 15-year cumulative risk using a discrete hazard loss function for 10 conditions, including angina (integrated area under the receiver operating characteristic curve at 15 years [iAUC ₁₅] = 0.73), CV disease death (iAUC ₁₅ = 0.83), congestive heart failure (iAUC ₁₅ = 0.85), coronary heart disease death (iAUC ₁₅ = 0.85), incident cognitive decline (iAUC ₁₀ = 0.65), diabetes (iAUC ₁₀ = 0.82), hypertension (iAUC ₁₀ = 0.79), obstructive sleep apnea (iAUC ₅ = 0.86), myocardial infarction (iAUC ₁₅ = 0.80), and stroke (iAUC ₁₅ = 0.78). We also show that these outcomes can be accurately predicted in independent cohorts, including CV disease death (iAUC ₁₀ = 0.79), coronary heart disease death (iAUC ₁₀ = 0.74), obstructive sleep apnea (iAUC ₅ = 0.77), and stroke (iAUC ₁₀ = 0.60). We report increased performance across all outcomes compared to other sleep foundation models, such as SleepFM. Through correlational analyses and explainability approaches, we illustrate features most informative for risk at different horizons. We further demonstrate SleepJEPA can effectively estimate sleep stages with high accuracy (F1 = 0.77 [95% CI: 0.76 - 0.77]), objective daytime sleepiness with modest performance (AUC = 0.64 [0.57 - 0.71]), and type 1 narcolepsy (AUC = 0.88 [0.68 - 0.97]), allowing for comprehensive labeling and disease risk assessment from hPSG signals.