A Zero-Burden Sleep Foundation Model Built on Cardiorespiratory Signals from 800,000+ Hours of Multi-Ethnic Sleep Recordings
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
BACKGROUND
Sleep disorders pose a major global health burden and are associated with a wide range of adverse health outcomes. Polysomnography (PSG) is the gold standard for sleep assessment, but it is impractical for home-based or long-term monitoring. We investigated whether zero-burden cardiorespiratory signals, when harnessed through foundation model approaches, can enable accurate sleep assessment at scale and capture broader dimensions of multi-organ health.
METHODS
We present SleepFounder, a foundation model for zero-burden sleep monitoring built upon cardiorespiratory signals. SleepFounder was developed on the largest curated multi-ethnic sleep dataset to date, comprising over 800,000 hours of recordings from 35 cohorts across the United States and China. We evaluated SleepFounder across downstream tasks ranging from conventional sleep analysis to emerging applications, including demographic profiling and multi-organ disease detection and prediction. We further conducted a real-world study using multi-center ballistocardiography (BCG) data collected with a custom-developed sleep mat system for external validation.
RESULTS
SleepFounder achieved strong performance across diverse downstream tasks and consistently outperformed baseline models, obtaining the best results in 14 out of 17 dataset-task pairs. For conventional sleep analysis and demographic profiling, averaged across external datasets, it achieved a Cohen’s Kappa of 0.671 (0.668-0.673) for five-class sleep staging, an area under the receiver operating characteristic curve (AUROC) of 0.917 (0.912-0.922) for moderate-to-severe obstructive sleep apnea detection, a mean absolute error of 6.727 (6.684-6.771) years for age prediction, and an AUROC of 0.865 (0.860-0.870) for sex classification. In multi-organ disease detection, representative AUROCs reached 0.943 (0.917-0.966) for Parkinson’s disease, 0.886 (0.841-0.928) for gastroesophageal reflux disease, and 0.881 (0.831-0.922) for heart failure. Additional conditions, including high cholesterol, coronary heart disease (CHD), bipolar disorder, and chronic pain, achieved AUROCs ranging from 0.811 to 0.830 in the held-out test set, with results further validated across five external cohorts. For future disease prediction, concordance indices reached 0.838 (0.797-0.873) for CHD death and 0.837 (0.806-0.865) for cardiovascular disease death, with corresponding metrics of 0.734-0.781 for congestive heart failure, stroke, and angina. In the real-world BCG study, SleepFounder maintained 94% of its performance on average relative to prior external validations conducted on PSG-based datasets.
CONCLUSIONS
SleepFounder establishes a foundation model that learns from cardiorespiratory signals to enable accurate, scalable, and zero-burden sleep assessment. By linking sleep physiology with multi-organ health, it bridges clinical and home settings and demonstrates that signals traditionally used for sleep monitoring can serve as powerful biomarkers of systemic function and disease risk. These findings highlight a new paradigm for zero-burden sleep and health monitoring in real-world settings.
