A Low-Burden Sleep Foundation Model Built on Respiratory and Heartbeat Signals from 780,000+ Hours of Multi-Ethnic Sleep Recordings
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
BACKGROUND
Sleep disorders represent a global health challenge, with obstructive sleep apnea alone affecting nearly 85 million Americans, about 80% of whom remain undiagnosed. Polysomnography (PSG), the gold standard for sleep analysis, is unsuitable for routine or home monitoring. Heartbeat and respiratory signals are core vital signs, broadly obtainable across settings. We investigate whether these signals, when leveraged with foundation model techniques, can enable accurate sleep analysis and provide insights into multi-organ health.
METHODS
We introduce SleepFounder, a foundation model leveraging heartbeat and respiratory signals for low-burden sleep monitoring. SleepFounder was developed from more than 780,000 hours of multi-ethnic sleep recordings across 28 cohorts in the United States and China, enabled by the broad accessibility of these signals. We evaluated SleepFounder on downstream tasks including sleep analysis, demographic profiling, and disease diagnosis across multiple physiological systems, using both curated PSG datasets and real-world, multi-center ballistocardiography (BCG) data collected with a custom-developed sleep mat.
RESULTS
SleepFounder achieved strong performance across all evaluated tasks and typically matched baseline models while using only 20% of the training data. Averaged across external PSG and BCG datasets, it outperformed baselines with Cohen’s κ of 0.65 for five-class sleep staging, intraclass correlation coefficient of 0.85 for apnea-hypopnea index estimation, mean absolute error of 6.90 years for age prediction, and area under the receiver operating characteristic curve (AUROC) of 0.85 for sex classification. In addition, SleepFounder demonstrated broad diagnostic potential: in the held-out test set, it achieved AUROCs of 0.89 for gastroesophageal reflux disease, 0.88 for heart failure, 0.83 for high cholesterol, 0.81 for coronary heart disease, chronic pain, and bipolar disorder, and 0.77 for atrial fibrillation and flutter. These findings were further validated in temporally and geographically external cohorts across five datasets.
CONCLUSIONS
Cardiorespiratory signals encode rich, system-wide physiological information, enabling both conventional sleep analysis and broader health assessment. Building on this insight, SleepFounder demonstrates that low-burden monitoring based solely on respiration and heartbeat can deliver accurate sleep care, highlighting its potential for scalable real-world deployment.
