Lead. Getting a reliable sleep assessment usually means spending a night in a lab wired to dozens of electrodes. Polysomnography (PSG), the gold standard for sleep staging and apnea diagnosis, is cumbersome, expensive, and ill-suited for routine screening or at-home use. Now a large international team has published results for BCGNet, a deep learning model that performs clinical-grade sleep staging and apnea detection using nothing more than a sensor mat slipped under the pillow. The study, appearing July 3 in NPJ Digital Medicine, reports that the model matches or outperforms many existing approaches while requiring no direct contact with the patient’s skin.
What they found. BCGNet is a two-stage transfer learning architecture developed by researchers from 14 institutions spanning China, Australia, and the United States, including Tsinghua University, Harvard Medical School, and the University of California, San Francisco. The team first pre-trained the model on 580,865 hours of polysomnography data, then fine-tuned it on 15,081 hours of ballistocardiography (BCG) recordings captured by the under-pillow device. The combined training set of roughly 596,000 hours is among the largest ever assembled for sleep staging research.
On four-class sleep staging (wake, light sleep, deep sleep, REM), the model achieved F1 scores between 0.710 and 0.817 across multiple validation cohorts. For estimating the apnea-hypopnea index at a 3% desaturation threshold (AHI3%), Pearson’s r exceeded 0.95, indicating near-perfect correlation with reference PSG measurements. Sleep continuity and architecture metrics, including total sleep time, sleep efficiency, and time spent in each stage, showed intraclass correlation coefficients and Pearson’s r values generally above 0.8.
The model generalized well across diverse external datasets that differed in patient demographics, recording equipment, and clinical settings. The authors also demonstrated strong performance on short daytime nap recordings, suggesting potential utility beyond overnight monitoring.
Why it matters. Obstructive sleep apnea affects an estimated 936 million adults worldwide, the vast majority of whom remain undiagnosed. Current home sleep tests, while more convenient than in-lab PSG, still require patients to wear sensors on the face, chest, or fingers. Many patients find these devices uncomfortable, and compliance with home testing remains inconsistent.
A truly contactless approach eliminates those barriers. The under-pillow BCG mat requires no setup by the patient, no cleaning between uses, and no wearing of any device. It can be deployed in homes, long-term care facilities, and hospital wards without disrupting the patient’s sleep environment. If the performance reported in this study holds up in prospective real-world deployments, the device could dramatically expand access to objective sleep assessment, particularly in settings where PSG is unavailable or impractical.
The model’s ability to estimate AHI3% with Pearson’s r above 0.95 is especially noteworthy. The apnea-hypopnea index is the primary metric used to diagnose and grade sleep apnea severity. A contactless sensor that can produce a clinically actionable AHI value could serve as a scalable screening tool, potentially identifying the millions of undiagnosed patients who would benefit from further evaluation and treatment.
Limits. The study was retrospective, and the authors acknowledge that prospective validation in real-world home environments is an important next step. While the external validation datasets were diverse, they were still curated research collections that may not be fully representative of the general population. The device captures only BCG signals; it cannot measure airflow, oxygen saturation, or EEG directly, meaning the model must infer these parameters indirectly. Performance in patients with complex comorbidities, severe cardiac arrhythmias, or unusual sleep architectures was not extensively evaluated. The patent is held by Five Seasons Medical, a Beijing-based company, and several authors are employees of the device manufacturer, a potential conflict of interest that should be weighed when evaluating the reported performance figures. The study was supported by the National Natural Science Foundation of China and the Ministry of Science and Technology of China.
Bottom line. BCGNet demonstrates that a deep learning model trained on an enormous bank of PSG data can transfer its knowledge effectively to a contactless ballistocardiography signal, producing sleep staging and apnea estimates that approach the accuracy of attended polysomnography. The work represents the latest and strongest evidence that wearable and contactless sensors, paired with sophisticated neural networks, may soon make reliable sleep assessment available to anyone with a bed.
Source. Chen S, Chen X, et al. BCGNet: an AI model trained on 600 K hours of sleep data for a novel under-pillow contactless monitoring device. NPJ Digit Med. 2026 Jul 3. doi: 10.1038/s41746-026-02885-y. PMID: 42399358.