Back

BCGNet: An AI Model Trained on 600K Hours of Sleep Data for BCG-based Contactless Monitoring

Chen, S.; Chen, X.; Huang, W.; Lei, F.; Shan, C.; Jin, Z.; Shi, Y.; Wang, Y.; Zhao, R.; Xu, X.; Lv, D.; Li, Y.; Westover, B.; Hong, S.; Zhang, C.; Yin, S.; Liu, C.; Yi, H.; Tang, X.; Leng, Y.

2025-08-26 health informatics
10.1101/2025.08.23.25334218 medRxiv
Show abstract

Sleep profoundly impacts health, yet current gold-standard Polysomnogram (PSG) is constrained by cost, discomfort, and limited scalability for longitudinal monitoring. Ballistocardiogram (BCG) offers a non-invasive and user-friendly alternative but often lacks the precision needed for reliable real-world applications. To address this gap, we propose BCGNet, a two-stage transfer learning model that is first pre-trained on 580,865 hours of PSG and then fine-tuned and validated on 15,081 hours of BCG (total 595,946 hours of recordings). Across multiple validation cohorts, BCGNet achieves strong performance in 4-class sleep staging (F1: 0.710 - 0.817), Apnea-Hypopnea Index (AHI3%) estimation (Pearsons r > 0.95), and robust quantification of sleep continuity and architecture (ICC and Pearsons r generally > 0.8). Notably, BCGNet maintains strong performance even on short daytime naps and demonstrates excellent generalizability across diverse external datasets. Deployed as a portable, contactless sleep tracking mat, BCGNet represents a major step towards scalable, user-friendly solutions for longitudinal home sleep monitoring, with important implications for population screening and personalized sleep medicine.

Matching journals

The top 6 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.