Background: Nocturnal heart rate variability (HRV) is a well-established physiological marker of autonomic function and widely used in sleep analysis, stress assessment, and recovery monitoring. Seismocardiography (SCG), a noninvasive sensing modality capturing subtle chest wall vibrations induced by cardiac mechanical activity, has emerged as a promising alternative for HRV estimation. However, most existing SCG-based methods rely on conventional signal processing approaches, which may be susceptible to noise and artifacts in long-term, real-world recordings. Deep learning methods, with their ability to model complex and noisy signals, remain insufficiently explored in nocturnal HRV monitoring. This study evaluates a previously developed deep neural network (DNN)-based method for HRV estimation from nocturnal SCG recordings.
Methods: A U-Net-based architecture, trained on short resting SCG recordings from over 5,500 individuals, was employed to estimate HRV from 3D SCG signals. The model performance was assessed using the publicly available NightbeatDB dataset, comprising 38 overnight sleep recordings (~291 hours) collected in home environments. Heart rate (HR) and HRV metrics were computed from 10-minute segments and compared with time-synchronized ECG-derived references.
Results: A total of 90.47% of the data were deemed analyzable using the proposed method. The model demonstrated excellent correlation with the reference ECG for HR estimation (r = 0.982), with a mean bias of -0.3 bpm and 95% limits of agreement (LoA; mean± 1.96×SD) from -1.7 to 1.0 bpm. The root mean square of successive differences (RMSSD) also showed high correlation (r = 0.929), with a mean bias of 0.6 ms and LoA ranging from -10.0 to 11.1 ms.
Conclusion: These findings support the feasibility of deep learning-based SCG analysis for accurate nocturnal HRV estimation in real-world sleep monitoring, highlighting its potential as a scalable and unobtrusive tool for continuous physiological monitoring.