Deep Learning-Based Heartbeat Detection from 3D Seismocardiography for Robust Heart Rate Monitoring

Accurate monitoring of heart rate (HR) is critical for assessing cardiac functions in a wide range of health and wellness applications. Seismocardiography (SCG), which captures subtle chest vibrations using wearable accelerometers, provides a non-invasive and cost-effective approach for resting and nocturnal HR monitoring. This study presents a deep learning-based approach for accurate heartbeat detection and HR estimation from three-dimensional SCG signals. The model was trained on a large-scale dataset of resting SCG signals collected from 6600 subjects and evaluated on an independent cohort of 947 individuals. For short-term (≤5 min) resting SCG recordings, the model achieved robust performance in heartbeat detection (PPV: 0.979, sensitivity: 0.916, F1-score: 0.946). HR estimation showed high accuracy, with a mean absolute error (MAE) of 0.27 bpm, root mean square error (RMSE) of 1.02 bpm, and correlation of 0.996 with the reference HR. To assess real-world applicability, the model was further evaluated on 28 nocturnal recordings acquired using Apple Watch accelerometer, yielding an MAE of 1.10 bpm, an RMSE of 1.88 bpm, and a correlation of 0.982. The proposed SCG-based deep learning model demonstrates robust and highly accurate HR monitoring in both resting and nocturnal conditions, highlighting its potential for integration with consumer-grade wearable devices in a server-based analysis pipeline.