Document Type : Original Article
Authors
1
Department of Mechacnical Engineering, Isfahan University of Technology, Isfahan, Iran
2
Department of Mechanical Engineering, Isfahan University of Technology, Isfahan, Iran
3
Department of Electrical and Computer Engineering, Isfahan University of Technology, Isfahan, Iran
4
Department of Bioengineering, Lehigh University, Bethlehem, PA 18015, USA
Abstract
Cardiovascular diseases remain the leading cause of mortality worldwide, highlighting the critical need for continuous and non-invasive monitoring of cardiac function to enable early detection and effective management. Ballistocardiography (BCG), which captures the mechanical forces associated with cardiac activity, holds great promise for unobtrusive heart monitoring in daily-life settings without requiring direct electrode contact. However, the inherent complexity and high susceptibility to noise in BCG signals make the accurate extraction of key cardiac parameters—particularly inter-beat intervals (IBIs)—a challenging task. This study presents a comprehensive evaluation of five distinct signal processing and deep learning approaches for IBI estimation from BCG signals, validated against synchronized electrocardiogram (ECG) recordings. In contrast to previous works, we employ a publicly available dataset distinct from those commonly used, enabling a broader assessment of method generalizability—particularly for the CLIE algorithm. The evaluated methods include: Continuous Local Interval Estimator (CLIE), CLIE with adaptive windowing, Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Bidirectional Long Short-Term Memory (BiLSTM) network. For the deep learning methods (MLP and CNN), we propose novel network architectures specifically tailored to the characteristics of BCG signals, leading to improved performance compared to conventional designs. Furthermore, our BiLSTM-based method not only incorporates testing on a dataset different from that of previous reference studies but also focuses on the accurate prediction of R-peak locations in the BCG signal, from which IBIs are subsequently derived. Evaluation based on Mean Absolute Error (MAE), 95th percentile error, and correlation coefficient shows that the CLIE method achieved the best overall IBI estimation accuracy, with an MAE of 28.7 milliseconds and the highest correlation coefficient (0.77). The BiLSTM method, while having a slightly higher MAE (40.1 milliseconds), demonstrated superior robustness to outliers by achieving the lowest 95th percentile error (9.5%). The MLP and CNN methods showed moderate performance, and the adaptive windowing variant of CLIE performed the worst. These findings demonstrate that accurate IBI extraction from BCG signals is feasible, and that both the CLIE and BiLSTM approaches are promising candidates for implementation in intelligent, home-based cardiac monitoring systems—offering, respectively, high accuracy and strong resilience to large errors.
Keywords
Main Subjects
)