Improving ballistocardiogram-based continuous heart rate variability monitoring: A self-supervised learning approach

Abstract

Heart rate variability (HRV) is a reliable measure of an individual’s physical and mental fitness. Monitoring HRV over time provides crucial indicators of various health issues and helps in the early implementation of preventive measures. While the ballistocardiogram (BCG) serves as an alternative to the electrocardiogram (ECG) for convenient at-home HRV monitoring, it is inevitable to face the presence of undesirable artifacts and irregular waveforms in BCG signals that can hinder accurate monitoring when appropriate interventions are unavailable. To address this challenge, we propose a self-supervised learning approach consisting of three stages: signal separation, heartbeat pattern detection, and peak identification. We introduce convolutional kernels of different sizes to construct adaptive soft filters that effectively extract the heartbeat waveform from the original signal. To discover the heartbeat patterns, we utilize the k-means algorithm for clustering the encoded peaks obtained from the masked autoencoder and then compute Kullback–Leibler divergences to determine the optimal number of clusters. Subsequently, a lightweight multi-feature cross-attention classifier is employed to locate the heartbeats. Although the classifier is trained using pseudo-labels generated in the previous stage, the inherent similarity of heartbeats can enhance its performance. Our approach accommodates diverse BCG signal modes and does not require an annotated dataset for model training. Through evaluations conducted on 24 healthy subjects using synchronized ECG references, our prototype system demonstrated an accuracy of approximately 99% in heartbeat identification and an interbeat interval RMSE of under 15ms. These results suggest the system’s practicality and suitability for real-world HRV monitoring.