BCG data imputation via multimodal feature alignment and semantic sequence prediction

The Ballistocardiogram (BCG) is a promising measure for long-term cardiovascular health monitoring. However, its unobtrusive data acquisition renders BCG vulnerable to interference and disruptions caused by involuntary body movements, resulting in fragmented signal records. To mitigate this challenge and obtain continuous BCG signals required by downstream applications, we propose an end-to-end framework for reconstructing missing data samples affected by abrupt body movements. Our approach leverages the multimodality of cardiovascular activities and the semantic interdependence of sequential heartbeats. The proposed data imputation framework comprises three stages: conversion from BCG to electrocardiogram (ECG), prediction of masked ECG heartbeats, and conversion from ECG back to BCG. By ensuring the alignment of latent features within the mutual conversion modules, our method can effectively capture the compact representations of heartbeats. Given its availability of public datasets and inherent signal stability, ECG is a suitable alternative to directly restoring corrupted BCG data. Evaluated on the Kansas public BCG dataset, our proposed method demonstrated superior performance in relative root mean square errors (rRMSE) and correlation metrics compared with existing techniques, thereby providing a practical solution for enhancing continuity in BCG signal sampling.