Accurate blood pressure waveform decomposition and feature extraction using Ballistocardiogram mechanism

Background Ballistocardiogram (BCG), a vibration signal generated by cardiac ejection, demonstrates promising potential for blood pressure (BP) estimation. Existing data–driven models estimate BP from BCG as black–box predictors and therefore generalize poorly across populations while offering little physiological insight. Methods Building on recent findings that BCG is a superposition of intra-aortic blood pressure waveforms (BP waveforms), we propose a B-spline based BP waveform decomposition (BPWD) to extract BP waveform features for BP estimation. By leveraging particle swarm optimization to refine B-spline control points, BPWD generates physiologically plausible BP waveforms whose superposition (generated-BCG) aligns morphologically with measured-BCG. However, between generated-BCG and measured-BCG, there is a scaling discrepancy (Sd) that affects BPWD. In our study, two BPWDs have been proposed: BPWD_avgSd uses the average Sd value of all measurements with known BP values; BPWD_optSd optimizes Sd in terms of both the pulse pressure difference calculated using a prior formula based on BCG waveform features and height and weight, and root-mean-square error between the generated BCG (scaled by Sd ) and the measured BCG. Results Based on 69 measurements from 23 participants, linear-regression models built on BP waveform features extracted by BPWDs outperformed the model built on BCG waveform features. The mean absolute error (mmHg) for diastolic/systolic/pulse pressure were BPWD_optSd – 5.9/8.9/7.4; BPWD_avgSd – 6.0/9.3/7.6; and BCG waveform features – 6.5/11.5/8.7, respectively. Conclusions BPWD effectively decomposes the BCG into BP waveforms whose features better support linear-regression BP estimation than BCG waveform features. The superior performance of BPWD_OptSd over BPWD_avgSd demonstrates that optimizing Sd further enhances BPWD performance.