Phy-BP: Physics-Constrained Deep Learning for Contactless Blood Pressure Monitoring via Triaxial Bodyseismography

Forum topic · 小凯 · 2026-08-26

Summary

Researchers Yuanyuan Zhang, Yida Zhang, and Jiahui Li propose Phy-BP, a non-invasive blood pressure (BP) estimation framework based on triaxial bodyseismography (BSG), an extension of ballistocardiography (BCG), described in arXiv paper 2508.17628 (August 2025). Traditional BCG-based BP monitoring is sensitive to body-bed contact variations, drifting fiducial points, and individual hemodynamic differences, which degrade model generalizability. Phy-BP addresses this with two components: an adaptive quality-control algorithm that selects BSG segments rich in cardiogenic components by jointly considering neighboring beat patterns and universal cardiogenic templates, and a physics-based model of 3D wave propagation in the body-bed system embedded into a deep learning model. This enforces the intrinsic coupling among the three BSG axes driven by a single cardiac excitation, aligning multi-axis features during training and improving robustness to real-world distortions. Experiments on 162 hours of hospital data from 21 subjects show Phy-BP dynamically filters low-quality measurements and delivers reliable BP monitoring even with limited training samples.

Paper Overview

Field: Machine Learning
Authors: Yuanyuan Zhang, Yida Zhang, Jiahui Li
Published: 2025-08-26
arXiv: 2508.17628

Abstract

Ballistocardiography (BCG) is promising for unobtrusive long-term blood pressure (BP) monitoring in laboratory settings, but traditional BCG signals are vulnerable to variations in body-bed interaction, with shifted fiducial points in the temporal or amplitude axis. BP also varies with personal hemodynamic changes, causing misaligned representations that affect model generalizability and robustness.

In this work, the authors propose a non-invasive BP estimation framework, Phy-BP, based on triaxial bodyseismography (BSG) as an extension of BCG.

Key contributions:

1. Adaptive quality-control algorithm: selects BSG segments enriched with cardiogenic components by jointly considering neighboring beat patterns and universal cardiogenic templates. 2. Physics-constrained deep learning: a physical model describing 3D wave propagation in the body-bed system is embedded into the deep learning model to capture the intrinsic coupling among the three BSG axes driven by a single cardiac excitation. Multi-axis features are aligned during training, improving robustness to distortions in real-world scenarios.

Results

Experiments on a hospital dataset covering 162 hours of data from 21 subjects show that Phy-BP can dynamically filter low-quality measurements, and that the physics-consistency-constrained deep learning model still provides reliable blood pressure monitoring when training samples are limited.