Physics-informed 3D source localization from multichannel seismocardiography

Seismocardiography (SCG) provides a non-invasive means of monitoring cardiac mechanical activity using wearable accelerometers. However, conventional SCG is typically recorded at a single chest location, limiting the spatial interpretation of the measured vibrations. In this study, we investigate whether multichannel SCG can be used to reconstruct the three-dimensional distribution of cardiac mechanical sources. We developed a physics-informed framework combining an electromechanical finite-element heart model, an attenuating transverse-shear radiation model, and a three-dimensional inverse reconstruction pipeline. Synthetic multichannel SCG signals without valve transients (hence not clinical SCG) were reconstructed using delay-and-sum (DAS), minimum variance distortionless response (MVDR), CLEAN, and sparse Bayesian learning (SBL). Three of the four recovered both position and shape on at least one array geometry. Our sparse Bayesian implementation localized the source but recovered no volume, collapsing onto a few voxels; thus, we report it as an approach we could not successfully adapt. Localization and robustness were evaluated against null controls and under varying noise levels, propagation uncertainties, and array geometries. With the source displaced 2 cm from the array center, a fully enclosing array, as a theoretical upper bound rather than a wearable configuration, recovered the centroid to 0.7–0.9 cm. On a more realizable one-sided array, the estimators separated: sparse deconvolution reached 1.0 cm and the adaptive beamformer 1.7 cm, both inside a 2.0 cm no-information baseline, while conventional delay-and-sum fell well outside it. Shape recovery cleared that baseline on both array geometries and collapsed when inter-channel phase was randomized. Reconstruction remained robust to measurement noise and moderate wave-speed mismatch but was more sensitive to geometric and registration errors. These findings indicate that multichannel seismocardiography, in principle, carries spatial information sufficient to localize cardiac mechanical activity in three dimensions. However, on a one-sided array, that recovery depends on the estimator, and centroid accuracy there requires an unmodeled local shear-speed contrast to stay within roughly 20– 30 % for the most tolerant estimator, and about 10 % for the adaptive one. Shape recovery is the more robust of the two. Forward and inverse models share the same physics, a limit that a structural mismatch test bounds but does not remove.