Multichannel Seismocardiography Reconstructs Simulated Three-Dimensional Cardiac Mechanical Sources

Multichannel seismocardiography can capture sufficient spatial information to reconstruct the three-dimensional distribution of cardiac mechanical vibrations in simulation, according to research conducted by Ruben Kuhlmann and Urs-Vito Albrecht at Bielefeld University. Their computational investigation revealed that while spatial localization is theoretically feasible, reconstructing the precise origin of mechanical activity from a practical chest sensor configuration depends substantially on the mathematical reconstruction algorithm employed and requires tight bounds on tissue wave-speed variations.

Seismocardiography provides a non-invasive methodology for monitoring cardiac mechanical activity by recording surface chest vibrations with wearable accelerometers. To overcome this limitation, the researchers evaluated whether an array of sensors recording across multiple channels could mathematically resolve the internal three-dimensional spatial distribution of mechanical cardiac sources.

To investigate the inverse problem, the authors established a physics-informed computational framework combining three core modeling stages: an electromechanical finite-element model of the heart, an attenuating transverse-shear wave radiation model through tissue, and a three-dimensional inverse reconstruction pipeline. The study evaluated synthetic multichannel seismocardiography signals generated without heart valve transients, explicitly establishing that the analyzed data represent simulated mechanical vibrations rather than clinical seismocardiography recordings.

The investigators reconstructed the synthetic signals using four distinct processing algorithms: standard delay-and-sum beamforming, minimum variance distortionless response adaptive beamforming, the CLEAN sparse deconvolution algorithm, and sparse Bayesian learning. Across the evaluations, three of the four estimators successfully recovered both the physical position and volumetric shape of the mechanical source on at least one sensor array geometry. In contrast, the sparse Bayesian learning implementation localized the source position but failed to recover source volume, collapsing entirely onto a few isolated voxels, which the authors documented as an approach they could not successfully adapt to the task.

Localization accuracy and algorithm robustness were systematically tested against null controls across varying noise levels, wave propagation uncertainties, and sensor array geometries. With the simulated cardiac source displaced two centimeters from the array center, a theoretical fully enclosing sensor array—modeled as an upper performance bound rather than a wearable clinical setup—successfully recovered the source centroid to within 0.7 to 0.9 centimeters.

Performance diverged markedly when the algorithms were applied to a physically realizable one-sided chest array. On this single-sided geometry, sparse deconvolution achieved a centroid localization error of 1.0 centimeter, while the adaptive minimum variance distortionless response beamformer attained an error of 1.7 centimeters. Both algorithms outperformed a predefined 2.0-centimeter no-information baseline, whereas conventional delay-and-sum beamforming failed to clear this threshold and fell well outside it.

Volumetric shape reconstruction proved more robust than centroid localization, successfully surpassing the no-information baseline across both enclosing and one-sided array geometries. However, shape recovery collapsed completely when the researchers randomized inter-channel phase relationships, confirming that spatial shape estimation relies strictly on coherent phase timing across the array. Additionally, overall reconstruction remained resilient against measurement noise and moderate wave-speed mismatches, but displayed heightened sensitivity to sensor registration and geometric positioning errors.

On a one-sided array, maintaining centroid localization accuracy required unmodeled local shear wave-speed contrast to remain within roughly 20 to 30 percent for sparse deconvolution and within approximately 10 percent for the adaptive beamformer. Furthermore, the authors noted an overarching limitation of the study design: the forward simulation and inverse reconstruction pipelines shared the same underlying physics throughout, a shared dependency that structural mismatch testing bounded but could not entirely remove.

Subject of Research: Three-dimensional cardiac mechanical source localization using multichannel seismocardiography

Article Title: Physics-informed 3D source localization from multichannel seismocardiography

Article References: Kuhlmann, R., & Albrecht, U.-V. (2026). Physics-Informed 3D Source Localization from Multichannel Seismocardiography. Frontiers in Signal Processing, 6, 1937659. https://doi.org/10.3389/frsip.2026.1937659

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1937659

Keywords: seismocardiography, multichannel seismocardiography, cardiac mechanics, inverse source localization, 3D source localization, finite-element heart model, transverse-shear radiation model, delay-and-sum beamforming, minimum variance distortionless response, CLEAN deconvolution, sparse Bayesian learning, wearable accelerometers, acoustic wave-speed mismatch, sensor array geometry, centroid accuracy, phase coherence, biomedical signal processing