A FastKAN-Based Spatiotemporal Inverse Network for Robust Epicardial Potential Reconstruction in Electrocardiographic Imaging

Date

2026-8-5

Author

Tüysüz, Yunus Emre

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

56
views

0
downloads

Cite This

Electrocardiographic imaging (ECGI) reconstructs epicardial potentials from body-surface measurements by solving an ill-posed inverse problem, where measurement noise, missing leads, and modeling errors can cause large errors. This thesis presents a learning-based ECGI model built around a forward-matrix-free Kolmogorov--Arnold (FastKAN) inverse mapping that directly estimates epicardial potentials from body-surface potentials. A FastKAN inverse block produces an initial estimate, which is refined by two lightweight residual modules that promote spatial coherence across the epicardial mesh and temporal continuity along reconstructed electrograms. Training further includes a differentiable activation-time consistency loss to support activation-time reconstruction and pacing-site localization. The model is evaluated on two torso-tank datasets: a canine dataset with simulated body-surface potentials and an available forward operator, and a multi-animal pig dataset with measured body-surface potential maps paced from multiple epicardial sites and no consistent forward matrix. Performance is assessed using temporal correlation, spatial correlation, activation-time correlation, and localization error under clean, additive-noise, and lead-dropout conditions. The proposed model is compared with learning-based baselines and, for the simulated dataset, classical Tikhonov- and Twomey-type solvers. The main finding is that waveform similarity and localization accuracy are not equivalent: the method with the highest correlation does not necessarily best recover the earliest-activation site. Across the tested dataset--condition combinations, the proposed model achieves the lowest localization error, including cases where baselines obtain higher waveform correlation. Ablation results show that the spatial and temporal refiners provide most of the localization gain, while the activation-time loss further reduces localization error, especially on measured data.

Subject Keywords

Electrocardiographic imaging, inverse problem, Kolmogorov–Arnold networks, activation-time mapping, deep learning.

URI

https://hdl.handle.net/11511/120644

Collections

Graduate School of Natural and Applied Sciences, Thesis