Clinical Validation of Multivariate Adaptive Regression Splines for Electrocardiographic Imaging

Date

2026-7-3

Author

Almus, İbrahim Canberk

Metadata

Show full item record

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

Item Usage Stats

39
views

11
downloads

Cite This

Electrocardiographic Imaging (ECGI) aims to non-invasively reconstruct cardiac electrical activity, which can provide guidance for localization of arrhythmias like Premature Ventricular Contractions (PVCs). To address the fundamental challenge of solving the ill-posed inverse problem of ECGI, this thesis presents the clinical validation of a data-driven framework based on Multivariate Adaptive Regression Splines (MARS). The methodology was evaluated using clinical data from 10 patients with spontaneous PVCs, and successful radiofrequency ablation sites served as ground truth. Patient-specific simulations were employed as training data. To maximize computational efficiency and prevent overfitting, an adaptive complexity constraint was proposed. This approach reduces average training time from 43.4 to 7.1 hours while achieving a 15.99 mm mean localization error. Furthermore, when validated on an unseen cohort across diverse geometries, this adaptive strategy outperformed the optimal static constraint, reducing the overall error from 29.35 mm to 27.88 mm. The robustness of the MARS framework was systematically investigated under varying noise levels and structural geometric variations, including homogeneous, inhomogeneous, and hybrid geometry configurations. Additionally, the study assessed the impact of diverse activation time mapping methods and source localization strategies, which reveals that the Region of Activation (ROA) strategy decreased the overall mean localization error from 26.8 ± 15.6 mm down to 19.0 ± 6.2 mm compared to traditional point-based methods. Finally, results demonstrate that the adaptively constrained MARS framework consistently outperforms zero-order Tikhonov regularization by achieving an overall mean localization error of 25.2 mm compared to 36.1 mm.

Subject Keywords

Electrocardiographic Imaging, Inverse Problem, Multivariate Adaptive Regression Splines, Clinical Data, PVC Localization

URI

https://hdl.handle.net/11511/119947

Collections

Graduate School of Natural and Applied Sciences, Thesis