Interpretable ECG-Based Classification of Accessory Pathway Laterality in Pediatric Wolff-Parkinson-White Using Cardiac Digital Twins

Sandra Perez-Herrero1, Guadalupe Garcia Isla2, Ferran Prats-Domenech3, Ferran Roses-Noguer3, Beatriz Trenor4, Javier Saiz5, Jorge Sanchez1
1Center for Research and Innovation in Bioengineering (Ci2B), Universitat Politecnica de Valencia, 2Universidad Politécnica de Valencia, 3Hospital Vall d'Hebron, 4Universitat Politècnica de València, 5Centro de Investigacion e Innovacion en Bioingenieri­a (Ci2B), Universitat Politecnica de Valencia


Abstract

Accessory pathway (AP) localization from the 12-lead ECG remains clinically relevant in pediatric Wolff–Parkinson–White (WPW) syndrome, as it can guide pre-procedural planning, catheter access, and identification of high-risk septal locations. However, accurate localization in children remains challenging due to age-dependent variability and the limited reliability of existing algorithms, which are often developed in adult populations and rely on subjective feature interpretation.

We propose an interpretable, simulation-driven framework for AP laterality classification based on decision trees trained on automatically extracted ECG features. A patient-specific cardiac digital twin was constructed from CT imaging, including ventricular geometry, fiber orientation, and a cardiac conduction system network. A dataset of 696 simulated ECGs was generated across 174 AP locations and four pre-excitation timings. Features capturing early activation were extracted from all leads, and ground-truth labels were defined directly from the simulations, enabling objective left-versus-right classification.

The model achieved high performance in simulated data (balanced accuracy 0.968, macro-F1 0.956), with only 18 misclassifications. The resulting decision tree was compact and interpretable, relying on a small number of features, primarily QRS balance in lead V1, complemented by V2 short-window slope and lead III QRS balance. When evaluated on an independent clinical dataset of 10 pediatric ECGs, the model achieved a balanced accuracy of 0.708 and macro-F1 of 0.697, indicating preliminary generalization to clinical ECGs.

These results show that ventricular laterality can be captured using a limited set of quantitative ECG descriptors derived from simulated data. The proposed framework provides interpretable decision rules linked to electrophysiological mechanisms and offers a reproducible alternative to subjective ECG assessment, representing a step toward integrating cardiac digital twins and machine learning for non-invasive pre-procedural planning in pediatric WPW.