Explainable Deep Learning-based Classification of Wolff-Parkinson-White Electrocardiographic Signals

TL;DR

The study tackles noninvasive localization of antegrade accessory pathways in WPW by casting it as a 24-class time-series classification problem learned from a large synthetic ECG database generated via a cardiac digital twin. It systematically compares three input formats (stacked univariate, multi-channel, and 2D image-like) and integrates post-hoc explainability methods (Guided Backpropagation, Grad-CAM, Guided Grad-CAM) to produce interpretable predictions. The best models achieve localization accuracies above $95\%$ with high sensitivity ($\approx94\%$) and specificity ($\approx99.8\%$), and XAI outputs align with known electrophysiological patterns; a lead-importance index identifies $V2$ as the most informative lead. The work demonstrates the feasibility of explainable deep learning combined with digital twins for precise, non-invasive WPW AP localization and offers practical guidance for lead selection and model design in clinical workflows.

Abstract

Wolff-Parkinson-White (WPW) syndrome is a cardiac electrophysiology (EP) disorder caused by the presence of an accessory pathway (AP) that bypasses the atrioventricular node, faster ventricular activation rate, and provides a substrate for atrio-ventricular reentrant tachycardia (AVRT). Accurate localization of the AP is critical for planning and guiding catheter ablation procedures. While traditional diagnostic tree (DT) methods and more recent machine learning (ML) approaches have been proposed to predict AP location from surface electrocardiogram (ECG), they are often constrained by limited anatomical localization resolution, poor interpretability, and the use of small clinical datasets. In this study, we present a Deep Learning (DL) model for the localization of single manifest APs across 24 cardiac regions, trained on a large, physiologically realistic database of synthetic ECGs generated using a personalized virtual heart model. We also integrate eXplainable Artificial Intelligence (XAI) methods, Guided Backpropagation, Grad-CAM, and Guided Grad-CAM, into the pipeline. This enables interpretation of DL decision-making and addresses one of the main barriers to clinical adoption: lack of transparency in ML predictions. Our model achieves localization accuracy above 95%, with a sensitivity of 94.32% and specificity of 99.78%. XAI outputs are physiologically validated against known depolarization patterns, and a novel index is introduced to identify the most informative ECG leads for AP localization. Results highlight lead V2 as the most critical, followed by aVF, V1, and aVL. This work demonstrates the potential of combining cardiac digital twins with explainable DL to enable accurate, transparent, and non-invasive AP localization.

Figures (10)

• Figure 1: insertion points grouped according to the defined regions (left) and bullseye representation of the anatomical region distribution (right). The regions were defined based on , including both rotational and longitudinal subdivision.
• Figure 2: Regional variation of the signals with in the . Mean signal (black) with shaded regions representing plus or minus 2 standard deviations across samples. Next to the leads, the insertion sites in the ventricles are displayed together with the region numbering.
• Figure 3: Regional variation of the signals with in the . Mean signal (black) with shaded regions representing plus or minus 2 standard deviations across samples. Next to the leads, the insertion sites in the ventricles are displayed together with the region numbering.
• Figure 4: Data misclassified by in Section \ref{['subsec:image']}. (a) of misclassified data are plotted with the color of the region they have been assigned to. (b) Comparison of a misclassified test sample (dashed) with a correctly classified training sample (continuous) assigned to the same class.
• Figure 5: (GC), (GB), and (Guided GB) saliency maps, and activation patterns across ECG leads vary with the anatomical location and conduction behavior of . (A) contributing to the ongoing wavefront highlights features in the QRS complex in multiple leads, especially aVL and aVF. (B) Right ventricular demonstrates localized activation in leads V1 and V2, consistent with activation disruptions from the RV free wall. (C) initiating early ventricular activation produces diffuse explainability across several leads. (D) in the RV free wall outside the His-Purkinje system reveals two distinct ECG regions: the delta wave and His-Purkinje capture.