Identifying critical ablation targets in ischemic ventricular tachycardia (VT) remains challenging when sustained arrhythmia mapping is not feasible. Substrate mapping acquired during sinus or paced rhythm provides rich structural and electrophysiological information, but inferring relevant regions from these measurements alone is non-trivial. We propose a graph-based learning approach to predict VT ablation targets by jointly modeling local substrate features and their spatial organization.
Electroanatomical maps are represented as graphs where nodes correspond to spatial sampling points and edges connect neighboring locations based on geometric proximity. Node features include voltage measurements, activation times, conduction velocity, and markers of abnormal electrograms such as late potentials and local abnormal ventricular activities. The task is formulated as supervised node classification, where labels indicate proximity to clinically delivered ablation lesions. We employ Graph Neural Networks (GNNs) within a message-passing framework to learn node representations that integrate substrate features with local spatial context while respecting permutation equivariance. We compare these models to non-relational baselines, including an XGBoost classifier trained on node features independently. The models were trained on 21 cases in which a VT isthmus could be mapped, at least partially, during sustained arrhythmia, and subsequently validated on an independent cohort of 55 cases.
Results show that incorporating spatial structure significantly improves predictive performance. The GNN achieves a ROC-AUC of 0.86, outperforming the feature-based XGBoost baseline (ROC-AUC 0.78). These findings suggest that local spatial interactions and relational inductive biases contribute to improved characterization of arrhythmogenic substrate patterns. This work demonstrates the potential of graph-based representation learning for substrate-guided VT ablation and suggests that exploiting spatial structure can enhance clinical decision support in electrophysiology.