A graph neural network trained on simulated body surface potential maps to predict atrial fibrillation ablation success
M Macarulla-Rodriguez, J Sanchez, C Fambuena-Santos, C Herrero-Martin, R Moreno-Lopez, M Martinez-Perez, I Llorente-Lipe, J Vicente, R Molero, I Hernandez-Romero, A M Climent, J Osca-Asensi, M S GuillemAbstract
Background
Success rates for atrial fibrillation (AF) catheter ablation remain suboptimal, with outcomes typically around 70% for paroxysmal AF and barely exceeding 50% for persistent AF, highlighting a critical need for tools that can non-invasively predict patient-specific outcomes to guide therapy. We hypothesized that a graph neural network (GNN), trained on simulated data, could learn to identify complex patterns in pre-procedural body surface potential maps (BSPMs) that are predictive of success after pulmonary vein and posterior wall isolation (PVI+PW).
Purpose
This study aims to develop a GNN model capable of predicting the success of PVI + PW ablation using non-invasive input data derived from BSPM (Figure 1).
Methods
A GNN was developed and trained exclusively on a large dataset of 13,447 in-silico (simulated) AF episodes. Ablation "success" for training was defined as arrhythmia termination following a virtual PVI+PW procedure. The GNN model's input consists of pre-ablation BSPMs, where torso electrodes are treated as nodes in a graph to capture complex spatiotemporal relationships. A proof-of-concept validation was then performed in a cohort of 12 patients undergoing PVI+PW ablation. The model generated a success score between 0 and 1 per patient, using 0.5 as the threshold to predict benefit from PVI+PW ablation, and was compared against the 6-month clinical outcome (arrhythmia recurrence) (Figure 2).
Results
At 6-month follow-up, 7 of the 12 patients (58%) were free from arrhythmia (success group), while 5 (42%) had recurrence (recurrence group). The GNN model correctly predicted the successful outcome (Figure 2) in 6 of the 7 patients in the success group (sensitivity: 86%). Conversely, it correctly predicted recurrence in only 2 of the 5 patients in the recurrence group (specificity: 40%). This resulted in an overall accuracy of 67% (8 out of 12 patients correctly classified).
Conclusion
We demonstrate the feasibility of a novel approach: training a GNN entirely on in-silico data to predict clinical AF ablation outcomes from non-invasive BSPMs. The model shows a strong ability to identify patients who will likely benefit from PVI+PW ablation. While validation in larger cohorts is required to improve specificity, this proof-of-concept establishes a powerful novel framework for developing AI-driven tools to improve patient selection for personalize AF ablation therapy.Purpose pictureresults picture