Identifying outcome predictors of radiofrequency ablation of atrial fibrillation using patient-specific computational models and machine learning

Abstract Up to 40% of atrial fibrillation (AF) ablations require a repeated procedure. Patient-specific computational models could predict pulmonary vein isolation (PVI) outcomes, but their value depends on which routinely acquired measurements inform the outcome-relevant parameters. We aimed to identify the electrophysiological determinants of arrhythmia sustainability and PVI benefit, and how informative each measurement is about them. We built 11 left atrial digital twins from late gadolinium enhancement MRI and electroanatomical (bipolar voltage and local activation time, LAT) mapping, using a five-parameter tissue model with spatial heterogeneity (20 parameters). For each anatomy, 600 parameter samples were simulated for arrhythmia and PVI and classified into four endpoints. Four classifiers (logistic regression, gradient boosting, random forest, and multilayer perceptron) predicted binary arrhythmia sustainability and PVI benefit, and Shapley and permutation importance ranked the predictors. We also simulated dense LAT mapping during coronary-sinus pacing and conduction velocity (CV) restitution from an S1S2 octa Ray protocol, and quantified how informative each protocol is about each parameter with a Gaussian-process Sobol analysis. Maximum conduction velocity (CVmax) and maximum action potential duration (APDmax), which together set the activation wavelength, were the leading determinants of sustainability across all classifiers. For PVI benefit, CVmax dominated in pre-ablation AF and left atrial surface area in atrial tachycardia, with tissue heterogeneity playing a secondary role. Dense LAT mapping was informative about CVmax, whereas the S1S2 protocol was jointly informative about APDmax and CVmax through the effective refractory period and CV restitution. Maximum conduction velocity and action potential duration are therefore the primary determinants of arrhythmia sustainability and, in pre-ablation AF, of PVI benefit, and two routinely acquired mapping protocols are jointly informative about these outcome-relevant parameters, supporting personalised, image-based ablation planning.