Identifying Outcome Predictors of Radiofrequency Ablation of Atrial Fibrillation Using Patient-Specific Computational Models and Machine Learning

Abstract Background: Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia. Catheter-based radiofrequency ablation that targets pulmonary vein isolation (PVI) is the primary interventional treatment. However, to 40% of patients require repeat procedures. Patient-specific computational models hold promise for predicting ablation outcomes, but their clinical utility depends on identifying which electrophysiological measurements that are routinely acquired effectively inform the most important model parameters. Methods: We constructed 11 adult human atrial digital twins from LGE-MRI images, voltage mapping, and local activation time (LAT) data. We used a 5-parameter phenomenological electrophysiological model. Spatially varying personalised left atrial electrophysiological characteristics were described with a total of 20 parameters. For each anatomy, we obtained 600 parameter samples and simulated arrhythmias both pre- and post-PVI. We classified each arrhythmia simulation into one of four endpoints (non-initiated, self-terminating, atrial tachycardia (AT), atrial fibrillation (AF)), and this data set was used to further score the effectiveness of PVI. Next, four classifiers (logistic regression with polynomial feature expansion, gradient boosting, random forest, and a multi-layer perceptron) were trained to predict (i) the binary sustainability of the pre-ablation arrhythmia (i.e. being sustained vs non-sustained) and (ii) the binary ablation benefit, stratified using pre-ablation arrhythmias (AT, AF). The most predictive parameters were identified with permutation importance and Shapley analysis. Finally, for each sample, we simulated: a dense electroanatomical mapping in response to pacing from the coronary sinus, and an S1S2 decremental pacing protocol, assuming stimulation and recordings were obtained using an octa Ray catheter. The dimensionalities of the simulated local activation times (LAT) maps and of the conduction velocity (CV) restitution curves were reduced by principal-component analysis, retaining three components per output. Using Sobol indices and Gaussian process emulators trained on these components, we quantified which parameters were identifiable from each protocol. Results: The four classifiers agreed on the leading predictive features. The homogeneous components of the maximum conduction velocity (CVmax) and of the maximum action potential duration (APDmax) were the primary determinants of atrial arrhythmia sustainability. In addition, the left atrial surface area was a determinant feature when the classifiers were trained on the aggregated dataset, i.e., when augmented with anatomy-dependent features. For PVI benefit in the setting of AF, CVmax was the leading parameter, while the left atrial surface area dominated in pre-ablation AT. Spatial heterogeneity played a secondary, modulatory role. Dense mapping was fundamentally linked to CVmax, and as expected, the effective refractory period of the S1S2 protocol was strongly modulated by APDmax. Conclusions: In our simulations, the maximum conduction velocity (CV) and the maximum action potential duration (APD) are the primary electrophysiological determinants of arrhythmia sustainability and PVI benefit. The left atrial surface area emerged as the dominant anatomical factor for ablation of pre-existing atrial tachycardia. Dense LAT mapping is informative about CVmax, while the S1S2 protocol (through ERP and CV restitution) is informative about the single and combined effects of APDmax and CVmax. In summary, results from two routinely acquired protocols jointly constrain the outcome-relevant parameters, supporting their combined use to personalise computational models for ablation plans and evaluation.