Predicting PVI Non-Responders using machine learning on left-atrial electroanatomic mapping data

Background/Introduction

Although pulmonary vein isolation (PVI) is an established ablation strategy for atrial fibrillation (AF), a substantial subset of patients experiences AF recurrence despite durable PVI. Early identification of such non-responders during the index procedure could guide additional substrate ablation or closer rhythm follow-up.

Objective

To develop and evaluate a machine-learning (ML) model that identifies patients likely to exhibit AF recurrence despite durable PVI, using left-atrial (LA) electroanatomic mapping (EAM) and demographic data.

Methods

We retrospectively analyzed AF patients presenting for catheter ablation. PVI non-responders were patients undergoing repeat ablation for recurrence, with all veins confirmed durably isolated at re-do. Controls were patients without AF recurrence or patients with recurrence at re-do and ≥1 reconnected vein.

Low-voltage areas (LVAs) [Bipolar (≤0.5 mV) and unipolar (≤2.5 mV)] were quantified globally and regionally across predefined LA segments (see figure 1). Uni–Bi mismatch metrics (relative and mean-scaled dispersion) were computed to capture voltage discordance. These mapping features, combined with demographic variables (age, sex, BMI), were used to train an XGBoost classifier (Figure 2).

Data were split into training and test sets, scaled, and oversampled to balance classes. Feature importance was derived using SHAP values. Model hyperparameters were optimized through randomized-search cross-validation. Performance was assessed using the area under the precision–recall curve (PR-AUC) and F1 score on the test set.

Results

We analyzed 175 patients; 61 (34.9%) had AF recurrence despite durable PVI (non-responders). The dataset was partitioned into training (n = 140) and test (n = 35) cohorts. Cross-validated training performance was PR-AUC = 0.97 and F1 = 0.92. On the held-out test set, performance was PR-AUC = 0.91 and F1 = 0.80, with specificity = 1.00, precision = 1.00, and sensitivity = 0.67, indicating accurate identification of non-responders with low false positives.

Conclusions

An ML model trained on LA EAM features and demographics identified patients who recurred despite durable PVI with high precision on a held-out test set. Findings support the feasibility of ML substrate-guided risk stratification at the time of ablation. Prospective, multicenter validation is required to confirm clinical utility, assess generalizability, and address potential model overfitting.Figure 1.LVA Burden Quantification.Figure 2.ML Model