# Machine Learning-Based Prediction of Three-Year Heart Failure and Mortality After Premature Ventricular Contraction Ablation

## TL;DR

This study uses machine learning to predict heart failure and mortality after PVC ablation, finding that LightGBM and logistic regression with ROSE provide the best performance.

## Contribution

The paper introduces a robust, clinically interpretable risk stratification model for post-PVC ablation outcomes using modern machine learning techniques.

## Key findings

- LightGBM with ROSE achieved the highest ROC AUC of 0.822 for predicting three-year heart failure.
- Logistic regression and LightGBM with ROSE showed balanced performance for predicting three-year mortality with ROC AUCs of 0.886 and 0.882.
- Age, prior heart failure, malignancy, and end-stage renal disease were the most influential predictors identified.

## Abstract

Introduction: Long-term heart failure and mortality after catheter ablation for premature ventricular contraction (PVC) remain underexplored. Methods: We retrospectively analyzed 4195 adults who underwent PVC ablation in a nationwide claims database. To address class imbalance, we used synthetic minority over-sampling technique (SMOTE) and random over-sampling examples (ROSE). Five supervised algorithms were compared: logistic regression, decision tree, random forest, XGBoost, and LightGBM. Discrimination was assessed by stratified five-fold cross-validation using the area under the receiver operating characteristic curve (ROC AUC). Because rare events can bias ROC, we also examined precision–recall (PR) curves. Results: For predicting three-year heart failure, LightGBM with ROSE achieved the highest ROC AUC at 0.822. For three-year mortality, logistic regression with ROSE and LightGBM with ROSE showed balanced performance with ROC AUCs of 0.886 and 0.882. Pairwise DeLong tests indicated that these leading models formed a high-performing cluster without significant differences in ROC AUC. Age, prior heart failure, malignancy, and end-stage renal disease were the most influential predictors by model explainability analysis. Discussion: Addressing class imbalance and benchmarking modern learners against a transparent logistic baseline yielded robust, clinically interpretable risk stratification after PVC ablation. These models are suitable for integration into electronic health record dashboards, with external validation and local threshold optimization as next steps.