B-PO05-180 MACHINE LEARNING PREDICTION OF PULMONARY VEIN STENOSIS IN PATIENTS UNDERGOING ATRIAL FIBRILLATION ABLATION

Abstract

Data regarding the predictors of pulmonary vein stenosis (PVS) after atrial fibrillation (AF) ablation is limited. Develop a machine learning (ML) prediction model for PVS after AF ablation and compare its performance with logistic regression (LR) and to determine the predictors of PVS. We identified patients undergoing AF ablation at our institution from 2016 to 2019. The study outcome was development of PVS. We divided the study dataset into 80% training and 20% testing sets. The features were selected using the random forest technique. Random forest (RF) machine learning and LR algorithm were used to train the models on the selected features for predicting PVS post-AF ablation. 4318 atrial fibrillation procedures were analyzed of which 18 (0.416%) developed PVS. A total of 82 candidate variables were included of which 9 features were selected using the RF technique. The RF algorithm significantly outperformed LR in estimating PVS (0.90 (95% CI, 0.84-0.95) vs 0.85 (95% CI, 0.82-0.88), p<0.001). The 9 features in the order of importance were presence of atrial scarring, CHA2DS2VASc score, NYHA class, coronary artery disease, redo pulmonary vein isolation, persistent atrial fibrillation, cryoballoon ablation, dissociated firing, and common ostium. Amongst these variables, CHA2DS2VASc score, cryoballoon, disseminated firing, and common ostium were inversely associated whereas the others were directly associated with the development of PVS. This is the first risk prediction model for estimating PVS post-AF ablation and this tool can be of aid to clinicians in prognostication.