Abstract 18852: Machine Learning to Predict Atrial Fibrillation Recurrence Using Pulmonary Vein Morphology

Abstract

Introduction: Predicting risk of recurrence and sites of recurrence after atrial fibrillation (AF) ablation help direct ablation strategies. While the left atrium (LA) has been implicated as a possible site of AF recurrence leading to ablation directed toward LA substrates, reconnection of pulmonary veins (PVs) has been observed as a mechanism of recurrence on repeat ablations. Hypothesis: Regions of statistical shape differences between AF recurrence (AF+) and non-recurrence patients (AF-) within the PV on a CT scan might be associated with sites of recurrence, that features extracted from these regions may be better predictors of AF recurrence. Methods: A Hierarchical Atlas construction method for PV branching shape analysis was developed. Patients with the median volume of LA plus PV in each cohort were selected as the templates. All CT scans in each cohort were registered to the cohort’s template. Then, each branch was individually registered to the branch’s template and resulting control points applied to the LA and other PV branches to maintain connectivity. This process was repeated with 6 iterations to better adjust branching. Two separate atlases of AF+ and AF- were constructed. The areas of significant shape differences on the PV, surfaces of interest (SOIs), between the two groups were obtained via a General Linear Model based t-test on the atlases. Normal and curvature-based features were extracted from SOI regions. A Gradient Boosting classifier was trained based on these features. Cross-validation and atlas construction involved 104 cases. 50 cases served as a holdout set. Results: The classifier trained on SOI features on CT scans achieved AUCs of 0.76 (cross-validation) and 0.74 (holdout set). SOIs include sites within the left superior PV and the right PV carina. Conclusions: This finding appears to suggest that SOI regions on PVs tend to be associated with recurrence and hence may be candidate sites for additional ablation for preventing recurrence of AF.