B-PO05-151 AUTOMATIC CLASSIFICATION OF MACRO-REENTRANT ATRIAL TACHYCARDIA MECHANISMS USING 12-LEAD ECG

Abstract

Atrial Flutter (AFl) is a common reentrant atrial tachycardia driven by self-sustainable mechanisms. Intracardiac electrophysiological mapping and catheter ablation are often performed without detailed prior knowledge of the mechanism perpetuating AFl in a given patient, prolonging the procedure time of these invasive interventions. We discriminated 3 types and direction of atrial tachycardia by a 7-class classifier: typical AFl, perimitral macro-reentry - both clockwise and counterclockwise; figure-8 macro-reentry around the pulmonary veins - in anterior and posterior direction; and others (e.g., focal source, micro- and scar-related reentry). Simulations of several atrial tachycardia scenarios were performed to compute 12-lead ECGs. The virtual study cohort yielded 2512 ECGs in total. For each ECG, we extracted 151 features. K-nearest neighbor classifier with leave-one-atria-out cross-validation was trained, validated, and tested on the simulated dataset. The resulting classifier was tested on a clinical dataset (15 patients - 2 for each class, except for "Others" with 3 patients). The most significant feature for classification was the F-wave duration. The classifier achieved an in-silico test accuracy of 74.4%, and a clinical test accuracy of 60.0% for a 7-class classification. A machine learning approach can potentially identify several AFl types and directions using the 12-lead ECG. Simulations can support training of machine learning when clinical ground truth data are scarce. This non-invasive method could aid clinical procedure planning for AFl ablation after validation in a bigger cohort.