Abstract 14742: Deep Learning of Intracardiac Electrograms in Atrial Arrhythmia

Abstract

Introduction: Intracardiac devices detect atrial fibrillation (AF) by rate and regularity, but any inaccuracies may cause inappropriate use of anticoagulants or anti-arrhythmic medications. Hypothesis: Machine learning of raw intracardiac electrograms can identify AF from other atrial arrhythmias better than traditional measures of rate or regularity and without using specific electrophysiological analyses such as dominant frequency (DF). Methods: In 86 persistent AF patients (25 female, age 65±11) we recorded 64 unipolar intracardiac electrograms over 60 seconds prior to ablation (fig A). We trained deep learning models (comprising 2X1D-convolutional layers and 2 dense layers) on successive 4-sec segments labelled AF or Flutter/tachycardia (AFL), using 10-fold cross-validation with 80% of patients for training and an independent 20% for testing. We compared results to classical statistical and machine learning (ML) analyses of electrograms featurized by 30 metrics of cycle length (CL), DF and autocorrelation-based metrics (AC; fig B). Results: Identification of AF varied between methods, but was modest for features of CL (c-statistic 0.70), DF (0.67) and AC (0.75). ML that combined features improved results: linear combination (c-statistic 0.95 ± 0.04), Bagged trees (0.92 ± 0.06), k-nearest neighbors (0.92 ± 0.06) and support vector machines (0.95 ± 0.04). Deep learning using raw electrograms as input (no featurization) provided AUC of 0.95 ± 0.05 (fig C). Conclusions: Detailed machine learning of raw intracardiac electrograms identified AF more accurately than traditional indices of rate, regularity, and dominant frequency. This approach could reclassify AF detection from devices to improve management, and may reveal novel AF phenotypes with distinct clinical courses.