B-PO02-044 ELECTROGRAM FINGERPRINTS OF ATRIAL FIBRILLATION

Abstract

It is unclear whether disorganized waves in atrial fibrillation (AF) produce distinct ‘signature shapes’ on the electrogram (EGM), reflecting specific patterns. We tested the hypothesis that deep learning (DL) could be used to identify specific EGM shapes pathognomic for AF, that could form the basis for future approaches to detect subtype and treat patients. We developed custom-built convolutional (CNN) and recurrent (RNN) neural networks using N=29,340 EGM segments (4 sec), from 86 patients with AF or AFL at ablation (25 female, 65±11 years), and using 10-fold cross-validation (80% patients for training, 20% for independent testing). We probed CNN by generating 29,880 EGM sequences from patient data, in which we successively controlled variables including cycle length (CL) and EGM shape. Trained CNN and RNN identified AF from other atrial arrhythmias with an AUC of 0.95 ± 0.05 and 0.97 ± 0.04 in the testing sets respectively (Fig. A). We identified EGM shapes that were classified as AF even if they had regular rhythm and slow rate (long CL). A systematic analysis of 101 such AF signature EGMs were classified as AF >80% by CNN regardless of rate and irregularity (fig B, orange), and also by the RNN on 30 of such EGM signatures. Signature AF electrograms were identical by both RNN and CNN classification for 15 shapes (fig C, red shapes). Deep learning identified signature AF electrograms. EGM signature shapes were more important to AF L classification than high rate or irregular timing. Extraction of complex electrophysiological information hidden to the expert eye may improve diagnosis and shed insights into mechanisms of EGM generation in AF.