B-PO04-062 COMPARING THE ACCURACY OF FEATURES TO IDENTIFY ATRIAL FIBRILLATION ON INTRACARDIAC ELECTROGRAMS

Abstract

Cardiac devices are challenged when separating atrial fibrillation (AF) from other tachycardias (AT) from single electrograms (EGM). For this reason, devices tend to group arrhythmias by a rate threshold. We set out to determine the accuracy of multiple clinically-intuitive EGM features to separate AF from AT, using a deep learning (DL) logic. In 86 patients (61 male, age 65±11) we obtained a total of 29,340 EGMs of AF (4 s) from intracardiac channels (fig A) and AT pre-ablation. EGM variability was seen in both arrhythmias. We comprehensively probed EGM using 49 clinically-intuitive features based on cycle length, frequency and amplitude (fig B), analysed individually and combined. We developed custom convolutional (CNN) and recurrent (RNN) neural networks. Classifiers were tested using the same 10-fold cross-validation folds (80% patients for training, 20% for testing). Single EGM features modestly separated AF from AT with c-statistics ranging from 0.55 (rate) to 0.82 (autocorrelation of EGM shape). Accuracy increased when features were combined (fig C) Linearly (c-statistic 0.95 ± 0.04), by Bagged Trees (0.95 ± 0.04), K-Nearest Neighbors (0.95 ± 0.04) or Support Vector Machines (0.95 ± 0.04). DL on raw EGMs provided higher accuracy without the need to calculate features, with c-statistics up to 0.97 ± 0.04 (RNN; fig C). Traditional EGM features separate AF from AT, but ‘high rate events’ as a sole criterion is poor. Combining shape and regularity indices provides the optimal approach. This limitation must be considered in future approaches to automatically detect AF from devices to guide clinical care.