Intra-cardiac Signatures of Atrial Arrhythmias Identified by Machine Learning and Traditional Features

Abstract

Intracardiac devices separate atrial arrhythmias (AA) from sinus rhythm (SR) using electrogram (EGM) features such as rate, that are imperfect. We hypothesized that machine learning could improve this classification.

In 71 persistent AF patients (50 male, 65 ± 11 years) we recorded unipolar and bipolar intracardiac EGMs for 1 min prior to ablation, providing 50,190 unipolar and 44,490 bipolar non-overlapping 4 s segments. We developed custom deep learning models to detect SR or AA, with 10-fold cross-validation, compared to classical analyses of cycle length (CL), Dominant Frequency (DF) and autocorrelation.

Classical analyses of single features were modestly effective with AUC ranging from 0.91 (DF) to 0.70 for other rate metrics. Performance increased by combining features linearly (AUC 0.991/0.987 for unipolar/bipolar), by Bagged Trees (0.995/0.991) or K-Nearest Neighbors (0.985/0.991). Convolutional deep learning of raw EGMs with no feature engineering provided improved AUC of 0.998/0.995 to separate AA from SR.

Deep learning of raw EGMs outperforms classic rule-based classifiers of SR or AA. This could improve device diagnosis, and the logic developed by deep learning could shed novel insights into EGM analyses beyond current classification based on EGM features and rules.