DEEP LEARNING IMPROVES ON CLASSICAL FEATURES FOR IDENTIFYING ATRIAL TACHYARRHYTHMIAS FROM INTRACARDIAC ELECTROGRAMS

Miguel Rodrigo, Albert J Rogers, Prasanth Ganesan, Chayakrit Krittanawong, TIna Baykaner, Mahmood Alhusseini, Sanjiv M Narayan

Abstract:

Background

Cardiac devices are increasingly used to identify atrial arrhythmias (AA) from benign rhythms such as sinus tachycardia (ST). However, this is based on electrogram (EGM) features predominantly based on ‘high rates’, that may be suboptimal.

Objective
To test the hypothesis that deep learning improves on traditional rate and regularity based identification of AA from ST.

Methods
In 71 persistent AF patients (50 male, 65±11 years) we recorded 18±5 unipolar intracardiac EGMs prior to ablation (fig A) providing 25,560 non-overlapping 4 sec segments. We developed convolutional neural networks (CNN) to detect AA from ST, with 10-fold cross-validation (80% patients for training, 20% for independent testing). CNN performance was compared to cycle length (CL), Dominant Frequency (DF) and autocorrelation of EGM shape (AC; fig B) alone and with complex classifiers that combined features linearly, or via Forest Trees and K-Nearest Neighbors (KNN).

Results
Classical analyses were modestly effective for identifying AA, with c-statistics of 0.70 (for CL), 0.67 (DF) and 0.75 (AC). Performance increased when features were combined (fig C) linearly (c-statistic 0.96 ± 0.03), by Forest Trees (0.95 ± 0.03) or KNN (0.94 ± 0.03). However, CNN of raw EGMs without any hypothesis-based features provided near perfect c-statistics of 0.996 ± 0.004 (fig C; p<0.001).

Conclusion
CNN of raw EGMs outperforms traditional rule-based identification of AA from ST. This approach could improve device diagnosis, and may shed insights into classification and characterization of AA beyond classical features such as rate and regularity.