Feasibility of Machine Learned Intracardiac Electrograms to Predict Postinfarction Ventricular Scar Topography.

Abstract

Background: Accurate delineation of scar patterns is valuable for guiding catheter ablation of ventricular tachycardia. We hypothesized that scar and its pattern of distribution can be determined from intracardiac electrograms using computational signal processing and that further improvements in classification can be achieved with a convolutional neural network.

Methods: A total of 5 sheep underwent anteroseptal infarction (plus 1 healthy control) with electroanatomic mapping (129±12 days post-infarct). A whole-heart histological model of the postinfarction scar was created and coregistered to ventricular electrograms. Electrograms were matched to scar pattern categories; no scar, at least endocardial scar: at least intramural scar (intramural scar sparing the endocardium), or epicardial-only scar (epicardial scar sparing the endocardium/intramural space). A suite of signal-processing features was extracted from bipolar electrograms. Furthermore, bipolar and unipolar electrograms were used to train a time series convolutional neural network (InceptionTime).

Results: A total of 11 551 electrograms were matched to 451 biopsies. Bipolar and unipolar voltage alone were poor classifiers of scar patterns. For each of the scar labels, 20 bipolar electrogram features (predominantly within the frequency domain) yielded an area under the curve of 0.815, 0.810, 0.704, and 0.681 to predict no scar, at least endocardial scar, at least intramural scar, and epicardial-only scar, respectively. Substantial improvement was achieved with a convolutional neural network trained on unipolar electrograms: areas under the curve and accuracy (averaged across wavefronts) were 0.977 and 0.929 for no scar, 0.970 and 0.919 for at least endocardial scar, 0.909 and 0.959 for at least intramural scar and 0.926 and 0.958 for epicardial-only scar.

Conclusions: Convolutional neural network-derived analysis of unipolar electrogram data has excellent predictive value for determination of scar patterns. Computational analyses of electrogram data beyond voltage and other time-domain features are necessary to improve the identification of arrhythmogenic sites in the ventricle.