Abstract 11651: Machine Learning of Action Potential Shape to Define Refractory Periods in Ischemiccardiomyopathy

Abstract

Introduction: The effective refractory period (ERP) can characterize arrhythmic risk. To measure it clinically, a pacing protocol with 8-10 beats is required, preventing beat to beat analysis. Monophasic action potentials (MAP) can indicate ERP on every single beat (Fig A), but are often difficult to annotate due to a wandering baseline (Fig B). Hypothesis: We hypothesized that machine learning of ventricular MAPs could help identify instantaneous repolarization, without requiring manual measurement and without measuring time to drop to a specific repolarization voltage. Methods: We analyzed 5706 MAPs from 42 patients (age 65±13 Y) with ischemic cardiomyopathy in pacing. A multilayer perceptron (MLP) was developed to input MAP shapes and output action potential duration at 30, 60, 90% (MAPD 30,60,90) repolarization, indicating ERP. We used 60% of data in training, 20% in validation, 20% in testing. Results were compared to manual measurement. To identify which MAP regions were key to machine learning, we systematically perturbed MAP regions with white noise and assessed the drop in performance. Results: Errors between APD 30,60 and 90 from MLP and manual marking were 9.3(2.0), 5.8 (1.1) and 5.8 (0.7) ms respectively (fig. C). Fig D shows that MAP phases 0 and 3 were used by MLP to predict APD, as seen in the significant increase in error with perturbations in those MAP regions ( P <0.05). Conclusions: Machine learning can directly indicate effective refractory periods by learning the electrogram shape of single beats, without extrastimuli or manual measurement. Explainability studies confirm that machines learn physiological data. This approach can be used to assess beat-to-beat characterization of atrial fibrillation patients.