B-PO01-088 LOCALIZATION OF OUTFLOW TRACT PREMATURE VENTRICULAR BEATS OR VENTRICULAR TACHYCARDIA IN SURFACE ELECTROCARDIOGRAMS USING A CONVOLUTIONAL NEURAL NETWORK

Abstract

ECG based algorithms for localization of outflow tract (OT) premature ventricular complexes (PVCs) have variable sensitivity and specificity. Machine learning may identify visually imperceptible ECG changes and augment predictive accuracy, especially in early disease stages. To compare a Convolutional Neural Network (CNN) with manual PVC localization of left vs. right ventricular (RV vs. LV) OT PVC. All patients with successful ablation of RV-LV OT from 1/2013-1/2020 were included if they had at least one standard 12-lead ECG recorded with a clinical PVC before ablation. PVC origin was defined by the site of successful ablation excluding LV summit, aorto-mitral continuity and para-hisian PVCs. Success was defined as absence of the targeted PVC for the 24 hr post procedure monitoring. We compared CNN performance to 3 OT PVC localization algorithms (Table 1). Patient-level ECG data were split into Training, Validation and Test Datasets in a ratio of approximately 7:1:2. Results are reported as averaged across 10 random splits of the data and model initializations for robustness. 308 ECGs (86 RVOT, 43 LVOT, 179 sinus) from 75 patients were used for CNN development. CNN classified RVOT PVC with similar sensitivity to algorithms but with higher specificity (Table 2). For LVOT PVC, CNN specificity was higher, and sensitivity was higher than all but one, of the manual algorithms. The CNN area under the receiver operating characteristic curve for LVOT and RVOT were 0.929 and 0.914, respectively. A CNN can achieve higher specificity at similar or higher sensitivity compared to most published algorithms to differentiate right from left OT PVCs.