PO-03-124 DEVELOPMENT AND VALIDATION OF A DEEP NEURAL NETWORK TO MEASURE QTC ON PEDIATRIC ELECTROCARDIOGRAMS

Abstract

A prolonged Bazett corrected QT (QTcB) either congenital or acquired is one of the primary risk factors in the development of ventricular arrhythmias and sudden cardiac death in children. QTcB determination is notoriously problematic for non-pediatric cardiologists and currently available automatic electrocardiogram (ECG) interpreters. We set out to develop and validate a deep neural network (DNN) to measure QTcB and diagnose prolonged QTc on pediatric ECGs. All resting ECGs from both inpatient and outpatient settings at a quaternary freestanding children’s hospital performed on children 0-18 years of age and read by a pediatric electrophysiologist (PEP) for which the QTcB was manually measured between January 2010 and December 2020 were included. Using the manual QTcB measurements as labels and matrices of all 15 lead voltages as features a convolutional DNN with residuals blocks was trained to measure the QTcB (unweighted DNN) and a dynamic bias was added to optimize the false negative rate and make the model more conservative (dynamically-weighted DNN). The QTcBs on a prospective sample of 200 randomly selected ECGs from 2021 were manually measured by each of 3 PEPs. The ECGs were further classified as pathologically prolonged if the QTcB >=460 ms, and the diagnostic accuracy of the algorithms and the PEPs validated against the PEP consensus estimate of the QTcB. The DNN was trained on a sample of 65,370 ECGs from 37,992 patients. The mean absolute error (MAE) of the PEPs against the consensus of all others was 17.3 (16.2-18.4) ms. Both the unweighted DNN (MAE 10.9 (10.3-11.5) ms) and the dynamically-weighted DNN (MAE 12.8 (12.0-13.6) ms) outperformed the PEPs. The sensitivity of individual PEPs to detect a prolonged QTc (>=460 ms) was 72.7% which was significantly better than the unweighted DNN (64.3%, p=0.045) but significantly worse than the dynamic DNN (92.8%, p<0.001). The positive likelihood ratio of the PEPs was lower than the uneweighted model (p=0.006) but higher than the dynamic model (p=0.002). The negative likelihood ratio of the dynamic DNN (1:12.8) outperformed both the unweighted DNN (1:2.7) and the PEPs (1:3.5). A dynamically-weighted DNN can serve as an automatic screening test to accurately detect pathologically prolonged QTc’s on pediatric ECGs given a higher sensitivity and lower negative likelihood ratio than pediatric electrophysiologists while achieving more accurate estimates of the QTc itself.