Cardiac Arrhythmia Prediction from Pediatric ECGs: A Novel Machine Learning Risk Model

Abstract

Machine Learning (ML) models are valuable tools in healthcare for early disease detection, risk stratification, and improved diagnostic accuracy, leading to more effective disease management. ML’s application in cardiology is especially promising, as the early detection and prediction of cardiac events can significantly improve clinical outcomes through timely intervention. However, the early detection of cardiovascular disease is limited in pediatric populations due to age-related physiological fluctuations in ECG metrics. This in-progress study addresses these challenges by developing a Pediatric Arrhythmia Risk Assessment Tool (PARAT) based on traditional ML models and deep learning approaches, including convolutional neural networks, to predict future arrhythmia onset in seemingly normal pediatric electrocardiograms (ECGs). The study first used Children’s National Hospital’s patient database to collect retrospective pediatric ECGs recorded between January 2000 and June 2025. The goal of data collection was to create a training dataset from two patient groups: a) control patients with no arrhythmia and b) patients with normal ECGs who later developed arrhythmia. For each ECG, information such as the reason for ECG, QT interval length, Bazett-corrected QT interval length, and heart rate was extracted. A total of 249,579 pediatric ECGs were collected, of which 54,216 were arrhythmogenic. The next steps involve training the ML models with the two datasets and evaluating their performance. Following its training and performance evaluation, PARAT aims to predict future arrhythmic events in pediatric patients, thereby improving diagnostic accuracy and disease management.