Abstract 15166: Automated Analysis of 12-lead Electrocardiogram for Pediatric Cardiac Disease Mass Screening in School-age Children by the Combined Use of Signal Processing and Deep Learning

Abstract

Introduction: Electrocardiogram (ECG) is useful for screening children at risk of sudden cardiac death. However, high resources for manual interpretation were a drawback when undertaking mass screening; the accuracy of conventional automated ECG analysis was unsatisfactory for this purpose. Hypothesis: A model consisting of signal processing and deep learning is newly developed and validated to analyze ECG in school-age children and has higher accuracy than a conventional automated model. Methods: We obtained 12-lead ECGs performed in consecutive patients at 6-18 years of age at the inpatient or outpatient clinic of a tertiary hospital in Japan during 2003-06. Patients were randomly assigned to training (83%) and test (17%) groups. Each ECG was labeled as normal or abnormal by 3 board-certified pediatric cardiologists in accordance with the JCS guideline 2016 for school ECG screening. A model to analyze several-second digital wave data of 12-lead ECG was developed by combining signal analysis and deep learning (fast Fourier transform and deep convolutional neural network) to predict abnormality. The trained model was evaluated in the test data using ROC curve and by comparing its accuracy with that of a conventional model (Minnesota code assigned by ECG-1400, Nihon kohden, Japan). Results: We included 1842 ECGs performed in 1062 patients (median age 11 years, IQR, 8-14; male 56%), in which 519 ECGs (28%) were labeled abnormal. Findings include ST-T abnormality (16%), right bundle branch block (7.8%), QRS axis abnormality (5.9%), right ventricular hypertrophy (3.6%) and left ventricular hypertrophy (3.3%). Of 310 ECGs for test (27% labeled as abnormal), 20% were considered abnormal by our model, while 73% by the conventional model. Our model showed AUC of .87 and higher accuracy than the conventional model (.85 vs .47; P <.001, McNemar test). Sensitivity and specificity were .60 and .95 in our model; .86 and .33 in conventional model, respectively. Conclusions: A model combining signal processing and deep learning is newly developed and validated for analyzing 12-lead ECG in school-age children, achieving higher accuracy than a conventional model. This proof-of-concept study warrants further studies toward deep learning-based school ECG mass screening.