Performance of a Machine Learning Algorithm for Electrocardiogram Interpretation in Pediatric Student Heart Check Screenings

Document Type

Conference Proceeding

Publication Date

9-2025

Publication Title

American Journal of Preventive Cardiology

Abstract

Background: The use of artificial intelligence (AI) to interpret electrocardiograms (ECGs) has the potential to enhance the efficiency of cardiac screenings in remote or resource-limited settings. This study evaluates the performance of an opensource AI-based algorithm validated in a population >16 years of age in a younger student population. Methods: We analyzed a sample of 61 ECGs from a student heart check program, comparing the on-site cardiologist interpretations to Yale University’s CarDS open AI algorithm (ECG-GPT©, www.cardslab.org/ecg-gpt). A third-party pediatric cardiologist resolved ECGs where the AI and on-site cardiologist disagreed. Results: 61 ECGs were analyzed (32 normal and 29 abnormal as determined by the on-site pediatric cardiologist). The mean age of participants was 15 years. In the overall cohort, AI agreement with the on-site read was 74%, with 47% false positive and 3.4% false negative for determining whether an abnormality was present or not. In the cohort with normal ECG findings, AI agreement with on-site cardiology read was 53%. In the cohort with abnormal ECG findings, there was 97% agreement that an abnormality was present. Sensitivity was 97%, specificity was 53%, positive predictive value was 65%, negative predictive value was 94%, and accuracy was 74%. Of the 28 ECGs that aligned in their abnormal reports, 7 had conflicting findings. Of the ECGs where there was disagreement between the AI and onsite cardiologist, a third-party pediatric cardiologist agreed with the on-site cardiologist for 48%, agreed with the AI for 13%, and disagreed with both for 39%. Conclusions: The CarDS open-AI demonstrated modest performance in this cohort, which is younger than the validation cohort for this algorithm. It was superior at detecting abnormal ECG findings with a low incidence of false negative rates compared to normal ECGs. There was high sensitivity for recognizing abnormal ECGs but high rates of disagreement about the type of abnormality.

Volume

23

Issue

Suppl

First Page

60

Comments

ASPC (American Society for Preventive Cardiology) 2025 Congress on CVD Prevention, August 1-3, 2025, Boston, MA

Last Page

60

DOI

10.1016/j.ajpc.2025.101203

Link to Full Text

Share

COinS