Comparative Interpretation of Pediatric Electrocardiograms by Large Language Models and a Pediatrician

Large language models (LLMs) have been explored for clinical applications, yet their reliability in pediatric electrocardiogram (ECG) interpretation remains unclear. We compared the diagnostic accuracy of three general-purpose artificial intelligence (AI) systems-ChatGPT, Gemini, and Copilot-with that of a pediatrician and characterized error patterns across the platforms.

A total of 178 children (median age 108 months) underwent standard 12-lead ECG recording at a single center between January 2024 and July 2025. One pediatrician and three LLMs interpreted each ECG using a zero-shot design; responses were compared against a definitive diagnosis established through expert consensus. Rhythms were classified as physiological or pathological, with subgroup analysis for rhythms calling for emergency intervention.

The pediatrician achieved the highest accuracy (78.7%), followed by ChatGPT (71.9%), Copilot (66.9%), and Gemini (50.0%). Area under the receiver operating characteristic (ROC) curve confirmed pediatrician superiority (0.79) over ChatGPT (0.70), Copilot (0.63), and Gemini (0.54). Agreement with definitive diagnosis was moderate for the pediatrician (κ = 0.566) and ChatGPT (κ = 0.410), but weak or nonsignificant for Copilot and Gemini. Error profiles differed markedly; Gemini showed a high false negative rate (39.9%), while Copilot showed a high false positive rate (24.7%).

The clinician outperformed the LLMs in pediatric ECG interpretation. Diagnostic accuracy varied markedly across the models, with reliability limits particularly evident in high-risk rhythms. In their current form, these systems may serve as supplementary tools under expert oversight rather than independent decision-makers.