AI-Based View Classification in Pediatric Echocardiography: Transferability of an Open-Source Framework for Adult Echocardiography

All articles of this category (opens in new window) Background: Echocardiography is the most widely used imaging modality in pediatric cardiology. For research applications, particularly those involving computer vision and other AI methods, echocardiography videos must first undergo manual view labeling. Given the size of datasets—often comprising thousands of videos—this process can be monotonous and time-consuming. AI-based view labeling methods have shown promise in prior studies using adult data. EchoPrime is an open-source vision-language model for adult echocardiography interpretation (arXiv: 2410.09704), reporting a one-vs-rest AUC of 0.997 for predicting standard echocardiographic views. This preliminary study evaluates the applicability of EchoPrime for a view classification of pediatric echocardiography videos. Methods: A pediatric specialist with more than 5 years of echocardiography experience labeled 245 echocardiography videos from nine randomly selected examinations of a pediatric aortic valve disease cohort. Views were assigned to one of ten categories supported by the framework’s classifier (A3C, A4C, A5C, apical Doppler, Doppler parasternal long, Doppler parasternal short, parasternal long, parasternal short, SSN, subcostal). Forty-one videos classified as “other views” were excluded, leaving 204 videos for AI analysis using the EchoPrime framework, which outputs both classifications and class probabilities. Probability thresholds of 85% and 90% were applied to ensure high-confidence assignments. Performance was evaluated by measuring accuracy against human labels. Results: Overall, the framework classified 160/204 videos (78%) at a confidence level of ≥ 85% and 152/204 (75%) with a confidence level of ≥ 90%. At these thresholds, the system achieved accuracies of 0.838 (85% threshold) and 0.862 (90% threshold), respectively. Total accuracy, including low-confidence classification (0% threshold), was 0.740. The human-labeled views could potentially fit different labels from EchoPrime (e.g., subcostal A4C could correspond to A4C or Subcostal), making a clear human classification not always possible. Conclusion: Our findings suggest that adult view classification models may be effectively adapted to pediatric echocardiography, enabling automated labeling with encouraging accuracy. Further refinement is needed by incorporating large datasets, assessing performance in complex congenital heart disease, and ultimately developing AI models specifically tailored to pediatric cardiology. Publication History Article published online: 17 February 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany