AI learning for pediatric right ventricular assessment: development and validation across multiple centers.

Congenital and acquired heart disease affects approximately 1% of children worldwide, and right ventricular (RV) dysfunction is a common and complex manifestation in conditions such as congenital heart disease, pulmonary hypertension, and prematurity. Accurate RV assessment remains difficult due to the ventricle's irregular geometry and morphological variability in pediatric patients. Using 24,984 echocardiograms from 3993 children across four tertiary centers in North America and Asia, we developed and validated a video-based deep learning framework for automated RV functional assessment. The model performs frame-level ventricular segmentation and beat-by-beat estimation of fractional area change (FAC), classification of RV-related disease, and exploratory prediction of left ventricular ejection fraction (LV EF). A U²-Net architecture achieved high segmentation accuracy (Dice = 0.86 [A4C], 0.88 [PSAX]) and classification performance (AUC = 0.95 U.S., 0.97 Asia). In LV EF prediction, the model outperformed previous methods across cohorts. This validated framework enables expert-level, real-time quantification of pediatric ventricular function, enhancing diagnostic consistency, reducing manual workload, and supporting earlier intervention for children with heart disease, particularly in resource-limited settings.