A Multi-Task Deep Learning Model for Pediatric Echocardiography Analysis

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Pediatric Imaging

Background:

• Congenital heart defects affect nearly 1% of newborns globally.
• Current deep learning models for echocardiography are limited to adult analysis or single tasks in pediatrics.
• Existing pediatric models analyze limited echocardiographic views, hindering comprehensive assessment.

Purpose of the Study:

• To introduce EchoAI-Peds, the first multi-task deep learning model for pediatric echocardiography.
• To develop a model capable of integrating information from multiple echocardiographic views simultaneously.
• To address the limitations of single-task and single-view models in pediatric echocardiography analysis.

Main Methods:

• A video-based vision transformer was trained to detect 28 abnormalities from complete pediatric echocardiography studies.
• The model integrates information from all available views for unified study-level predictions.
• Trained on over 700,000 videos from 11,000+ studies, validated on internal and external datasets.

Main Results:

• Achieved macro-averaged AUROC of 0.91 (internal) and 0.89 (external).
• Significantly outperformed adult-based echocardiography foundation models (p < 0.001).
• Demonstrated robust performance across diverse patient demographics and study types.

Conclusions:

• Multi-task deep learning models hold significant potential for aiding pediatric echocardiogram interpretation.
• Results emphasize the critical need for AI models specifically tailored to pediatric cardiac conditions.
• EchoAI-Peds represents a significant advancement in automated analysis of pediatric echocardiography.