Using Artificial Intelligence to Improve the Diagnosis of Pediatric Heart Disease
by Micaela Harris-Kim
June 28, 2026
Congenital heart defects are one of the most common structural abnormalities at birth, resulting in nearly 1% of births worldwide. Though, congenital heart defects are exceptionally complex, varying considerably in their anatomical complexity, impact on heart function, and long-term clinical outcomes. Clinically, congenital heart defects can be challenging to assess and diagnose, due to pediatric cardiac growth and postoperative changes in anatomical structure.
Echocardiography remains the primary procedure in assessing congenital heart defects due to its noninvasive nature, accessibility, and real-time results. Though, interpreting pediatric echocardiograms requires substantial expertise and a lot of time. Due to limited specialized clinicians and variability in image interpretation, there can be delays and inconsistencies in diagnoses, creating further hurdles in pediatric care.
With the rise of deep learning, there is considerable promise for enhancing assessment and diagnosis of congenital heart defects through automating and improving analyses. In adult populations, deep learning has progressed significantly, demonstrating its capabilities to support a broad range of analyses across echocardiographic views. However, comparable progress has yet to be made in pediatric populations.
A team of investigators at Stanford University, led by Joseph Cho, Mrudang Mathur, PhD, and William Hiesinger, MD, recently published in Circulation, sought to address this gap by developing EchoAI-Pads, a deep learning model designed for pediatric populations. Unlike previous approaches focusing only on single diagnoses, this model simultaneously detects 28 different congenital heart defects (CHD), structural and functional abnormalities, repairs, and interventions directly from complete pediatric echocardiography studies. EchoAI-Peds, like specialized pediatric cardiologists’ integration of information, combines data from across the entire study to generate a comprehensive assessment of CHD.
The model was developed using 12,000 studies at Stanford Medicine, encompassing a broad range of CHDs, with structural and functional abnormalities, surgical repairs, and interventions. In addition, the investigators evaluated the model’s performance with a distinct patient cohort at the Children’s Hospital of Philadelphia. Excitingly, EchoAI-Peds addresses the complex clinically relevant conditions in pediatric CHD, defying what has historically been a restriction in pediatric clinical utility. By leveraging large language models, the researchers were able to efficiently generate large-scale training data to capture the broad range of findings. Additionally, this approach increases deep learning potential for hospital use worldwide.
In the future, the team imagines EchoAI-Peds to be used in hospitals with limited pediatric cardiology expertise, serving as an automation tool to flag potential clinically relevant abnormalities. Additionally, they hope it can also serve as a secondary reader to assist experts in their interpretations of pediatric CHD data. The investigators hope to improve performance on existing diagnoses, broadening range of conditions, and enabling the model to provide diagnostic classifications and quantitative measurements. This deep learning model is an important step toward progression in automated pediatric echocardiography analysis and highlights the necessity of developing Artificial Intelligence systems specifically for pediatric populations.
Additional Stanford Cardiovascular Institute members include Dhamanpreet Kaur, Matthew Duda, Aravind Krishnan, Matthew Leipzig, Rohan Shad, Robyn Fong, Abhinav Kumar, Cyril Zakka, and Curtis Langlotz.
Joseph Yoshimura Cho
Dr. Mrudang Mathur
Dr. William Hiesinger