Children's hospitals are turning open source AI frameworks from NVIDIA into clinical tools for some of the hardest problems in pediatric cardiology. According to NVIDIA, Children's Hospital of Philadelphia is using open source AI tools to model children's hearts in seconds, with the goal of safer, more precise care for kids born with congenital heart disease.

The work spans several institutions and several open source projects, but the pattern is the same: tasks that once took specialists half an hour of manual image processing now run in seconds, and the models are being built on frameworks anyone can download.

Confirmed

The most concrete published result comes from Cincinnati Children's, where cardiologists and researchers are using deep learning to match heart donors with recipients. The traditional method compares donor and recipient height and weight — an inexact approach that can lead to unnecessarily rejected matches. A newer method calculates Total Cardiac Volume (TCV) from MRI or CT scans, but doing it manually means downloading and processing more than 300 images, which takes about 30 minutes, according to Ryan Moore, director of Digital Health Innovation at Cincinnati Children's Heart Institute.

The team's deep learning model does the same calculation in seconds. In the group's published study, deep-learning-generated TCV identified matches with 94.5% accuracy — equal to manual identification — and the accuracy held across different ages, weights, and heights. Moore brought the concept to NVIDIA, which introduced the team to Project MONAI (Medical Open Network for Artificial Intelligence), a set of open source frameworks for medical imaging AI. The work is supported by a National Institutes of Health grant led by David Morales, executive co-director of the Heart Institute.

A second thread runs through Boston Children's Hospital, where cardiologist John Triedman and cardiology fellow Joshua Mayourian have launched the Congenital Heart Artificial Intelligence (CHAI) Lab. Its focus is the AI-powered electrocardiogram: cheap, widely available ECG readings analyzed by models that can detect conditions the human eye misses. One model, described in The Lancet: Digital Health, predicted ventricular dysfunction — a heart not squeezing normally — from ECG data alone. Another, published in JACC: Clinical Electrophysiology, interpreted heart rhythms at an expert level and outperformed commercial ECG software at spotting Wolff-Parkinson-White syndrome and long QT syndrome.

The motivation is global. Studies cited by the lab suggest as many as 90 percent of children in some low- and middle-income countries receive limited heart care, largely because advanced diagnostics never reach them. An ECG machine, by contrast, is affordable and nearly universal.

On the simulation side, a Stanford University team has built a one-dimensional physics-informed model using MeshGraphNet, a graph neural network architecture, implemented in NVIDIA's open source PhysicsNeMo framework. The model simulates patient-specific blood flow in near real time, trained on data from 3D finite-element simulations run in the SimVascular software suite across eight cardiovascular geometries from the Vascular Model Repository. According to the team, the GNN surrogate outperformed physics-driven one-dimensional reduced-order models, particularly on complex geometries with multiple junctions and stenoses.

Unknown

None of these tools is yet a deployed, regulated clinical product at scale. The Cincinnati TCV model is still gathering cardiac scans for validation; wider deployment is planned through NVIDIA's federated learning network, which trains models across institutions without moving patient data, but that rollout has no stated date. The Stanford blood-flow surrogate remains research-stage — the team is extending it toward 3D models and integrating it into SimVascular, but it is not a clinical tool today.

Accuracy figures like 94.5% come from the research groups' own published studies, and independent replication across other transplant centers has not yet been reported. Moore himself flags the implementation gap: MONAI's application packages make integration feasible, but each hospital still needs someone to do the heavy lifting of fitting AI into clinical workflows.

Why it matters

The open source angle is the story's real lever. A proprietary cardiac AI model locks smaller hospitals out; MONAI and PhysicsNeMo let a children's hospital with one data scientist build on work done at Cincinnati or Stanford. For transplant matching specifically, speed has clinical weight — donor decisions happen in the middle of the night, under time pressure, and a 30-minute manual process competes with a narrowing window. Morales says the approach has already started to demonstrate improved waitlist survival, though that outcome data is still early.

Our take

The honest bottleneck here is not model quality — it is deployment. Every institution in this story describes the same gap between a validated model and a plug-and-play clinical tool, and the federated learning plan only works if peer hospitals can actually integrate what arrives. Watch whether the promised learning network materializes; that, not another accuracy point, is what determines whether this stays a showcase or becomes standard pediatric cardiac practice.

Sources