Largest AI-Echo Model Shows Promise for Increasing Workflow Efficiency

A new artificial intelligence (AI)-based model called EchoPrime, trained on more than 10 times more data than existing models, appears to provide more comprehensive interpretation of echocardiograms, according to a new study.

The model can synthesize multiple videos of different views, and it’s hoped that EchoPrime, which is not autonomous, will be integrated into clinical workflows to help analyze echocardiographic data—both video and text—alongside physicians, said senior author David Ouyang, MD (Cedars-Sinai Medical Center, Los Angeles, and Kaiser Permanente Northern California, Pleasanton, CA).

“Right now, there’s actually a shortage of sonographers and cardiologists that read echo,” Ouyang told TCTMD. “We hope that a technology like this will very much democratize access to cardiovascular care.”

Clinical testing is underway on about 1,200 echocardiograms, with results expected by late 2026, he reported.

Previous models like EchoNet, from Ouyang’s team, and the one studied in the PROTEUS trial, among others, have paved the way for AI in cardiac imaging, but EchoPrime stands out because of the size of its training dataset and its ability to provide comprehensive reports, say researchers.

“This is a vision-language model, meaning that it has the diversity of natural language in its output,” Ouyang explained. “A lot of other models will either say yes or no to a limited set of tasks, whereas this model uses kind of free text to really describe what’s happening. And the other piece is that this is a model that’s . . . 10 to 100 times bigger than all prior models.”

The study was published online last week as an accelerated article in Nature, with first author Milos Vukadinovic (Cedars-Sinai Medical Center).

EchoPrime was trained with 12,124,168 echocardiography videos and paired text reports from 275,442 studies across 108,913 patients treated at Cedars-Sinai Medical Center. Compared with other previous models, it resulted in a mean area under the curve (AUC) of 0.92 for identifying cardiac structure and pathophysiology across a range of benchmarks in the training dataset.

When the researchers evaluated the model on videos from four outside institutions—Stanford Healthcare (n = 91,746), Beth Israel Deaconess Medical Center (n = 75,768), Chang Gung Memorial Hospital (24,724), and Kaiser Permanente (n = 201,752)—the mean AUCs were 0.89, 0.85, 0.86, and 0.88, respectively.

Without any additional fine-tuning or training, EchoPrime outperformed or matched existing single-task prediction models, including several EchoNet algorithms assessing LVEF, tricuspid regurgitation, and mitral regurgitation, as well as PanEcho, EchoCLIP, and BiomedCLIP.

The researchers say EchoPrime will be open source so that it “can be a resource to the medical and AI communities,” they write. “The Achilles’ heel of medical imaging lies in human heterogeneity, and opportunities in the future lie in the potential integration of AI into the healthcare system. Our results represent an important step towards the automated evaluation of cardiac ultrasound.”