The newly developed open-source EchoNet-AS deep learning model accurately assessed aortic stenosis (AS) by combining B-mode videos, Doppler images and peak aortic velocity measurements, according to research published Oct. 5 in JACC: Cardiovascular Imaging. Its strong performance, validated in an external cohort, indicates its potential as a future clinical decision-support tool.

"Given that existing approaches do not fully leverage the complementary information available from multiple videos and Doppler measurements, we hypothesized that an integrated framework combining all relevant echocardiographic information could improve the accuracy and clinical reliability of AS assessment," model developers Hirotaka Ieki, MD, PhD, and colleagues write.

EchoNet-AS combines B-mode videos from parasternal long-axis (PLAX), parasternal short-axis (PSAX) and apical five- and three-chamber views and their corresponding PLAX, PSAX and apical Doppler images, as well as peak aortic jet velocity measurements.

Ieki and colleagues trained the model on 210,193 transthoracic echocardiographic videos from 16,076 studies involving 15,213 patients (mean age, 73 years; 48% women) from the Kaiser Permanente Northern California (KPNC) health system. It was validated using 1,588 held-out test studies and a temporally distinct cohort of 19,202 studies from KPNC, followed by external validation at Stanford Health Care (SHC; n=2,415) and Cedars-Sinai Medical Center (CSMC; n=9,038). While gender distribution, racial and ethnic composition, LVEF and major comorbidities were similar across cohorts, patients in the final validation cohorts from the external health systems were younger.

Results showed that in the held-out studies, EchoNet-AS achieved an area under the receiver-operating characteristic curve (AUROC) of 0.955 for ≥moderate AS and 0.964 for severe AS. In the temporally distinct cohort, AUROCs were 0.983 and 0.986, respectively. Performance remained strong in the SHC and CSMC cohorts: AUROCs for ≥moderate AS were 0.989 and 0.978, respectively, and AUROCs for severe AS were 0.985 and 0.989, respectively.

The new model outperformed multiview image-only models, including B-mode + color Doppler, B-mode-only and PLAX-only approaches, as well as the EchoPrime and PanEcho foundation models. The model maintained strong performance across subgroups, including low-flow, low-gradient AS and patients with concomitant valvular disease.

A one-category difference accounted for 95% of misclassifications. Patients who were misclassified were more likely to be older (mean age, 79 years vs. 73 years; p<0.001) and to have coexisting left-sided valvular disease (22.4% vs. 12.5%; p<0.001).

Artificial-intelligence (AI)-predicted severity of ≥moderate AS demonstrated a higher mortality hazard ratio than report-based assessment (1.33 vs. 1.29; p<0.001). Investigators found that peak aortic velocity was the model's most influential feature, with the PLAX view contributing most strongly among the echocardiographic views.

"Can AI read the echocardiogram like or better than an expert?" Jae K. Oh, MD, FACC, and Chieh-Ju Chao, MD, ask in an accompanying editorial comment. Rather than giving a binary answer, they explore how the current study paves the way for future approaches: "EchoNet-AS is a carefully validated, openly released demonstration that the path forward for valvular AI may lie not in ever larger generalist models but in disciplined, physiology-aware systems that fuse what we see with what we measure, much as we were trained to do."

Interested in advancing the field of cardiovascular imaging? Explore the call for applications for the next JACC: Cardiovascular Imaging Editor-in-Chief.