Surgery
Partner With Us
Advertisement
From the Journals Conference News The American Society of Echocardiography

AI Tops Guidelines in Classifying Aortic Stenosis

Gradient-boosted model shows higher diagnostic accuracy—especially for severe AS—in low-flow cases where guideline-based metrics often yield discordant results.

Conexiant
A gradient-boosted model leveraging convolutional neural network–derived echocardiographic measurements achieved 72% accuracy in classifying aortic stenosis severity, compared with 45% for a deterministic algorithm based on current guideline thresholds.
 
The research, conducted by Hema Krishna, MD, and colleagues, was chosen as a 2025 Arthur E. Weyman Young Investigator’s Award Competition Finalist. Findings were presented at the American Society of Echocardiography (ASE) 2025 annual meeting and as an abstract in the Journal of the American Society of Echocardiography.
 
“Unlike many ‘black box’ AI tools, our approach is more transparent. We used an FDA-cleared echocardiography platform (Us2.ai) to automate standard hemodynamic measurements - cardiologists can see and adjust inputs, while the model quickly processes calculations like valve area and stroke volume,” said Dr. Krishna, of the University of Illinois, Chicago, in an exclusive interview with Conexiant.
 
For identifying severe aortic stenosis (AS), the AI approach delivered 91% sensitivity and 96% specificity, outperforming the deterministic algorithm’s 63% sensitivity while matching its 96% specificity.

Performance Across Patient Populations

The gradient-boosted model (GBM) was trained on 537 patients and validated on 214 within a 968-patient cohort who underwent echocardiograms from 2005 to 2023. The data set was enriched for low-flow AS (stroke volume index ≤35 mL/m²), a diagnostically challenging scenario.
 
Feature-importance analysis identified mean gradient, peak aortic jet velocity, and dimensionless index as the most salient markers, all measured automatically by convolutional neural networks (CNNs). In the low-flow subset, the GBM maintained 72% accuracy, whereas the guideline-based algorithm showed variable sensitivity across severity grades.
 
In the low-flow subset, the GBM achieved 72% accuracy even in hemodynamically complex cases where guideline-based approaches showed variable sensitivity across severity grades and particular weaknesses in intermediate classifications, noted Krishna, from the University of Illinois, Chicago, and colleagues.

Clinical Outcomes and Concordance

The GBM was closely aligned with cardiologist classifications, with only 38 discrepant cases across the cohort. However, cardiologist assessment remained superior for predicting clinical outcomes—aortic valve replacement, heart-failure hospitalization, mortality, and the composite endpoint.
 
“This model will not replace expert judgment. It’s designed to work alongside cardiologists, providing near-instant diagnostic support,” said Dr. Krishna.

Clinical Implications

The CNN-augmented GBM addresses a key limitation in current AS classification—evaluation in low-flow states that often produce discordant results. This approach “may produce time savings and improve reproducibility of classification across readers and sites, performing well even in complex hemodynamic states," noted Dr. Krishna and colleagues.

“Although this was a retrospective cohort, the data included a broad mix of vendors, acquisition settings, and image quality, which supports confidence in generalizability. Looking ahead, two steps are essential: prospective, multicenter validation to confirm accuracy across diverse populations; and workflow studies to define what happens when AI and cardiologists disagree—safeguards such as clinician override and clear visibility of input parameters are critical to build trust. Our goal is not only higher accuracy, but also safe integration into daily practice,” said Dr. Krishna.

Disclosures were not available at press time.

Source: Journal of the American Society of Echocardiography