Researchers have developed an AI model that uses imaging data to deliver accurate risk assessments of patients with aortic stenosis (AS), sharing their findings in JACC: Cardiovascular Imaging.
The group hopes its work can help optimize the timing of aortic valve replacement (AVR).
“Interest has increased in alternative methods for risk stratifying patients with AS, including CT assessments of the valve (the aortic valve CT calcium score) and cardiovascular magnetic resonance (CMR) assessments of the myocardium (myocardial fibrosis and left ventricular remodeling), both of which appear to provide improved prognostic information,” wrote first author Partho P. Sengupta, MD, a cardiologist at West Virginia University Heart and Vascular Institute, and colleagues. “However, these imaging modalities are expensive, not widely available, and involve either radiation exposure or the administration of intravenous contrast agents. There is, therefore, a need for accurate yet simple methods to improve risk assessment in AS.”
Sengupta et al. explored echocardiography (ECHO) measurements from more than 1,000 patients—taken from a previous study of Canadian patients—to separate patients into different groups based on the severity of their AS. They then developed a “supervised machine-learning classifier,” verifying its performance with data from 752 patients who underwent ECHO and CT and another 160 patients who underwent ECHO and CMR.
Overall, the team found that 68% of the 1,964 patients included in the analysis had either nonsevere or discordant AS severity. The machine-learning identifier determined that 57% of patients had high-severity AS and the remaining 43% had low-severity AS.
The AI model was associated with more accurate risk assessments for patients in the ECHO, CT and CMR cohorts. It correctly classified approximately 99% of all patients in the high-severity AS subgroup, reclassifying patients with nonsevere or discordant AS severity as necessary.
“Together, our study findings demonstrate that our open-access machine-learning model can integrate ECHO features readily and meaningfully with robust performance across diverse international patient populations and provide powerful prediction of clinical events,” the authors wrote. “This approach holds major promise in optimizing the timing of AVR, particularly for patient groups where traditional echocardiographic assessments are inconclusive.”
The authors have made their classifier available to the public. Information on where to find that classifier, as well as the team’s full analysis, is available here.
Michael has more than 16 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.