Prediction of aortic stenosis progression using artificial intelligence: a machine learning model based on echocardiographic data

Aortic Stenosis (AS) is amongst the most common valvular heart diseases, particularly in older individuals. Left untreated, the condition carries a significant risk of morbidity and mortality. Current guidelines for monitoring AS progression focus on serial echocardiographic assessment, which is resource-intensive and subject to variability. Artificial intelligence (AI) may offer an opportunity to enhance the early identification of patients at risk of developing severe AS.

Our objective was to create an echo-based model that can predict whether a patient will deteriorate from mild/moderate AS to severe AS.

We retrospectively analyzed a single-center database of 529,751 echo exams and identified 9,330 echocardiograms performed on patients initially diagnosed with mild or moderate AS who had progressed to severe AS within 5 years of follow-up, as defined by an expert echocardiographer. We developed a model agnostic to any patient data outside the scope of the echocardiography report. Performance was assessed for accuracy, AUC-ROC, and calibration. SHAP values provided interpretability for the model’s predictions.

During the follow-up period, 1,625 (47%) patients developed severe AS during the follow-up period. The model demonstrated a strong predictive performance, manifested by an AUC-ROC of 0.93, (Figure 1) an accuracy of 85%, a sensitivity of 86%, and an Integrated Calibration Index = 0.0721 (Figure 2). Key predictors for severe AS development included maximal and mean systolic aortic valve gradients, calculated aortic valve area, and ascending aortic diameter. The model successfully identified patients at high risk of progression, with robust calibration and generalizability confirmed through cross-validation.

Our novel, echocardiography-focused AI model is a reliable tool for the early identification of patients at risk of progression to severe AS. Pending future, multi-center, prospective validation, such models may facilitate personalized follow-up strategies and timely interventions, ultimately leading to improved patient outcomes resource utilization.

ROC Curve