Comprehensive Summary
Upton et al. developed an artificial intelligence (AI) system for automated interpretation of stress echocardiography studies to detect severe coronary artery disease (CAD). The model was trained on 578 patients from the multicenter Echocardiography Value and Accuracy at REst and STress (EVAREST) study and extracted 31 important geometric and kinematic features from the left ventricular motion of patient hearts. In an independent validation with 154 U.S. patients, the AI model achieved an AUROC of 0.927, with a specificity of 92.7% and sensitivity of 84.4%. A randomized crossover reader study showed that use of AI model classifications increased sensitivity of clinician reads by 10% (85.0% to 95.0%) without loss of specificity, while also improving inter-reader agreement and diagnostic reliance.
Outcomes and Implications
AI-assisted stress echocardiography showed a meaningful sensitivity gain without loss of specificity, suggesting it could reduce false negatives and improve consistency in CAD detection. Because they enhance inter-reader agreement and diagnostic reliability, AI tools may help standardize interpretation in clinical practice, especially in lower-expertise or high-volume settings. However, this model was trained on a small-scale dataset and validated in a singular cohort, and its binary classification system may oversimplify disease complexity. Reliance on visually adjudicated angiography also raises questions about the reference standard. Larger prospective trials are essential in determining whether AI integration improves outcomes or reduces unnecessary angiography before routine adoption.