Development of Artificial Intelligence to Predict Coronary Revascularization using Exercise ECG

Abstract

Background: Exercise stress electrocardiography (ExECG) is widely used for coronary artery disease evaluation, but its interpretation remains challenging due to variable diagnostic accuracy. I aimed to develop and validate an explainable artificial intelligence (AI) model to enhance the prediction of coronary revascularization need based on ExECG findings. Methods: The study included 20,534 patients who underwent ExECG using the modified Bruce protocol. I developed an explainable AI framework that first extracted clinically relevant ECG features using variational autoencoders and then trained a prediction model for coronary revascularization within 90 days after ExECG. Model performance was compared against clinicians and Duke Treadmill Score. Results: The pre-trained VAE model extracted clinically relevant ECG features by representing high-dimensional ExECG data with a small number of latent variables across exercise Stages (13–17 per Stage). The AI model demonstrated superior performance with an area under the receiver operating characteristic curve of 0.84 (95% CI: 0.80–0.88) compared to clinicians (AUROC, 0.75; 95% CI 0.71-0.80), and the Duke Treadmill Score (AUROC, 0.78; 95% CI 0.73-0.82). The odds ratio of coronary revascularization cases defined by AI was 12.37 (8.43-18.49), whereas the odds ratios determined by Duke Treadmill Score and physician diagnosis were 5.65 (3.02-9.40) and 19.65 (13.56-28.65). The model identified ST-segment depression in the mid-recovery phase as the most significant predictor of coronary revascularization need. Conclusion: I developed and validated an explainable artificial intelligence to predict coronary revascularization by using large-scale ExECG. By integrating advanced AI predictions with interpretable ECG feature analysis, my model may improve the diagnostic utility of ExECG in clinical practice.