Enhancement of Stress ECG Performance with Machine Learning

Background: Exercise stress electrocardiogram (ECG) (ESE) is a widely used, noninvasive diagnostic tool for detecting coronary artery disease (CAD). Despite its widespread use, the diagnostic accuracy of ESE remains suboptimal. Objectives: This study aimed to develop and evaluate an artificial intelligence (AI) model, using a transformer-based architecture, to enhance the diagnostic performance ofESEs. Methods: Patients who underwent coronary angiography within 2 months of the ESE were eligible for inclusion. An AI model processed exercise stress ECG images into time-series data. A transformer-based architecture was employed to integrate temporal ECG features and predict CAD. Model performance in predicting severe CAD was first evaluated using 5-fold cross-validation on a test subset from the original cohort, and subsequently on a second validation cohort. Results: We developed a model using a total of 1,200 ECGs. An additional validation cohort of 91 patients was also analyzed. On the initial test subset, the AI model demonstrated a sensitivity of 93.6%, specificity of 93.2%, and overall accuracy of 93.4%. Notably, the model improved sensitivity with an absolute increase of 40.9% in women and 44.6% in men. In the second validation cohort, the model achieved an accuracy of 78%, with a sensitivity of 64.6% and a specificity of 93%. Conclusions: This study presents a proof of concept demonstrating that an AI-based model for stress ECG interpretation is feasible and shows acceptable performance.