Machine learning-driven ECG screening for early coronary artery disease detection

Background Electrocardiography (ECG) is inexpensive and widely available, yet its ability to detect early and nonobstructive coronary artery disease (CAD) is limited when interpreted using conventional criteria. We evaluated whether a deep learning model applied to standard 12-lead-ECG images can identify angiographically confirmed CAD and enable clinically meaningful risk stratification in a real-world population. Methods In this multicentric study, 15,890 ECG–angiography pairs (12-lead ECG paired with coronary angiography [CA] or coronary CT angiography [CCTA]) were included between November 2021 and January 2024. A DenseNet-121 convolutional neural network was trained to predict CAD from ECG images. Temporal validation was performed on a prospectively collected, independent cohort of 2168 samples from February to May 2024. Model performance was assessed using receiver operating characteristic curve (ROC-AUC), PR-AUC, and standard diagnostic metrics. Robustness across varying disease prevalence was evaluated using simulation, and prevalence-dependent predictive performance was summarized using PPV and NPV prevalence curves (AUCppv and AUCnpv). Results The model was trained on 13,299 samples, validated on 1,002, and tested on 1339 (CAD prevalence 52.9%). On the test set, ROC-AUC was 0.80 and PR-AUC 0.83. At the optimal 0.50 threshold, sensitivity was 87% (618/708 CAD cases detected) and specificity 50% (317/631 non-CAD correctly classified). On the temporal validation set ( n = 2168; 64% CAD), ROC-AUC was 0.81 and PR-AUC 0.83, with sensitivity 0.67, specificity 0.78, PPV 0.84, and NPV 0.57. Simulated prevalence analyses demonstrated robust PPV/NPV trends across distributions. Conclusion Deep learning analysis of routinely acquired ECG images can identify patients at increased risk of CAD and may support low-cost, scalable CAD risk stratification.