Deep learning ECG detection of wall motion abnormalities suggesting silent coronary artery disease
M Soh, Y Lee, H Hong, S Kang, H Lim, T Yu, Y Na, H An, S Eom, Y Ji, S Joo, J Park, J ShinAbstract
Background
Regional wall motion abnormalities (WMA) reflect myocardial ischemia or prior infarction and are associated with adverse cardiovascular outcomes. WMA may be present even in patients with preserved global ejection fraction, representing subclinical or silent coronary artery disease (CAD). Detection typically requires echocardiography or advanced imaging, limiting feasibility for large-scale screening. We evaluated whether a deep learning model could detect WMA directly from standard 12-lead electrocardiograms (ECGs).
Methods
We retrospectively analyzed paired ECG–echocardiography examinations collected between 2020 and 2023 at a tertiary academic medical center. WMA was defined as a wall motion score index >1.0 based on standard 17-segment echocardiographic analysis. ECG–echocardiography pairs were matched within ±14 days. A Vision Transformer architecture pretrained on unlabeled ECGs was fine-tuned to classify WMA using paired ECG–echo data. Data were split at the patient level into training, validation, and test sets (7:1:2). After matching and exclusions, 106,410 ECG–echocardiography pairs from 43,010 patients were included. The prevalence of WMA in the test cohort was 14.2%.
Results
The model achieved an AUROC of 0.919 (95% CI 0.910–0.928) in the test set. Using a threshold determined in the validation cohort where sensitivity equaled specificity, the model demonstrated sensitivity 0.838 (95% CI 0.815–0.858) and specificity 0.841 (95% CI 0.832–0.850). The positive predictive value was 0.466, while the negative predictive value reached 0.969, indicating strong potential as a rule-out screening tool.
Conclusion
A deep learning model using standard 12-lead ECGs demonstrated high accuracy for detecting regional wall motion abnormalities. ECG-based artificial intelligence may enable scalable, non-invasive screening for silent coronary artery disease and facilitate earlier identification of patients who may benefit from further cardiac evaluation.ROC curve of AI prediction for WMADiagnostic metrics of AI-ECG model for WMA