Abstract：Objective To investigate the performance of different artificial intelligence (AI) models in predicting severe stenosis (≥90%) of major coronary arteries or their main branches in patients with apparently normal ECGs. Methods The preoperative ECGs of patients undergoing selective coronary angiography were collected. From these ECGs we screened out 392 apparently normal ECGs, matched them with coronary angiography results, and then constructed a dataset. The enrolled patients were grouped by the degree of stenosis in major coronary arteries or their main branches. There were 138 patients with severe stenosis (the degree of stenosis≥90%, case group) while there were 254 cases with mild or without stenosis (the degree of stenosis＜50%, control group). We then constructed three deep neural networksbased models: Model 1, Model 2 and Model 3, based on training from scratch using ECG, training from scratch using ECG and clinical information, and transfer learning from a pre-trained ECG classification model, respectively. Results Model 1 eventually achieves an AUC value of 0.65, a sensitivity of 62.5%, a specificity of 66.7%, and F1 score of 0.483. The above performance parameters of Model 2 are separately 0.66, 63.0%, 69.2% and 0.563. The above performance parameters of Model 3 are separately 0.63, 55.6%, 53.8%, and 0.438. Conclusion Model 2 outperforms Model 1 while Model 3 performs slightly worse than Model 1. AI models have potential value in predicting severe coronary stenosis using apparently normal ECGs.
