Abstract
BACKGROUND: This study aimed to develop a deep learning model (DLM) for rapid screening of coronary heart disease (CHD) using "pseudo-normal" electrocardiograms (ECGs), particularly focusing on patients who present with normal or near-normal ECGs at admission. METHODS: This study utilized standard 12-lead ECGs from CHD and non-CHD patients collected at the Second Affiliated Hospital of Nanchang University (SAH) between September 2017 and May 2019. These data sets were employed for training and validation of the DLM. For external testing, ECGs from CHD patients who underwent revascularization from January 2020 to May 2020 at the First Affiliated Hospital of Gannan Medical College (FAH) were used. The DLM's diagnostic performance was evaluated by analyzing the area under the receiver operating characteristic curve (AUC), using eigenvalue-based visual cluster analysis to categorize ECG diagnoses. RESULTS: The model was developed using a dataset comprising 21,240 ECGs from 15,995 patients at SAH, with 4248 ECGs serving as the internal testing set. Additionally, 2572 ECGs from FAH were utilized as the external testing set. The AUC for the SAH model was 0.913 (95% confidence interval [CI]: 90.4-92.1%), and for FAH, it was 0.936 (95% CI: 0.925-0.945). Remarkably, the DLM exhibited an AUC of 0.721 (95% CI: 0.665-0.773) for identifying CHD in patients with "pseudo-normal" ECGs. Certain parameters within the DLM demonstrated potential as significant indicators for screening "pseudo-normal" ECGs. CONCLUSION: Our deep learning model effectively facilitates rapid screening for CHD using ECGs, demonstrating particular efficacy in identifying CHD in patients with "pseudo-normal" ECGs. This approach offers a swift and accurate method for detecting CHD, augmenting traditional diagnostic techniques.