Abstract
INTRODUCTION: Transthoracic echocardiography (TTE) is the current standard for detecting tricuspid regurgitation (TR); however, it incurs additional costs and is dependent on the operator's skill. In contrast, the 12-lead electrocardiogram (ECG) is widely available during initial evaluations. This study aimed to develop deep learning (DL) models using 12-lead ECG signals and clinical features to detect significant TR. METHODS: Between 2017 and 2019, a total of 5432 patients who underwent both 12-lead ECG and TTE were eligible for this study. Significant TR was identified in 570 of these patients. The DL model architecture was based on a combination of one-dimensional convolutional neural network, efficient channel attention block, and Multihead Attention modules. The model was trained on data from 3910 patients, tested on 435 patients, and validated using both internal and external cohorts. RESULTS: The diagnostic performance of the DL model using ECG signals, age, and sex to predict significant TR was as follows: an accuracy of 0.762, sensitivity of 0.809, specificity of 0.756, and an area under the curve (AUC) of 0.857. After incorporating additional factors such as RR interval, QRS duration, corrected QT interval, atrial fibrillation, and hypertension into the DL model, the diagnostic performance remained substantial, with an accuracy of 0.762, sensitivity of 0.836, specificity of 0.752, and an AUC of 0.866. External validation of the DL model showed satisfactory results. CONCLUSIONS: Implementing the DL model for ECG interpretation could facilitate the diagnosis of significant TR. However, the clinical utility of this DL model requires further validation and exploration.