Abstract
<title>Abstract</title> <p>Background Accurate grading of aortic regurgitation (AR) is essential for follow-up, treatment planning, and prognostic assessment. Transthoracic echocardiography is the first-line modality for AR evaluation, but conventional grading depends on image quality, operator experience, and multiparametric interpretation, leading to subjectivity and interobserver variability. This study aimed to develop and validate a Swin Transformer-based model for automated AR severity classification using standardized apical five-chamber color Doppler echocardiographic images. Methods This multicenter retrospective study included 2,575 adult patients with AR from three medical centers between January 2022 and August 2025. Data from Center A were divided at the patient level into a training set (n = 1,347), validation set (n = 337), and internal test set (n = 422). Data from Centers B and C formed an independent external validation set (n = 469). AR severity labels were assigned according to the 2017 American Society of Echocardiography recommendations by experienced echocardiographers after standardized training. Each patient contributed one representative apical five-chamber color Doppler image. Performance was assessed using accuracy, receiver operating characteristic analysis, F1-based metrics, and confusion matrices, and compared with ResNet18 and VGG. Results In the internal test set, ResNet18, VGG, and Swin Transformer achieved accuracies of 93.05%, 91.71%, and 93.36%, respectively. In external validation, the corresponding accuracies were 80.81%, 87.63%, and 92.11%, indicating more stable cross-center generalization for the Swin Transformer. Confusion matrices showed fewer adjacent-grade misclassifications. Gradient-weighted Class Activation Mapping (Grad-CAM) visualization demonstrated that the model primarily focused on color Doppler regions corresponding to the AR regurgitant jet. Conclusions A Swin Transformer model based on standardized apical five-chamber color Doppler images showed promising performance for automated AR severity classification and stable external generalizability. The model may serve as an adjunctive tool for echocardiographic AR assessment, although further validation with multiview imaging, quantitative parameters, and higher-level reference standards is warranted.</p>