Deep Learning-Based Multiclass Classification of Mitral Valve Etiologies Using Limited B-Mode and Color Doppler Echocardiography: Internal and External Validation.

BACKGROUND

Accurate etiologic classification of the mitral valve (MV) is essential for guiding clinical management but remains dependent on expert visual interpretation. Despite advances in artificial intelligence-based quantitative analysis, automated morphologic interpretation under routine imaging conditions remains limited.

OBJECTIVES

The aim of this study was to develop and validate a deep learning framework for multiclass classification of major MV etiologies using limited routine transthoracic echocardiography (TTE) views.

METHODS

A multiview deep learning model was developed to classify 5 MV etiologies (normal, rheumatic, degenerative, prolapse, and functional). The developmental data set comprised 4,344 TTE examinations from a nationwide multicenter registry. Validation was performed using an internal test data set and an independent external test data set (2,262 TTE examinations). Prespecified subgroup analyses were conducted according to mitral regurgitation (MR) severity and automated image quality.

RESULTS

The model demonstrated robust performance across all MV etiologies in both internal and external data sets. In the internal test data set, area under the receiver operating characteristic curve values ranged from 0.968 to 0.997 across etiologies, with higher performance observed for normal valves and rheumatic disease. In the external test data set, discriminative performance remained preserved (area under the receiver operating characteristic curve, 0.931-0.992), despite differences in disease distribution and MR severity. Sensitivity for MV prolapse increased markedly with moderate or greater MR compared with mild MR, whereas degenerative disease showed persistently lower sensitivity across MR severity. Diagnostic performance remained stable across image quality strata, with comparable accuracy and macro-F1 scores in all-adequate and partially suboptimal examinations. In post hoc analyses of cases with multiple MV etiologies, the model correctly identified at least 1 expert-assigned etiology in 85.7% of cases.

CONCLUSIONS

Deep learning-based analysis of limited, routinely acquired TTE views enables reliable multiclass classification of MV etiologies. This approach may complement quantitative automation and expert visual assessment, supporting more consistent and scalable MV evaluation in routine echocardiographic practice.