AI Detects Mitral Valve Disease from Ultrasound

A deep learning (DL) model trained on standard heart ultrasound scans can distinguish between different causes of mitral valve disease with high accuracy. The research, led by D. Jeong and colleagues, was published as a preprint on medRxiv.

The mitral valve regulates blood flow between the heart’s left chambers. Determining the exact cause of its dysfunction is vital for treatment but currently relies on expert cardiologists visually analyzing echocardiograms. This new work suggests artificial intelligence could provide a consistent, automated second opinion using the same limited views acquired in a routine exam.

How the AI Model Classifies Heart Valve Disease

The research team built a multi-view DL framework to classify the mitral valve into five categories: normal, rheumatic, degenerative, prolapse, and functional. They developed the model using 4,344 transthoracic echocardiography (TTE) examinations from a nationwide multicenter registry in South Korea.

A critical step was testing the model’s generalizability. The team used both an internal test set and a completely independent external test set of 2,262 exams from a different institution. This external validation is essential to see if an AI tool works outside the specific hospital system where it was created. The model analyzed only the limited B-mode and color Doppler views that are standard in a routine TTE, making it potentially easy to integrate into existing clinical workflows.

Robust Performance Across Different Hospitals and Image Quality

The model’s diagnostic performance was strong. In the internal test dataset, the area under the receiver operating characteristic curve (AUROC) ranged from 0.968 to 0.997 across the five etiologies. Performance was highest for identifying normal valves and rheumatic disease.

More importantly, results held up in the external dataset. AUROC values there ranged from 0.931 to 0.992, despite differences in patient demographics and disease severity between the institutions. The model’s sensitivity for detecting mitral valve prolapse was significantly higher in cases with moderate or greater mitral regurgitation compared to mild cases. Sensitivity for degenerative disease, however, was lower across all severity levels.

Notably, the model’s accuracy and macro-F1 scores were comparable between exams with fully adequate image quality and those with partially suboptimal quality. This suggests the AI can work reliably with the imperfect images commonly encountered in daily practice.

Implications for Scalable, Consistent Heart Care

For the field of cardiovascular health and longevity, consistent, early, and accurate diagnosis is a cornerstone of effective intervention. This DL approach offers a tool that could make expert-level morphological assessment more accessible.

“This approach may complement quantitative automation and expert visual assessment, supporting more consistent and scalable MV evaluation in routine echocardiographic practice,” the authors write. It could assist less-experienced sonographers or cardiologists, serve as a quality check, or help triage cases in areas with limited specialist access. In complex cases where multiple issues were present, the model identified at least one of the expert-confirmed etiologies 85.7% of the time in a post-hoc analysis.

Important Limitations and the Preprint Status

This study is a preprint and has not undergone peer review. Its findings require independent validation and may be revised. Several limitations are noted. The model was trained and tested on data from South Korea, and its performance in populations with different genetic backgrounds and disease prevalence patterns is unknown. The “functional” etiology category had fewer cases, which can affect model learning. Furthermore, the AI classifies etiologies but does not specify the exact anatomical location of abnormalities, such as which segment is prolapsing.

If validated, this technology would not replace cardiologists but act as a diagnostic support system. Integrating such a tool into clinical decision-making pathways would require careful study to ensure it improves patient outcomes without introducing unintended bias or over-reliance.

Source:
Deep Learning-Based Multiclass Classification of Mitral Valve Etiologies Using Limited B-Mode and Color Doppler Echocardiography: Internal and External Validation (medRxiv preprint, 2026-04-23)