A Deep Learning Model to Identify Mitral Valve Prolapse From the Echocardiogram.

BACKGROUND

Mitral valve prolapse (MVP) has a prevalence of 2% to 3% and increases risk of heart failure and sudden death, but diagnosis by transthoracic echocardiography requires time and expertise.

OBJECTIVES

This study aims to develop a deep learning model DROID-MVP (Dimensional Reconstruction of Imaging Data-Mitral Valve Prolapse) to classify MVP from digital echocardiogram videos.

METHODS

DROID-MVP was trained and validated using 1,043,893 echocardiogram videos (48,829 studies) from 16,902 cardiology patients at MGH (Massachusetts General Hospital), and externally validated in 8,888 MGH primary care patients and 257 primary care patients at BWH (Brigham and Women's Hospital). The authors tested associations among DROID-MVP predictions (range: 0-1), mitral regurgitation (MR) severity, and mitral valve repair or replacement (MVR).

RESULTS

Of 16,902 patients (6,391 [38%] women; age 61 ± 16 years) in the derivation sample, 783 (4.6%) had MVP. DROID-MVP accurately identified MVP across the MGH cardiology internal validation set (area under the receiver-operating characteristic curve [AUROC]: 0.947 [95% CI: 0.910-0.984]; average precision [AP]: 0.682 [95% CI: 0.565-0.784]; prevalence: 0.036), MGH primary care external validation set (AUROC: 0.964 [95% CI: 0.951-0.977]; AP: 0.651 [95% CI: 0.578-0.716]; prevalence: 0.022), and BWH primary care external validation set (AUROC: 0.968 [95% CI: 0.946-0.989]; AP: 0.774 [95% CI: 0.666-0.797]; prevalence: 0.113). A high (>0.67) vs low (<0.33) DROID-MVP score was associated with moderate or severe MR (adjusted OR: 2.0 [95% CI: 1.1-3.8]; P = 0.030) and future MVR (adjusted HR: 3.7 [95% CI: 1.5-8.9]; P = 0.004).

CONCLUSIONS

A deep learning model identifies MVP from echocardiogram videos, and model predictions are associated with clinical endpoints including MR and future MVR. Deep learning can automate MVP diagnosis and potentially generate digital markers of clinically significant MVP.