Deep Learning for Echo Analysis, Tracking, and Evaluation of Mitral Regurgitation (DELINEATE-MR)

• DOI: 10.1161/circulationaha.124.068996
• Publication Date: 2024-06-17T09:01:40Z
• Authors: Aaron Long; Christopher M. Ha...; Joshua Finer; Dustin Hartzel; Linyuan Jing; Azadeh Keivani; Christopher Kelsey; Daniel Rocha; Jeffrey Ruhl; David vanMaanen; Gil Metser; Eamon Duffy; Thomas Mawson; Mathew Maurer; Andrew J. Einstein; Ashley Beecy; Deepa Kumaraiah; Shunichi Homma; Qi Liu; Vratika Agarwal; Mark Lebehn; Martin Leon; Rebecca Hahn; Pierre Elias; Timothy J. Poterucha
• Keywords: Echo (communications protocol); Functional mitral regurgitation

BACKGROUND: Artificial intelligence, particularly deep learning (DL), has immense potential to improve the interpretation of transthoracic echocardiography (TTE). Mitral regurgitation (MR) is most common valvular heart disease and presents unique challenges for DL, including integration multiple video-level assessments into a final study-level classification. METHODS: A novel DL system was developed intake complete TTEs, identify color MR Doppler videos, determine severity on 4-step ordinal scale (none/trace, mild, moderate, severe) using reading cardiologist as reference standard. This tested in internal external test sets with performance assessed by agreement cardiologist, weighted κ, area under receiver-operating characteristic curve binary classification both moderate or greater severe MR. In addition primary model, 6-step assessment model studied intermediate classes mild-moderate moderate-severe exact ±1 step clinical interpretation. RESULTS: total 61 689 TTEs were split train (n=43 811), validation (n=8891), (n=8987) an additional set 8208 TTEs. The had high (exact accuracy, 82%; κ=0.84; curve, 0.98 MR) 79%; κ=0.80; MR). Most (63% 66% external) misclassification disagreements between none/trace mild accuracy slightly higher TTE views (accuracy, 82%) than only apical 4-chamber 80%). subset analyses, accurate secondary lower cases eccentric analysis system, 80% 76% 99% 98% set, respectively. CONCLUSIONS: end-to-end can entire echocardiogram studies accurately classify may be useful helping clinicians refine assessments.