Abstract 4362373: Strain and Deep Learning-Derived Mitral Regurgitation Features Are Independently Associated With Rheumatic Heart Disease - an analysis from the GOAL Trial

Background: Rheumatic heart disease (RHD) affects over 50 million people globally. Identifying latent RHD in asymptomatic children by echocardiography (echo) enables early secondary prophylaxis. The GOAL (Gwoko Adunu pa Lutino) trial demonstrated that monthly penicillin prevents disease progression in Ugandan children with screen-detected RHD. Recently a deep learning (DL) model trained on pediatric echo from GOAL with expert-adjudicated labels achieved strong performance compared to expert review (1). However, the performance of DL derived features against other emerging metrics such as left ventricular/left atrial strain has not been assessed. This study determined whether strain could provide additive value to assessment of RHD. Methods: Retrospective cohort analysis of completion studies from 435 children in GOAL trial (<20 years of age). RHD status (latent or worse) was identified by expert clinicians. Mitral regurgitation (MR) features, including jet length, velocity and duration were extracted using the previously described DL model which was trained on pediatric echo images from high-quality standard portable echo with expert-adjudicated labels . Left atrial (LA), left ventricular (LV), and right ventricular (RV) strain parameters were manually quantified from apical views using Philips Ultrasound Workspace. Univariable and multivariable logistic regression were used to evaluate associations with RHD. Results: Two-hundred and forty-five children (56%) were identified to have RHD at study completion. Univariable logistic regression (Table 1) identified RHD associations with both DL-derived MR features and manual strain metrics, including maximum MR jet length (OR: 3.71, p < 0.001), MR jet-to-LA atrium length ratio (OR: 2.87, p < 0.001), MR duration (OR: 2.64, p < 0.001), lower LV global longitudinal strain (OR: 0.79, p = 0.005), and lower LA global strain (OR: 0.79, p = 0.02). In multivariable stepwise regression (Table 1), three features remained independently associated: lower LV global longitudinal strain (OR: 0.76 [0.61–0.95], p = 0.016), greater MR jet length (OR: 3.06 [2.24–4.19], p < 0.001), and longer normalized MR jet duration (OR: 1.63 [1.20–2.22], p = 0.016). Conclusions: Strain and DL-based classification of MR are independently associated with RHD. This study supports the biological validity of DL-driven RHD detection, also suggesting that strain measurements provides additive value in the assessment of RHD.