Abstract 4138955: Artificial intelligence-guided screening of rheumatic heart disease from single-view two-dimensional echocardiography

Introduction: Rheumatic heart disease (RHD) is the most common acquired heart disorder in children and adolescents worldwide. We developed and validated an automated artificial intelligence (AI)-guided RHD screening algorithm adapted for point-of-care ultrasonography (POCUS) in school-aged children. Methods: We employed a cross-domain transfer learning approach, in which a 3D convolutional neural network (CNN) was first trained to detect structural RHD deformation of the mitral or aortic valves in 244,523 videos, representing all views from 5,614 adult transthoracic echocardiograms (1:5 age and sex-matched cases and controls; median age 69 [58-80] years, 76.4% female) in a large US health system. The model was fine-tuned for stage ≥B (“definite”) RHD in 21,472 POCUS videos (2D parasternal and apical acquisitions) from 5,525 studies (75% training, 25% validation) in a pediatric screening program (median age 11 [IQR 10-13] years, 54.6% female) in Brazilian low-income schools. Testing was performed in a held-out set of 1,966 parasternal long-axis (PLAX) videos from 1,138 studies in Brazil (14 [1.2%] with stage ≥B RHD) as well as in an external pediatric screening set in Uganda consisting of 249 videos from 96 studies (34 [35.4%] with stage ≥B RHD) ( Fig. 1 ). Results: Our model ( Fig. 2 ) achieved a study-level AUROC (area under the receiver operating characteristic curve) of 0.88 across the held-out/external testing sets for identifying stage ≥B RHD from cardiac POCUS ( Fig. 3A ). On a video-level the model learned a continuous spectrum of phenotypes on PLAX acquisitions spanning stage ≥B (“definite”) and stage A (“borderline”) cases, ranging from a median video-level AI probability of 0.13 [0.01-0.73] for stage ≥B to 0.00 [0.00-0.01] for non-RHD POCUS ( Fig. 3B ). At the threshold that maximized Youden’s J in the held-out Brazil set, our algorithm’s performance in the set from Uganda showed 97% recall (sensitivity), a positive predictive value (precision) of 46%, and a negative predictive value of 95%. Conclusions: A transfer learning approach that employs multi-view learning achieves excellent performance for RHD on single-view two-dimensional cardiac POCUS without Doppler. Our study suggests a scalable approach to AI-enabled RHD detection with images that can be acquired by individuals with modest training.