Interpretable AI for echocardiography-based early screening of rheumatic heart disease in Ethiopia: a cost-effective strategy to tackle a persistent global health disparity

Rheumatic heart disease (RHD) is a serious condition that develops from untreated streptococcal throat infections and causes long-term damage to the heart valves. Although it is become rare in wealthier countries, it still affects many people in low- and middle-income countries (LMICs), like sub-Saharan Africa, where fragmented healthcare infrastructure and limited access to specialized diagnostics exacerbate diagnostic delays. This study presents an interpretable artificial intelligence (AI) framework designed to automate early RHD screening using echocardiographic imaging, with a focus on addressing critical diagnostic gaps in resource-constrained settings such as Ethiopia. Our study explores how Interpretable AI based approach can help identify early signs of RHD using echocardiographic images, primarily mitral regurgitation (MR). Specifically, we focus on detecting mitral regurgitation in two views (A​pical​ ​five chamber (AP5) and Parasternal long axis (PSLAX) views) during the heart’s systolic cycle and measuring the thickness of the anterior mitral valve leaflet (AMVL) from the PSLAX view as part of the old criteria. To do this, we trained and compared two AI models to segment the left atrium (LA) and mitral valve (MV) using annotated echocardiographic datasets. By tracking vol​ume changes in the segmented LA, we identified the frames corresponding to the systolic cycle. To detect MR, which is a key pathological hallmark of RHD, we then overlaid Color Doppler blood flow information on the segmented LA regions of the systolic frames. Additionally, the MV segmentation model was used to measure the AMVL thickness during systole, a critical morphological indicator of chronic valvular damage which was not part of our primary focus. Importantly, this AI framework was developed primarily for the detection of World Heart Federation (WHF) significant MR. While AMVL thickness was examined as an additional morphological parameter, its measurement was performed during sys​tole for methodological convenience and has not been validated for the assessment of mitral stenosis. Furthermore, the framework does not evaluate restricted mitral leaflet motion. Therefore, leaflet thickness measurements should be regarded as exploratory findings and not as diagnostic outputs of the proposed system. Trained on annotated echocardiographic datasets and tested on an independent pilot test set of real-world scans from Ethiopian regional hospitals (patients not seen during training or validation), the AI model demonstrated promis​ing results, achieving precision (1.0), recall (0.7), F1-score (0.82), and accuracy (0.7) metrics that underscore its potential for reliable deployment in low-resource environments. The system’s interpretable design, which visualizes segmentation boundaries, Doppler flow patterns, and anatomical measurements, prioritizes clinical transparency, enabling healthcare workers to validate AI outputs against familiar diagnostic benchmarks. By automating labor-intensive echocardiographic analyses, this AI-driven strat​egy could revolutionize early RHD detection in LMICs, reducing reliance on scarce specialist expertise and paving the way for scalable, cost-effective screening programs. Our findings highlight the transformative role of context-specific AI solutions in bridging global health inequities.