Uncertainty-Driven Rheumatic Heart Disease Detection through Active Learning

In resource-scarce areas, rheumatic heart disease (RHD) poses a significant health challenge. We are developing AIdriven tools for RHD screening in resource-constrained settings. Integrating active learning (AL) with these tools can enhance performance and reduce the need for extensive data annotation and resources. We propose a novel automatic RHD detection framework that integrates AL and uncertainty assessment. Our method leverages a 3D residual network to identify relevant systolic frames in cardiac color Doppler ultrasound. We analyzed 1,022 echocardiograms from 511 Ugandan children. Testing 100 cases within this cohort yielded an accuracy of 0.85, sensitivity of 0.79, and specificity of 0.92 for detecting RHD. This accuracy is comparable to that of expert cardiologists. Remarkably, our AL framework achieved similar performance using only 43% of the data (p-value = 0.15), highlighting its efficiency in significantly saving annotation efforts, which are costly and unfeasible due to the specialized human skills required.