Interpretable machine learning and signal processing for automated reading and quality control of lateral flow tests for schistosomiasis

Abstract

Abstract There is a lack of automated pipelines for diagnostic classification of point-of-care tests for neglected tropical diseases. Here, we present an end-to-end automated pipeline for the analysis of point-of-care circulating cathodic antigen tests for schistosomiasis. We incorporated deep learning for cassette segmentation with signal processing. Automated classifications were compared to quantitative readings from calibrated antigen samples examined in lateral flow readers and visual readings from highly trained field and senior technicians. The pipeline was evaluated for 3188 individuals within the SchistoTrack cohort in rural Uganda. Our quantitative classifications were on par with a lateral flow reader and showed 86.6% sensitivity and 96.5% specificity with visual readings from a senior technician, which was an improvement on the visual readings from field technicians. Automated classifications were possible in as little as five min after test preparation for high antigen concentrations. We showed that visual trace uncertainty can be resolved with signal processing, indicating that visual traces should be classified as negative. Our pipeline will aid in advancing diagnostics to meet the World Health Organization target product profile for schistosomiasis, provide quantitative assessments for other diagnostics, enable large-scale surveillance in areas targeting elimination and provide real-time quality control for diagnostics introduced into primary healthcare facilities.