AI-supported versus manual microscopy of Kato-Katz smears for diagnosis of soil-transmitted helminth infections in a primary healthcare setting
Joar von Bahr, Antti Suutala, Hakan Kücükel, Harrison Kaingu, Felix Kinyua, Martin Muinde, Kevan Osundwa, Wigina Ronald, Jackson Muinde, Billy Ngasala, Mikael Lundin, Andreas Mårtensson, Nina Linder, Johan Lundin
Abstract
Soil-transmitted helminths primarily comprise Ascaris lumbricoides, Trichuris trichiura, and hookworms, infecting more than 600 million people globally, particularly in underserved communities. Manual microscopy of Kato-Katz thick smears is a widely used diagnostic method in monitoring and control programs, but is time-consuming, requires on-site experts and has low sensitivity, especially for light intensity infections. In this study, portable whole-slide scanners and deep learning-based artificial intelligence (AI) were deployed in a primary healthcare setting in Kenya. Stool samples (n = 965) were collected from school children and Kato-Katz thick smears were digitized for AI-based detection. Light-intensity infections accounted for 96.7% of cases. Three diagnostic methods - manual microscopy, autonomous AI and human expert-verified AI - were compared to a composite reference standard, which combined expert-verified helminth eggs in physical and digital smears. Sensitivity for A. lumbricoides, T. trichiura and hookworms was 50.0%, 31.2%, and 77.8% for manual microscopy; 50.0%, 84.4%, and 87.4% for the autonomous AI; and 100%, 93.8%, and 92.2% for expert-verified AI in smears suitable for analysis (n = 704). Specificity exceeded 97% across all methods. The expert-verified AI had higher sensitivity than the other methods while maintaining high specificity for the detection of soil-transmitted helminths in Kato-Katz thick smears, especially in light-intensity infections.