Predictive modeling based on machine learning for mapping risk areas of human sporotrichosis in southeastern Brazil

Abstract

Sporotrichosis, a zoonotic mycosis with a growing public health impact, requires innovative methods to map risk areas. This study applied machine learning techniques, Artificial Neural Networks (ANN), and Decision Trees (DT) to integrate sociodemographic, epidemiological, environmental, and urban data from Contagem, Minas Gerais, Brazil. Both models exhibited high predictive capacity, with complementary performances: the ANN stood out in class discrimination (mean accuracy of 0.9106, AUC of 0.939, MSE of 0.4040, RMSE of 0.6313, R² of 0.5955), while the DT demonstrated greater consistency and lower errors (mean accuracy of 0.9185, AUC of 0.9147, MSE of 0.0695, RMSE of 0.2625, R² of 0.6285). The DT also identified key risk factors, such as the presence of parks, squares, soccer fields, positive cats, health facilities, and suburban clusters. Spatial analysis reinforced the findings, with the comparative map showing high similarity between actual and predicted data: of the 884 census sectors, 221 (25 %) recorded positive human cases against 219 (24.78 %) predicted by the ANN. These results highlight the potential of the techniques used to optimize the monitoring and control of sporotrichosis, enriching the understanding of its epidemiology and providing robust instruments for developing more effective control strategies, promoting significant advances in public health and animal welfare.