Evaluating the contribution of weather variables to machine learning forecasts of visceral leishmaniasis in Brazil

Visceral leishmaniasis (VL), a deadly neglected tropical disease, remains a persistent public health challenge in Brazil, where transmission is shaped by interacting climatic, environmental, and sociodemographic factors. Despite evidence that weather conditions influence VL dynamics, they remain underutilized for outbreak prediction. This study evaluates whether climate-informed machine learning can support early warnings for VL in Brazil. We developed machine learning models to forecast monthly VL case counts and classify outbreak risk using data from 2007 to 2024 across 113 Brazilian municipalities. A cutting-edge sliding window approach enabled models to capture both short- and long-term trends using lagged meteorological data combined with land-use and sociodemographic variables. Risk classification models were developed for a subset of 22 municipalities following the Brazilian Ministry of Health's prioritization framework to enable direct policy alignment. Predictive performance and variable importance were evaluated across locations. Weather patterns and indicators of human land-use pressure consistently ranked among the strongest predictors of VL risk. However, the relative importance of predictors varied across municipalities, reflecting local differences in transmission dynamics. Overall, forecasting models successfully captured long-term trends in observed case counts, and risk classification models, offering particularly timely and actionable signals for targeted intervention, achieved area under the curve scores above 0.80 in 86% of municipalities. Weather-informed machine learning models can provide timely, locally tailored predictions of VL risk in Brazil. As weather variability intensifies, integrating environmental data into existing surveillance systems may improve preparedness and reduce disease burden in vulnerable communities.