Predicting cutaneous and visceral leishmaniasis outbreak occurrence in Brazil: A machine learning model of forest cover, climate, and land use

Cutaneous and visceral leishmaniasis (CL and VL) are neglected vector-borne diseases of growing concern in Brazil, where deforestation, reforestation, and urban expansion are reshaping vector habitats and disease risk. While previous research has identified forest cover as a primary environmental predictor of CL, the effects of dynamic forest variables—particularly reforestation—remain under explored. Our analyses aim to address the central question: whether environmental factors can help predict occurrence of leishmaniasis in Brazil. Here, we apply machine learning-based models to test how climate, land-use, and forest change variables influence outbreak occurrence of CL and VL across Brazil’s Amazon from 2001 to 2018. Using Extreme Gradient Boosting (XGBoost) with Bayesian optimization and spatial subsetting, we leveraged over a decade of municipality-level case data and high-resolution geospatial covariates to assess variable importance and characterize nonlinear relationships via partial dependence plots. As for the precision of our CL model, we found that out-of-sample area-under-the-curve (AUC) was 0.770, and for VL, our model’s out of sample AUC was 0.729. We found that, in line with previous ecological work, forest cover was a strong predictor of CL outbreak occurrence but weak for VL. Contrary to expectations, deforestation was positively associated with CL risk, while reforestation was negatively associated—suggesting that habitat disruption may increase CL transmission through human-vector contact. For VL, forest-related variables exhibited hump-shaped or declining effects, highlighting complex, potentially nonlinear interactions between land use and synanthropic vector ecology. Temperature, urban cover, and population also emerged as important predictors for both diseases. Our results support the ecological relevance of land-use change in shaping leishmaniasis risk and introduce reforestation as a novel covariate with differential predictive effects across disease types. These findings underscore the value of machine learning in disentangling high-dimensional, nonlinear disease-environment relationships and call for further integration of ecological, vector, and host data in predictive modeling of leishmaniasis under ongoing landscape transformation.