Yellow fever (YF), a mosquito-borne flavivivirus, re-emerged in Brazil in the early 2000s, and is spreading southeast from the Amazon. Current YF risk maps have aggregated environmental conditions over time muting any differences in seasonality across different Brazilian biomes, and do not recommend vaccination in the densely populated coastal region. Here we present a bagged logistic regression model that accounts for temporal heterogeneities to predict the spatio-temporal propensity of YF spillover in Brazil by municipality (sub-state administrative unit) based on monthly average environmental and socio-demographic covariates. We then allowed for differences in the spillover process between two contiguous regions split by high (>5) and low non-human primate (NHP) reservoir richness. The regions were created based on a natural break in the distribution of NHPs. Models were built using 70% of the dataset, stratified by space and time. The model output was the average of 500 model iterations of the bagged data. The withheld 30% of the dataset was used to assess model performance via AUC. Finally, we compared the predictions for the high and low NHP richness areas (HR and LR, respectively) of the regional model to the national model.
Results/Conclusions
Only 77 of the total 5,560 municipalities reported YF cases from 2001 to 2013. Municipalities reporting cases in multiple months were extremely rare. Model accuracy was high (0.78, 0.85, and 0.78 AUC for the national, HR, and LR models, respectively). Variable importance by permutation ranked population density and NHP richness highly in all models. The remaining variables changed rank between models suggesting spillover process changes by region. Modeling spillover risk in by region led to different spatio-temporal patterns in predictions. The largest change is in the Amazon, which shifts from a consistent high risk area in the national model to cycling between high and low risk through the year. This cycling is similar to the dynamics in the LR region, and national model. However, different seasonality of risk between the regions of the regional model implies that, at times, spillover risk can be higher in the urban coastal regions then the Amazon river basin which is counterintuitive based on current YF risk and vaccine recommendation maps. Understanding the spatiotemporal risk of YF in Brazil can better inform vaccination strategies and resource allocation, such as prioritizing urban and rural areas at different times of the year.