Spatio-temporal spillover risk of yellow fever in Brazil

Abstract

Background: Yellow fever virus is a mosquito-borne flavivirus that persists in an enzoonotic cycle in non-human primates (NHPs) in Brazil, causing disease in humans through spillover events. Yellow fever (YF) re-emerged in the early 2000s, spreading from the Amazon River basin towards the previously considered low-risk, southeastern region of the country. Previous methods mapping YF spillover risk do not incorporate the temporal dynamics and ecological context of the disease, and are therefore unable to predict seasonality in spatial risk across Brazil. We present the results of a bagged logistic regression predicting the propensity for YF spillover per municipality (administrative sub-district) in Brazil from environmental and demographic covariates aggregated by month. Ecological context was incorporated by creating National and Regional models of spillover dynamics, where the Regional model consisted of two separate models determined by the regions' NHP reservoir species richness (high vs low).

Results: Of the 5560 municipalities, 82 reported YF cases from 2001 to 2013. Model accuracy was high for the National and low reservoir richness (LRR) models (AUC = 0.80), while the high reservoir richness (HRR) model accuracy was lower (AUC = 0.63). The National model predicted consistently high spillover risk in the Amazon, while the Regional model predicted strong seasonality in spillover risk. Within the Regional model, seasonality of spillover risk in the HRR region was asynchronous to the LRR region. However, the observed seasonality of spillover risk in the LRR Regional model mirrored the national model predictions.

Conclusions: The predicted risk of YF spillover varies with space and time. Seasonal trends differ between regions indicating, at times, spillover risk can be higher in the urban coastal regions than the Amazon River basin which is counterintuitive based on current YF risk maps. Understanding the spatio-temporal patterns of YF spillover risk could better inform allocation of public health services.

Keywords: Arboviruses; Brazil; Risk mapping; Spatial epidemiology; Vectors; Yellow fever.

Fig. 1

Conceptual diagram of modeling methods. The dataset was aggregated by month and municipality (top panel) before being split into training (70%) and withheld testing (30%) datasets. Models were fit to 500 data subsamples, which consisted of 10 spillover events and 100 background observations (lower panel). The bagged logistic model predictions are the average of subsampled dataset models. Spatial dependence was not considered in the model

Fig. 2

Distribution of NHP species richness by municipality. Plot of distribution of non-human primate species richness per municipality, colored by the break used to determine areas of high reservoir richness (purple) and low reservoir richness (orange). Inset is a map of the two regions

Fig. 3

Predicted spatial risk of yellow fever spillover. Propensity of yellow fever spillover in January, June, and September of 2008. Raw outputs of the model for each municipality-month are rank-ordered to allow for comparison across models. Results from the National model are on the top row and the Regional model are on the bottom row. Black outline represents the split between HRR (northwest) and LRR (southeast) regions. The outline in the national model is for reference only. See supplemental video for entire time series. Map projection: SAD69 Brazil Polyconic. Data source: 2001 municipality boundaries, Brazilian Institute of Geography and Statistics

Fig. 4

Variation of yellow fever spillover intensity in space and time. Plots of variance of the predicted spillover intensity throughout the 13-year time series from the National model (a), low reservoir richness Regional model (b), and high reservoir richness Regional model (c). Darker municipalities are predicted to have greater seasonality in spillover risk than lighter municipalities. The seasonal pattern in model predictions are shown by monthly averages of predicted spillover intensity across the entire study area of Brazil for the National model (d), within the low reservoir richness Regional model (e), and within high reservoir richness Regional model (f). Gray lines represent an individual year of data with overall mean in black. Rug along x-axis represents true spillover events, with larger and darker shapes representing more spillover events during that calendar month. Map projection: SAD69 Brazil Polyconic. Data source: 2001 municipality boundaries, Brazilian Institute of Geography and Statistics

Fig. 5

Rank order of median variable importance. The median variable importance was calculated for the National, low reservoir richness (LRR), and high reservoir richness (HRR) Regional models based on 100 permutations per variable within a model. The variables were ranked from most important (1) to least important (12) for model accuracy

Fig. 6

Variable importance for the National and Regional model. The median variable importance was calculated for the National, low reservoir richness (LRR), and high reservoir richness (HRR) Regional models based on 100 permutations per variable within a model. Values are the decline in AUC (∆AUC) due to permutation from the original model scaled to the largest ∆AUC within each model