The potential distribution of Bacillus anthracis suitability across Uganda using INLA

Abstract

To reduce the veterinary, public health, environmental, and economic burden associated with anthrax outbreaks, it is vital to identify the spatial distribution of areas suitable for Bacillus anthracis, the causative agent of the disease. Bayesian approaches have previously been applied to estimate uncertainty around detected areas of B. anthracis suitability. However, conventional simulation-based techniques are often computationally demanding. To solve this computational problem, we use Integrated Nested Laplace Approximation (INLA) which can adjust for spatially structured random effects, to predict the suitability of B. anthracis across Uganda. We apply a Generalized Additive Model (GAM) within the INLA Bayesian framework to quantify the relationships between B. anthracis occurrence and the environment. We consolidate a national database of wildlife, livestock, and human anthrax case records across Uganda built across multiple sectors bridging human and animal partners using a One Health approach. The INLA framework successfully identified known areas of species suitability in Uganda, as well as suggested unknown hotspots across Northern, Eastern, and Central Uganda, which have not been previously identified by other niche models. The major risk factors for B. anthracis suitability were proximity to water bodies (0-0.3 km), increasing soil calcium (between 10 and 25 cmolc/kg), and elevation of 140-190 m. The sensitivity of the final model against the withheld evaluation dataset was 90% (181 out of 202 = 89.6%; rounded up to 90%). The prediction maps generated using this model can guide future anthrax prevention and surveillance plans by the relevant stakeholders in Uganda.

Figure 1

Distribution of anthrax presence and pseudo-absence locations across Uganda. The navy-blue circles show wildlife cases (n = 294) used for model training, the blue triangles show livestock cases (n = 171), and the red diamonds represent human cases (n = 32). The blue polygons show the locations of the 50 km, 75 km, and 100 km buffers which were constructed around the wildlife cases with a distance of 10 km between the buffers and the presence locations. The pink dots show the pseudo-absence points selected within the 50 km buffer, the orange dots show the pseudo-absence points selected within the 75 km buffer, and the white dots show the pseudo-absence points selected within the 100 km buffer. Prediction maps were developed using the Quantum Geographic Information System software (QGIS). URL: https://qgis.osgeo.org (2020).

Figure 2

Results of correlation between covariates using Pearson’s correlation test. Correlation between covariates was shown by red numbers (negative correlation) and blue numbers (positive correlation). BIO1 = Annual Mean Temperature, BIO2 = Mean Diurnal Range, BIO3 = Isothermality, BIO4 = Temperature Seasonality, BIO5 = Max Temperature of Warmest Month, BIO6 = Min Temperature of Coldest Month, BIO7 = Temperature Annual Range, BIO8 = Mean Temperature of Wettest Quarter, BIO9 = Mean Temperature of Driest Quarter, BIO10 = Mean Temperature of Warmest Quarter, BIO11 = Mean Temperature of Coldest Quarter, BIO12 = Annual Precipitation, BIO13 = Precipitation of Wettest Month, BIO14 = Precipitation of Driest Month, BIO15 = Precipitation Seasonality, BIO16 = Precipitation of Wettest Quarter, BIO17 = Precipitation of Driest Quarter, BIO18 = Precipitation of Warmest Quarter, BIO19 = Precipitation of Coldest Quarter.

Figure 3

(A) The fixed effect and credible intervals for the linear covariates, (B) the smoothed fits of elevation (in m in x-axis) and linear fits of: (C) distance to water (in km in x-axis), (D) soil calcium (in cmolc/kg in x-axis), (E) soil water (in v% in x-axis) variables. The y-axis shows the estimated probability of presence. The shaded grey polygons represent 95% credible intervals.

Figure 4

(A) The posterior predictive mean (B) posterior standard deviation (C) posterior lower credible interval (2.5% quantile) and (D) posterior credible upper interval (97.5% quantile) of the probability of B. anthracis suitability across Uganda from wildlife data (2004–2010). Prediction maps were developed using the Quantum Geographic Information System software (QGIS). URL: https://qgis.osgeo.org (2020).

Figure 5

A binary map of the posterior predictive mean of the probability of B. anthracis suitability across Uganda using the threshold for maximum sensitivity and specificity (0.52). Prediction maps were developed using the Quantum Geographic Information System software (QGIS). URL: https://qgis.osgeo.org (2020).