A global model of avian influenza prediction in wild birds: the importance of northern regions

Abstract

Avian influenza virus (AIV) is enzootic to wild birds, which are its natural reservoir. The virus exhibits a large degree of genetic diversity and most of the isolated strains are of low pathogenicity to poultry. Although AIV is nearly ubiquitous in wild bird populations, highly pathogenic H5N1 subtypes in poultry have been the focus of most modeling efforts. To better understand viral ecology of AIV, a predictive model should 1) include wild birds, 2) include all isolated subtypes, and 3) cover the host's natural range, unbounded by artificial country borders. As of this writing, there are few large-scale predictive models of AIV in wild birds. We used the Random Forests algorithm, an ensemble data-mining machine-learning method, to develop a global-scale predictive map of AIV, identify important predictors, and describe the environmental niche of AIV in wild bird populations. The model has an accuracy of 0.79 and identified northern areas as having the highest relative predicted risk of outbreak. The primary niche was described as regions of low annual rainfall and low temperatures. This study is the first global-scale model of low-pathogenicity avian influenza in wild birds and underscores the importance of largely unstudied northern regions in the persistence of AIV.

Figure 1

Density plots for important variables. Density plots for the variables with the five highest variable importance scores as calculated by Random Forests in the accuracy of the predictive model of avian influenza in wild birds (A-E). The density, or the likelihood of a variable to take on a value, of AIV-positive samples for each variable is represented by red dotted lines, the AIV-negative by black solid lines.

Figure 2

Partial dependence plots for important variables. Partial dependence plots for the variables with the five highest variable importance scores as calculated by Random Forests in the accuracy of the predictive model of avian influenza in wild birds (A-E). Plots show the partial dependence of a high Relative Occurrence Index value for avian influenza on each predictor variable.

Figure 3

Global map of the predicted relative occurrence of avian influenza virus (AIV) in wild birds. The predictive model was constructed using the Random Forests algorithm on 41 predictor variables. The dots on the map represent all samples in both the testing and training databases. Locations where one or more AIV-positive samples were collected are shown as black dots; locations where no positive samples were collected are marked with white. A single dot may represent multiple samples taken at that location. This map is presented in Robinson (sphere) projection, central meridian 145°.