Spatial risks of Orthoebolavirus spillover vary based on outbreak type.

OBJECTIVES: We develop and test a risk map for Orthoebolaviruses which are emerging infectious pathogens primarily concentrated in sub-Saharan Africa. The accuracy of predictive models and risk maps has been limited thus far by uncertainty in mechanisms underlying spread, low number of known outbreaks, and in how well various drivers predict different types of outbreaks (e.g. human vs epizootic outbreaks, and outbreaks of different viral species).

METHODS: Here, we explore frugivory and other factors as mechanisms of Orthoebolavirus spread and demonstrate statistical methods with repeated cross-validation that can be used even with very small data sets to explore how different factors influence different classes of events using ensemble machine learning logistic regression.

RESULTS: We show that covariates predicting outbreaks with the highest discrimination power are frugivore richness (area under curve [AUC] = 0.95) and fruit tree (Ficus) habitat suitability (AUC = 0.94). We found that Ficus distributions contributed to predictions of past Orthoebolavirus outbreaks relatively equally, regardless of type, based on feature contributions estimated using Shapley value calculations. In contrast, frugivore richness was a better contributor of predictions of epizootic than human outbreaks. Hunting activity was a poor predictor overall (AUC = 0.85) but contributed to some Sudan outbreaks predictions.

CONCLUSIONS: Our results suggest that different drivers best influence different classes of Orthoebolavirus outbreaks and models taking into account a variety of factors are needed to predict future spillover events.