Explainable AI Unravels Local Factors Driving Extraordinary Outbreaks of West Nile Virus in Europe

Year-to-year emergence of West Nile virus has been sporadic and notoriously hard to predict. In Europe, 2018 saw a dramatic increase in the number of cases and locations affected. In this work, we demonstrate a novel method for predicting outbreaks and understanding what drives them. This method creates a simple model for each region that directly explains how each variable affects risk. Behind the scenes, each local explanation model is produced by a state-of-the-art AI engine. This engine unpacks and restructures output from an XGBoost machine learning ensemble. XGBoost, well-known for its predictive accuracy, has always been considered a "black box" system. Not any more. With only minimal data curation and no "tuning", our model predicted where the 2018 outbreak would occur with an AUC of 97%. This model was trained using data from 2010-2016 that reflected many domains of knowledge. Climate, sociodemographic, economic, and biodiversity data were all included. Our model furthermore explained the specific drivers of the 2018 outbreak for each affected region. These effect predictions were found to be consistent with the research literature in terms of priority, direction, magnitude, and size of effect. Aggregation and statistical analysis of local effects revealed strong cross-scale interactions. From this, we concluded that the 2018 outbreak was driven by large-scale climatic anomalies enhancing the local effect of mosquito vectors. We also identified substantial areas across Europe at risk for sudden outbreak, similar to that experienced in 2018. Taken as a whole, these findings highlight the role of climate in the emergence and transmission of West Nile virus. Furthermore, they demonstrate the crucial role that the emerging "eXplainable AI" (XAI) paradigm will have in predicting and controlling disease. HighlightsO_LIThis study shows that the extraordinary 2018 West Nile virus outbreak in Europe was likely due to cross-scale effects between large climatic systems and local mosquito vector populations C_LIO_LIWe found that large areas in Europe are similarly vulnerable to large and sudden outbreaks C_LIO_LIThese findings were powered by a novel AI-driven engine for deriving locally precise models; this explanatory engine was supported by a high-performance XGBoost model (97% AUC). C_LIO_LIAI-driven local models allow for high-power statistical analyses, including: hypothesis testing,, standardized effect size calculation, multivariate clustering, and tertiary inferential modeling C_LI