Using Graph Neural Networks to Forecast West Nile Virus with Trap and Weather Data

We present a graph neural network model to generate short-term forecasts of the location of mosquito populations positive for West Nile virus in Illinois. We construct spatial input graphs by applying a k-nearest neighbors algorithm to geospatial data. In contrast to the majority of machine learning methods previously applied to similar problems, our model effectively accounts for the spatial dependence of the input data. This is reflected in our experiments for forecasts 1 to 7 weeks into the future, which show that our model outperforms state-of-the-art models such as XGBoost, fully-connected neural networks, and logistic regression. Curbing transmission of West Nile virus by controlling disease vectors, primarily Culex mosquitoes, remains challenging and resource intensive. Therefore, accurate forecasts of mosquito populations and disease have the potential to assist the targeting of these measures, and reduce the social and economic burdens of West Nile virus.