Using mosquito and arbovirus data to computationally predict West Nile virus in unsampled areas of the Northeast United States.

Predicting and projecting risk of West Nile virus (WNV) to humans in areas without mosquito surveillance data is a key limitation of many WNV surveillance programs. To better inform risk of WNV, we analyzed 20 years (2001-2020) of point-level mosquito surveillance data from Connecticut (CT), United States, using machine learning methods to determine the most informative weather variables and land cover classes associated with monthly collections and WNV detections in mosquitoes. All training models were assessed based on explained deviance, root mean square error, and parsimony of included variables then optimized using a backward selection process. We used these training models to create a predictive mapping framework that could spatially extrapolate the monthly risk of WNV activity in mosquitoes across the entirety of the Northeast United States (CT, Maine, Massachusetts, New Hampshire, New Jersey, New York, Rhode Island, and Vermont) at a 4 × 4 km resolution. We then validated WNV detection probabilities against observed human cases at the town level in CT and the county level for northeastern states using generalized linear (mixed effects) models. Our predicted town- and county-level WNV detection probabilities in mosquitoes were significantly associated with the odds of a human case occurring within the town and/or county. This methodology increases the utility of point-source mosquito surveillance data by creating a flexible workflow for predicting risk of WNV to humans across the Northeast United States using easily accessible online data sources.