Integrated Forecasts Based on Public Health Surveillance and Meteorological Data Predict West Nile Virus in a High-Risk Region of North America

Abstract

BACKGROUND: West Nile virus (WNV), a global arbovirus, is the most prevalent mosquito-transmitted infection in the United States. Forecasts of WNV risk during the upcoming transmission season could provide the basis for targeted mosquito control and disease prevention efforts. We devel-oped the Arbovirus Mapping and Prediction (ArboMAP) WNV forecasting system and used it in South Dakota from 2016 to 2019. This study reports a post hoc forecast validation and model comparison. OBJECTIVES: Our objective was to validate historical predictions of WNV cases with independent data that were not used for model calibration. We tested the hypothesis that predictive models based on mosquito surveillance data combined with meteorological variables were more accurate than models based on mosquito or meteorological data alone. METHODS: The ArboMAP system incorporated models that predicted the weekly probability of observing one or more human WNV cases in each county. We compared alternative models with different predictors including a) a baseline model based only on historical WNV cases, b) mosquito models based on seasonal patterns of infection rates, c) environmental models based on lagged meteorological variables, including temperature and vapor pressure deficit, d) combined models with mosquito infection rates and lagged meteorological variables, and e) ensembles of two or more combined models. During the WNV season, models were calibrated using data from previous years and weekly predictions were made using data from the current year. Forecasts were compared with observed cases to calculate the area under the receiver operating characteristic curve (AUC) and other metrics of spatial and temporal prediction error. RESULTS: Mosquito and environmental models outperformed the baseline model that included county-level averages and seasonal trends of WNV cases. Combined models were more accurate than models based only on meteorological or mosquito infection variables. The most accurate model was a simple ensemble mean of the two best combined models. Forecast accuracy increased rapidly from early June through early July and was stable thereafter, with a maximum AUC of 0.85. The model predictions captured the seasonal pattern of WNV as well as year-to-year variation in case numbers and the geographic pattern of cases. DISCUSSION: The predictions reached maximum accuracy early enough in the WNV season to allow public health responses before the peak of human cases in August. This early warning is necessary because other indicators of WNV risk, including early reports of human cases and mosquito abun-dance, are poor predictors of case numbers later in the season.