Evaluating West Nile Virus Forecasts in an Endemic Region of North America: A Retrospective Model Comparison and Validation

Abstract:

West Nile virus (WNV) remains a persistent public health hazard in many parts of the U.S., and there is need for knowledge of when and where WNV outbreaks will occur to target public health responses. To address this need, we developed the Arbovirus Monitoring and Prediction (ArboMAP) system. ArboMAP produces weekly, county-level forecasts of WNV risk using remotely sensed environmental data combined with entomological and surveillance data. We made prospective WNV forecasts in South Dakota, the U.S. state with the highest incidence of WNV, from 2016-2019.

ArboMAP uses a data-driven approach in which human WNV cases are modeled as a function of meteorological variables and mosquito infection status. Gridded environmental data, including air temperature, vapor pressure deficit, and precipitation, were derived from NASA’s North American Land Data Assimilation System. Mosquito surveillance was conducted by local trapping programs, and mosquitoes were pooled and tested weekly for WNV infection. We used distributed lags to model the delayed effects of environmental fluctuations, and a log-linear function to model growth of the infection rate in the early transmission season. ArboMAP is implemented using the R software environment for data processing, modeling, and reporting combined with a Google Earth Engine application for environmental data access.

The specific models used for forecasting have evolved over time as we have evaluated forecasting results. Here, we conducted a retrospective analysis of historical forecasts to determine the approaches that consistently provide the highest accuracy. We compared models based on different meteorological variables data transformation (raw data versus seasonal anomalies), underlying algorithms (cubic versus thin-plate splines), and model form (fixed versus seasonally varying environmental relationships). The best models were based on temperature and atmospheric moisture variables transformed into seasonal anomalies and accounted for seasonally varying environmental effects. Prediction accuracy was lowest at the beginning of the WNV season and increased as more data were incorporated into the forecasts. We were able to obtain accurate predictions of WNV outbreaks 1-2 months before they were captured by human case surveillance.