Forecasting human West Nile Virus cases using eco-climate and demographic surveillance data

Abstract

West Nile Virus (WNV) is the most common mosquito-borne arbovirus in the United States and continues to cause seasonal outbreaks that strain public health resources. Early forecasting of WNV cases could support proactive vector-control interventions, but operational forecasting is constrained by the fact that the only publicly available source of county-level human WNV case data, CDC ArboNET, reports cases annually rather than weekly. We hypothesized that warmer prior-year annual temperatures are associated with higher county-level human WNV case counts the following year, and tested this using machine learning across six geographically diverse U.S. counties (Boulder, CO; Cook, IL; Dallas, TX; Larimer, CO; Los Angeles, CA; Maricopa, AZ) from 2005 to 2024. Three machine learning models (linear regression, random forest, and gradient boosting with a Poisson loss objective) were compared against four baselines (global mean, per-county climatology, persistence, and three-year rolling mean) under leave-one-year-out cross-validation, with all features lagged by at least one year to prevent data leakage. Gradient Boosting achieved the lowest mean absolute error (MAE = 49.4 cases per year), a 13% improvement over the strongest baseline (per-county climatology, MAE = 56.6). However, RMSE remained substantially higher (175.2) and R² was near zero across all models, indicating that a small number of large outbreak years disproportionately affected performance. When disaggregated by outbreak status, Gradient Boosting was most accurate in non-outbreak years (MAE = 26) but least accurate in outbreak years (MAE = 960). Weather features, particularly growing degree days and maximum temperature, ranked highly in feature importance but did not improve forecast accuracy, suggesting that much of the apparent weather signal reflects stable between-county climatic differences already captured by county fixed effects. Annual-scale ArboNET-based forecasting is therefore insufficient for outbreak warning.