Using a neural network for mining interpretable relationships of West Nile risk factors

The West Nile Virus (WNV) is an infectious disease spreading rapidly throughout the United States, causing illness among thousands of birds, animals, and humans. Yet, we only have a rudimentary understanding of how the mosquito-borne virus operates in complex avian-human environmental systems. The four broad categories of risk factors underlying WNV incidences are: environmental (temperature, precipitation, wetlands), socioeconomic (housing age), built-environment (catch basins, ditches), and existing mosquito abatement policies. This research first built a model incorporating the non-linear relationship between WNV incidences and hypothesized risk factors and second, identified important factor(s) whose management would result in effective disease prevention and containment. The research was conducted in the Metropolitan area of Minnesota, which had experienced significant WNV outbreaks from 2002. Computational neural network (CNN) modeling was used to understand the occurrence of WNV infected dead birds because of their ability to capture complex relationships with higher accuracy than linear models. Further a detailed interpretation technique, based on weights and biases of the network, provided a means for extracting relationships between risk factors and disease occurrence. Five risk factors: proximity to bogs, lakes, temperature, housing age, and developed medium density land cover class, were selected by the model. The detailed interpretation indicated that temperature, age of houses, and developed medium density land cover were positively related, and distance to bogs and lakes were negatively related to the incidence of WNV. This paper provides both applied and methodological contributions to the field of health geography. The relationships between the risk factors and disease occurrence could contribute to vector control strategies such as targeted insecticide spraying near bogs and lakes, mosquito control treatments for older houses, and extensive mapping, inspection, and treatments of catch basins. The proposed interpretation technique expanded the role of CNN models in health sciences as both predictive and explanatory tools.