Environmental surveillance modeling: A predictive respiratory alert model for the Shenandoah Valley, Virginia

Chronic Obstructive Pulmonary Disease (COPD) is a serious respiratory ailment that affects millions of Americans. Several studies have shown that weather conditions and pollution can increase the occurrence of respiratory distress. The goal of the work described in this paper was to determine if the relationships between environmental variables and admissions rates for COPD were strong enough to enable the development of a surveillance system that could alert the population of potentially hazardous conditions. To conduct this study we obtained data on COPD admissions in the Shenandoah Valley of Virginia, an area of approximately 33,705 km<;sup>2<;/sup>. The data were coded at the zip code level (approximately 250 km<;sup>2<;/sup>). We obtained data for weather variables from 6 monitoring stations and used Kriging to estimate their values at the zip code level. We controlled for the effects of influenza in admission rates, although this required smoothing methods to impute missing values. We also controlled for different types of land use. To predict COPD admissions we developed three types of models: generalized linear models (GLM), multivariate adaptive regression splines (MARS), and random forests. All models showed that temperature or season was a significant (p <; 0.05) predictor for COPD admissions. In terms of predictive accuracy the random forest model provided the best performance measured by the receiver operations characteristic (ROC) and can provide the basis for strategic planning rather than tactical alerting.