Modeling Environmental Exposures to Predict Health Outcomes in the Southeastern US

Abstract

Predicting health outcomes for private well users is difficult due to the lack of data associated with well user habits and private well location. Depending on the habits of a private well user, they may be exposed to constituents in the groundwater associated with poor health outcomes. Environmental exposures are modeled in numerous ways including geospatial and machine learning models, often at population scales like the state or country. Here, we used Poisson regression to estimate the association between percent private well use and several cardiovascular health outcomes for Alabama census block groups (CBGs), modifying that relationship by the physiographic region, nitrate concentration, and percent agricultural land use, respectively. Next, sample data used to predict CBG nitrate concentration for that study were repurposed to develop several random forest machine learning models. Unique random forest models were tuned, trained, and tested for 18 southeastern US aquifers, the Coastal Plain aquifer system, the Piedmont and Blue Ridge crystalline rock aquifer, and the Mississippi River Valley Alluvial aquifer. We compared variable importance, R-squared, and RMSE values for these models and cross validated each using testing data from the other models. We estimated an inverse association between private well use and hypertensive heart disease mortality and stroke/cerebrovascular disease mortality, respectively, and found the association between private well use and stroke/cerebrovascular disease is positive when modified by percent agricultural land use one standard deviation above the mean. We also found that single aquifer models performed poorly when predicting nitrate for other lithologies whereas the 18-aquifer model predicted nitrate concentration with higher R-squared values and lower RMSE values when cross validated on the single aquifer models.