Predicting arsenic and manganese contamination in private well water with Machine Learning: An integrated analysis of geologic, well construction, and permitting data.

Exposure to arsenic (As) and manganese (Mn) through private well water is a pressing public health issue in North Carolina (NC). The goal of this study was to identify geologic, geographic, well construction, and water characteristics (e.g. pH) that are predictive of As and Mn concentrations in private wells. Data collected from over 700 private well water tests from Union County, NC, were analyzed, including thirteen variables spanning well construction and water characteristics that were joined with geologic and geographic features. These data were used to develop random forest (RF) and support vector machine (SVM) models to predict whether wells exceeded their respective maximum contaminant level (MCL) or health advisory level (HAL) (10 ppb for As; 300 ppb for Mn) using four different variable combinations (use cases). Major findings included that As and Mn contamination was prevalent with 18 % and 12 % of tested wells exceeding the MCL and HAL, respectively. Most algorithms across the use cases had generally high predictivity (AUC > 0.7). The top performing models used a linear kernel SVM that leveraged all thirteen variables for As (AUC = 0.80) and the geologic/geographic variables for Mn (AUC = 0.74). When classifying As concentrations, longitude, flow rate, pH, and land use were the most important predictors of contamination. When classifying Mn concentrations, casing depth, longitude, and latitude were the most important predictors. In conclusion, these findings can be used to inform the identification of private wells at-risk for metal contamination, prioritize intervention efforts for existing wells, and provide safer parameters for newly constructed wells.