Multivariable Investigation of Physicochemical Drivers Influencing Opportunistic Pathogens in Drinking Water

Abstract

Opportunistic pathogens (OPs) in drinking water distribution systems (DWDS) and premise plumbing pose health risks. Effective management requires understanding the associations between physicochemical parameters and OP occurrence. This study applied machine learning (ML) models to explore relationships between physicochemical parameters (e.g., pH, temperature), OPs (Legionella spp., Pseudomonas spp., and Mycobacterium spp.), and the free-living amoeba Vermamoeba vermiformis. eXtreme Gradient Boosting, Random Forest (RF), and Support Vector Regression were optimized using cross-validation to predict qPCR-derived cell-equivalent (CE) concentrations, and Shapley Additive exPlanations (SHAP) were used to interpret predictor contributions. For Legionella spp., the RF model achieved the best test performance (R2 = 0.80). SHAP analysis identified pH, free chlorine, and water temperature as the most influential predictors of Legionella spp. CE concentrations. Lower free chlorine concentrations and pH values above the pKa of OCl– were associated with higher Legionella spp. CE concentrations. V. vermiformis and Mycobacterium spp. showed a strong association in the ML models, consistent with previously reported biological interactions. Across individual and combined data sets, free chlorine consistently influenced model outputs, whereas differing pH and temperature ranges altered their relative importance. Overall, ML provided a framework for identifying associations between environmental conditions and OP occurrence in DWDS and premise plumbing.