Holistic monitoring framework for drinking water management: Microbiome data-assisted machine learning assessment for disinfectant byproducts

Though disinfection byproducts (DBPs) in finished drinking water are a pervasive health hazard, utilities still rely on the occasional passive monitoring of these compounds. In this study, we present an integrative data-driven machine learning (ML) modeling approach that integrates microbiome information with routinely measured water quality data for the predictive assessment of DBP levels. The optimized ML model was streamlined via systematic model tuning and feature selection, which improved (47.6 % of R²) prediction performances compared to baseline linear models. Three independent explainable artificial intelligence diagnostic tools identified two commonly measured water quality metrics (turbidity and residual chlorine) and four bacterial taxa as robust explanatory predictors. Interactive response plots revealed that the microbiome indicators tracked DBP levels independently of the chlorine concentration, thus decoupling DBPs from their precursor. Phylogenetic analysis of the bacterial indicators revealed that they frequently occurred in the finished drinking water in a number of countries, supporting their generalized use for risk assessment. The proposed ML modeling approach can be developed to shift current DBP monitoring practices from retrospective compliance testing to proactive risk management, while also guiding the review and adjustment of operational conditions to mitigate the formation of DBPs. This integrative data-driven ML modeling approach holds great promise as a holistic monitoring framework simultaneously surveilling unregulated pathogenic hazards and occasionally measured health hazards such as DBPs, potentially complementing existing monitoring practices for finished drinking water.