Machine learning-driven prioritization of micropollutants in river water: Integrating ecological and human health risks from a two-year monitoring campaign

Abstract

In this study, the prioritization of micropollutants in river water was performed via a two-year monitoring campaign of 117 organic micropollutants (OMPs) (December 2022-August 2024) and machine-learning models. A total of 82 OMPs including 17 pharmaceuticals, 26 pesticides, 14 organophosphate flame retardants, 14 perfluoroalkyl substances, 10 plasticizers and 1,2,3-benzotriazole, were detected at least once with total concentrations ranging from 242 ng/L to 4788 ng/L. Pharmaceuticals (average: 710 ng/L) were the most predominant compounds followed by pesticides (460 ng/L), organophosphate flame retardants (276 ng/L), and plasticizers (219 ng/L) over this period. To systematically prioritize OMPs that are relevant to both ecological and human exposure, a machine learning (ML)-optimized Toxicological Prioritization Index (ToxPi) was applied. The ToxPi model combined various variables like occurrence characteristics (detection frequency, mean concentration), physicochemical properties (biodegradation half-life, log K-ow, water solubility, bioconcentration factor, drinking water treatment plant treatability), and toxicity indicators (predicted no-effect concentration, median lethal dose (LD50), carcinogenicity). Random forest regression was employed to objectively optimize the weighting of risk-related variables. For both of ecological and human exposure scenarios, the selected model demonstrated stable performance (R-2 > 0.76 on both the training and test datasets). The model identified detection frequency and concentration as the most important variables in ecological exposure risk. An embedded molecular structure-based model was also used to predict the treatability of detected compounds in drinking water treatment plants. Consequently, for human exposure, carcinogenic potential, acute toxicity (LD50) and water treatability were the key variables influencing prioritization. By integrating these weighting criteria, compounds such as metformin, carbendazim, alachlor, metolachlor, bentazone, T3CPP, and PFHpA were found to pose a high risk to both ecological and human health and were categorized as the top 20 % high-priority OMPs.