Nationwide Prioritization and Machine Learning-Based Risk Prediction of Pharmaceuticals in China's Surface Waters

Pharmaceuticals, ubiquitous in human, veterinary and agricultural use, are prevalent emerging contaminants in Chinese surface waters. Although not highly persistent, their low removal in conventional wastewater treatment leads to continuous discharge, creating "pseudo-persistence." This chronic exposure poses significant ecological and human health risks, including hormonal disruption of female reproduction and antibiotic-induced gut microbiota alterations and antimicrobial resistance in aquatic biota.

Numerous pharmaceuticals (>100) have been detected in China's surface waters. However, clear regulatory priorities are lacking, and nationwide monitoring is insufficient, leaving many regions without concentration or risk data. This study aims to: (1) identify pharmaceuticals posing the highest human and environmental hazards; (2) develop nationwide predictive concentration models using machine learning; and (3) generate a health risk map for pharmaceuticals in China's surface waters.

Through systematic keyword searches in Web of Science and CNKI, we compiled data from 227 peer-reviewed articles (2010-2023), covering approximately 13,000 sampling sites across China's nine major river basins. Pharmaceutical concentrations, detection frequencies, and sampling metadata were extracted. To assess environmental behavior and risks, four key indicators were selected: octanol-water distribution coefficient (LogDow) for bioaccumulation potential, degradation half-life (T1/2) for persistence, predicted no-effect concentration for aquatic ecosystems (PNECeco) for ecotoxicity, and predicted no-effect concentration for human exposure (PNEChum) through drinking water and fish consumption.

Principal component analysis (PCA) integrated four indicators into a composite hazard score (HP) and to combine concentration and detection frequency into an exposure potential score (EP). Pharmaceuticals were preliminarily screened based on reference thresholds for HP and EP values, and then ranked by the product of HP and EP to establish priority control lists for each river basin. Roxithromycin and erythromycin, exhibiting high toxicity and extensive data, ranked highest across all basins. Antibiotics were consistently high-priority in all nine basins. In densely populated basins (Haihe, Yangtze, Pearl), bezafibrate, indomethacin, and ibuprofen require additional attention. Hormones (estrone, estriol, ethinylestradiol) showed elevated concentrations and risks in Songhua/Liao basins. Increased monitoring is strongly recommended for data-scarce inland basins.

Four representative pharmaceuticals (erythromycin, ciprofloxacin, norfloxacin, carbamazepine), selected based on high toxicity or exposure potential, were modeled nationally. Predictors included 27 variables across five categories: Socioeconomic, Healthcare, Agricultural and aquacultural, Natural environmental, and Water quality indicators. Seven machine learning algorithms were evaluated (DT, ExtraTrees, GB, KNN, RF, SVM, XGBoost). RF demonstrated superior performance and was selected for feature selection (via weighted backward stepwise regression) and hyperparameter tuning (grid search with 10-fold CV). The optimal model was chosen based on R² and RMSE.

Predicted concentrations were then input into the USEPA-recommended human health risk assessment model. Carbamazepine, ciprofloxacin, and norfloxacin exhibited low risks nationwide (HQ < 1). Erythromycin exceeded safe levels (HQ > 1) in eastern regions (Yangtze River Delta, Bohai Rim, Pearl River Delta). Spatially, erythromycin and norfloxacin risks displayed a distinct east-west gradient (higher east), while carbamazepine and ciprofloxacin showed minimal spatial variation.