Prediction of trihalomethane occurrence and cancer risk using interpretable machine learning and virtual data augmentation

Trihalomethanes (THMs) in drinking water are regulated for carcinogenic health risks. However, frequent water quality monitoring imposes significant resource burdens. This study proposes a framework integrating interpretable machine learning (ML) with virtual data augmentation to predict THM occurrence and associated cancer risks. Based on 146 real samples, this study uniquely uses COD_Mn (per Chinese standards) as a proxy for THM precursors. The ML results show that the CatBoost model with Bayesian optimization achieves the best R² values of 0.805–0.960 for THM species and cancer risk using six cost-effective input parameters. SHAP-based feature selection simplifies the model to four key parameters (COD_Mn, temperature, chlorine, and nitrate), while still maintaining R² values of 0.803 and 0.915 for T-THMs and cancer risks, respectively, highlighting its effectiveness for low-cost monitoring. Furthermore, a data augmentation method based on Uniform Manifold Approximation and Projection (UMAP) is proposed. By preserving the manifold during dimensionality reduction and then generating high-quality virtual data through neighborhood-based interpolation, this method reduced RMSE and MAE by 9.64–12.28 %, outperforming the baseline method using Generative Adversarial Network. These findings support data-driven soft sensing for water quality and health risk management in data-limited areas.