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

Area of Science:

• Environmental Chemistry
• Water Quality Monitoring
• Machine Learning Applications

Background:

• Trihalomethanes (THMs) in drinking water pose carcinogenic health risks, necessitating regular monitoring.
• Current water quality monitoring methods are resource-intensive, limiting frequent analysis.
• Predictive models are needed for efficient and cost-effective THM assessment.

Purpose of the Study:

• To develop an interpretable machine learning (ML) framework for predicting THM occurrence and associated cancer risks.
• To integrate virtual data augmentation for enhancing predictive model performance in data-limited scenarios.
• To establish a cost-effective soft sensing approach for water quality and health risk management.

Main Methods:

• Utilized 146 real water samples, employing COD_Mn as a proxy for THM precursors.
• Implemented interpretable ML models, including CatBoost with Bayesian optimization, for THM prediction.
• Applied SHAP for feature selection and Uniform Manifold Approximation and Projection (UMAP) for virtual data augmentation.

Main Results:

• The CatBoost model achieved high R² values (0.805-0.960) for THM species and cancer risk prediction.
• Simplified models using four key parameters (COD_Mn, temperature, chlorine, nitrate) maintained strong predictive accuracy (R²=0.803-0.915).
• UMAP-based data augmentation outperformed Generative Adversarial Networks, reducing RMSE and MAE by 9.64-12.28%.

Conclusions:

• The proposed ML framework offers an effective, low-cost solution for monitoring THMs and cancer risks.
• UMAP-based data augmentation significantly improves model performance, especially in data-scarce environments.
• This approach supports data-driven soft sensing for enhanced water quality and health risk management.