Prediction of Pseudomonas aeruginosa abundance in drinking water distribution systems using machine learning

The detection of Pseudomonas aeruginosa is a challenging but crucial task to ensure the bio-safety of drinking water. The current cultivation and molecular qPCR methods are costly, laborious and time-consuming, leading to inaccuracies and delayed monitoring. In this study, three machine learning (ML) models, including eXtreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Regression (SVR), were developed, interpreted, and validated for their ability to predict P. aeruginosa abundance in both urban and rural drinking water distribution systems (DWDS). To ensure the reliability and robustness of ML models, data leakage management for data pre-processing, 5-fold cross-validation and grid search for hyperparameters tuning were utilized during the training phase. To control overfitting issues, feature selection using embedded method was implemented to exclude three low-contributing input variables of oxidation-reduction potential (ORP), total chlorine, and heterotrophic plate counts (HPC). The XGBoost model outperformed RF and SVR models in terms of accuracy and generalizability in predicting P. aeruginosa abundance, achieving training/testing R2 of 0.92/ 0.85 in urban system, and 0.94/0.87 in rural system, respectively. Feature importance analysis revealed that water temperature, dissolved oxygen (DO), residual chlorine, and NO3--N were key variables for the prediction. The validation experiments, by randomly sampling from both urban and rural DWDS, demonstrated acceptable relative errors of 10.77 % and 8.86 %, respectively. Overall, this study provides an applicable ML modeling framework for the accurate and fast prediction of P. aeruginosa abundance in DWDS, potentially reducing laborious experiments in future.