Pathogen Risks Predicted in Drinking Water Using Machine Learning Models

Machine Learning Framework Predicts Drinking Water Pathogen Risks

Environmental measurements and machine learning algorithms can successfully estimate multiple microbial pathogens and public health risks in drinking water sources, according to research published on July 10, 2026, in Biocontaminant. Changzheng Cui’s team at the East China University of Science and Technology developed the predictive framework to address the limitations of conventional surveillance, which often relies on fecal indicator bacteria that fail to consistently reflect specific viral and bacterial pathogens.

Traditional culture methods for water monitoring can take several days, while molecular testing requires expensive equipment and specialized technicians. Seeking a way around these time constraints, the investigators gathered 95 surface water samples spanning five designated monitoring spots across two drinking water reservoirs in an eastern Chinese city during the period from May 2024 to December 2025. They tracked seven water quality parameters—including temperature, turbidity, conductivity, dissolved oxygen, and total dissolved solids—alongside six meteorological variables. Membrane filtration quantified three indicator bacteria, while quantitative polymerase chain reaction tracked six pathogens: Pseudomonas aeruginosa, Salmonella spp., Shigella spp., adenovirus, norovirus, and enterovirus.

Algorithm Performance Across Microbial Targets

Testing six different algorithms—spanning linear regression, boosting, decision trees, support vector machines, random forests, and neural networks—revealed distinct performance variations. Random forests and decision trees proved most effective for most targets, achieving coefficients of determination above 0.75. Notably, the decision tree model for P. aeruginosa exceeded 0.90. Independent temporal validation using additional samples collected in January and February 2026 generally supported these predictive capabilities, although uncertainty remained higher for several pathogens.

SHapley Additive exPlanations analysis showed that turbidity served as a dominant contributor to indicator bacteria predictions. Meanwhile, pathogen models relied on differing combinations of environmental variables. The team then incorporated the predicted concentrations into a quantitative microbial risk assessment using 10,000 Monte Carlo iterations to characterize uncertainty. The resulting disease burden distributions broadly matched calculations derived from directly measured pathogen concentrations.

Health Risk Projections and Treatment Efficiency

Under the study’s modeled exposure and treatment assumptions, Salmonella, Shigella, and enterovirus showed estimated probabilities of 31.68%, 59.70%, and 20.86%, respectively, of exceeding the established health burden benchmark. Disinfection efficiency emerged as the strongest driver of risk variability across the models. Researchers noted that molecular detection measures genetic material and does not independently establish pathogen infectivity or actual illness among consumers.

Frequently Asked Questions

Why can’t standard indicator bacteria reliably predict viruses in drinking water?

Conventional surveillance commonly relies on fecal indicator bacteria, but these organisms do not consistently reflect the presence of specific pathogens, especially viruses, prompting the need for direct molecular and machine learning estimation.

Which algorithm performed best for predicting Pseudomonas aeruginosa?

The decision tree model for P. aeruginosa achieved a coefficient of determination exceeding 0.90, making it one of the most accurate single targets in the study.

What factor drove the highest variability in predicted health risks?

Disinfection efficiency operated as the strongest driver of risk variability when researchers ran 10,000 Monte Carlo iterations to characterize uncertainty.

What is the next step for this machine learning framework?

Further validation across independent watersheds, seasons, and treatment settings is required before deployment, with potential integration into real-time monitoring platforms for targeted testing and timely interventions.