Researchers have developed a machine learning and quantitative microbial risk assessment framework—known as ML-QMRA—to predict pathogen concentrations in drinking water source monitoring, according to a study published in Biocontaminant. By analyzing routinely measured water quality indicators, the data-driven approach aims to help water managers estimate health risks faster than conventional microbial testing allows.
How the ML-QMRA Framework Predicts Water Contamination Risks
“Routine monitoring already generates a large amount of environmental information,” said Changzheng Cui, corresponding author of the study from East China University of Science and Technology. “Our goal was to determine whether these commonly measured variables could also help us anticipate microbial contamination and translate those predictions into meaningful health risk estimates.”
To build the framework, researchers collected 95 surface water samples from two drinking water sources in a major city in Eastern China between May 2024 and December 2025. The monitoring tracked three indicator bacteria—fecal coliforms, Escherichia coli, and Enterococcus faecalis—alongside six pathogens: Pseudomonas aeruginosa, Salmonella spp., Shigella spp., adenovirus, norovirus, and enterovirus.
Did you know? While fecal indicator bacteria strongly correlate with one another, the study revealed that their relationships with viral pathogens are often weak or inconsistent, meaning bacterial indicators alone may not always accurately reflect viral contamination.
Comparing Machine Learning Models for Water Quality Monitoring
The research team evaluated six different machine learning approaches to see which could best map routine physical and chemical indicators to actual pathogen concentrations. According to the study, the tested models included Multiple Linear Regression, Least Squares Boosting, Decision Tree, Support Vector Machine, Random Forest, and Multilayer Perceptron models.
The Random Forest and Decision Tree models emerged as top performers, with all optimized models achieving R² values above 0.75. Notably, the Decision Tree model demonstrated exceptional predictive strength for P. aeruginosa, scoring an R² above 0.90. Independent data collected in January and February 2026 provided temporal validation, confirming that most models could maintain predictive accuracy outside their original training periods.
Translating Pathogen Predictions into Health Risks
To evaluate what these contamination levels mean for human health, the researchers linked their predicted pathogen concentrations to quantitative microbial risk assessment (QMRA) calculations expressed as disability-adjusted life years (DALYs), according to the study. Most estimated risks stayed below the World Health Organization benchmark of 10⁻⁶ DALYs per person per year.
However, under unfavorable exposure conditions, Salmonella spp., Shigella spp., and enterovirus showed measurable probabilities of exceeding this safety benchmark. The analysis further identified disinfection efficiency as the dominant factor driving estimated health risks, underscoring the critical need for stable, effective drinking water treatment plants.
Understanding Predictive Importance Using SHAP Analysis
To demystify how the machine learning algorithms make decisions, the researchers applied SHapley Additive exPlanations (SHAP). The analysis showed that turbidity served as the single strongest predictor for fecal indicator bacteria, accounting for 41.6% to 62.1% of predictive importance in those models. Meanwhile, temperature, dissolved oxygen, and rainfall contributed varying degrees of predictive weight across individual pathogens.
Despite these promising findings, the authors emphasize that the ML-QMRA framework requires further validation across different watersheds, seasonal shifts, treatment systems, and land-use scenarios before broad deployment.
Frequently Asked Questions
What is the ML-QMRA framework in drinking water monitoring?
It is a data-driven approach that combines machine learning algorithms with quantitative microbial risk assessment to predict pathogen levels and human health risks using routine water quality measurements.
Why aren’t traditional fecal indicator bacteria enough to detect viruses?
The study found that while fecal coliforms, E. coli, and Enterococcus faecalis correlate with each other, their statistical relationships with viral pathogens like enterovirus and norovirus are weak and inconsistent.
Which machine learning models performed best in the study?
Random Forest and Decision Tree models performed the best, with all optimized versions achieving R² values above 0.75.
What water quality variable was most important for predicting indicator bacteria?
Turbidity accounted for 41.6% to 62.1% of predictive importance in the fecal indicator bacteria models.