Objective: Safe and reliable drinking water remains central to Sustainable Development Goal and is a national priority in New Zealand, where agricultural intensification and climate variability increase vulnerability to microbial contamination. This study develops an integrated machine learning (ML) framework to predict microbial contamination and campylobacteriosis risk across the full drinking-water pathway, from source to health outcome.

Methods: A three-stage ML modelling framework A three-stage modelling framework (source contamination, treatment performance, and health outcomes) was implemented using Boosted Regression Trees, Random Forests, and XGBoost to capture nonlinear environmental, infrastructural, and sociodemographic interactions. Models were spatially harmonised at meshblock level and interpreted using SHAP (Shapley Additive Explanations) to identify key predictors and equity implications.

Results: E. coli detections in treated water displayed spatial clustering in agriculturally intensive regions, including the Canterbury Plains and parts of the central North Island, while testing frequency was concentrated in major urban centres. SHAP analysis of the Boosted Regression Tree model indicated that meshblocks with higher testing frequency tended to have lower predicted risk, reflecting stronger surveillance intensity and system oversight. A SHAP dependence plot focusing on the mean E. coli detection proportion shows that as contamination increases from 0 to around 0.2, the SHAP value rises sharply — indicating a strong positive contribution to predicted campylobacteriosis notifications.

Conclusion: This multi-stage framework links environmental drivers, infrastructure performance, and health outcomes, providing an interpretable and policy-relevant tool to strengthen early warning, equity-focused surveillance, and infrastructure investment planning under changing climatic conditions.