Beyond point predictions: Quantifying uncertainty in E. coli ML-based monitoring

Machine learning regression models are increasingly used to improve management, decision-making, and monitoring of drinking water quality, leveraging growing data from real-time sensors and laboratory analyses. However, most models provide only point predictions, ignoring inherent uncertainty caused by unobserved factors that can produce varying outcomes under similar conditions. This study benchmarks state-of-the-art regression algorithms and uncertainty quantification methods for predicting E. coli concentrations in a drinking water catchment. Gradient-boosted decision trees (GBDT) proved effective for real-time tracking, with CatBoost achieving the lowest error (RMSLE = 0.877), improving on the naïve baseline (1.160) and outperforming Random Forest by 5 %. Uncertainty quantification techniques successfully generated valid prediction intervals to identify high-risk contamination events, with Conformalized Quantile Regression emerging as the most reliable method. By combining accurate GBDT predictions with well-calibrated uncertainty estimates, this approach enhances microbial water quality forecasting, offering improved risk assessment and supporting more robust decision-making in drinking water management.