10th U6+ Consortium of African Universities International Conference

Access to safe drinking water remains a pressing challenge across rural South Africa, where resource constraints severely limit the frequency and coverage of water quality monitoring. The Water Quality Index (WQI), while providing an accessible summary metric of water safety, requires extensive laboratory analyses, limiting its operational scalability. This study develops, validates, and interprets a parsimonious machine learning surrogate for WQI prediction using routinely measurable physicochemical parameters collected from 18 villages in the Eastern Cape Province. Four regression algorithms: Ridge Regression, Random Forest, Extra Trees, and XGBoost, were benchmarked under recursive feature elimination with cross-validation. Ridge Regression emerged as the optimal model, achieving cross-validated R² = 0.977 ± 0.028 and test-set R² = 0.887, RMSE = 168.83, and MAE = 74.58 using only seven predictors: pH, turbidity, free chlorine, phosphate, nitrates, ammonia, and potassium. Global model interpretability was assessed using SHapley Additive exPlanations (SHAP), which identified turbidity, nitrates, and ammonia as the three dominant predictors; followed by pH, free chlorine, phosphate, and potassium, consistent with established hydrochemical theory. Distribution-free uncertainty quantification via conformal prediction yielded 92% empirical coverage, approaching the nominal 95% target. Spatial generalisation was evaluated through leave-one-village-out cross-validation, yielding a mean R² of 0.849 ± 0.290. The surrogate equation, WQI ≈ −231.26 + 33.97pH + 12.47Turbidity − 40.32Free Chlorine − 14.04Phosphate − 3.80Nitrates + 63.93Ammonia + 4.58K, provides a transparent, deployable tool for cost-effective water quality screening in resource-limited settings.