Answers are grounded in this paper and cite their sources.
Binary classification is the standard approach in machine learning-based water quality monitoring, but its false negative rates remain a practical concern-contaminated samples misclassified as safe go undetected.This study applies selective prediction to water potability screening and characterises the resulting Accuracy-Safety Decoupling, demonstrating that contamination detection can be substantially improved independently of overall classification accuracy through a confidence-based triage framework.Rather than relying on binary decision boundaries, the framework sorts predictions into safe, uncertain, and contaminated zones using dual probability thresholds, directing uncertain cases to human review instead of automatic clearance.A Random Forest classifier was trained on 46 features engineered from 15 physicochemical parameters in a synthetic water quality dataset of 2,500 samples generated under controlled conditions to isolate framework performance from dataset-specific artifacts.The model achieved 79.2% accuracy and 86.79% ROC-AUC.The triage system raised the contamination capture rate from 75.21% to 98.3%, a triage-level measure of samples not auto-cleared as safe, an improvement confirmed by McNemar's test (χ² = 26.04,p < 0.0001), A paired risk difference of 23.14 percentage points (95% CI [15.63, 30.65]), and non-overlapping bootstrap confidence intervals.SHAP and LIME analysis showed moderate-to-strong agreement (Spearman ρ = 0.608), with chloramines, turbidity, and pH emerging as the top predictive factors.The dataset was generated under a documented parametric specification with an explicit rule-based labelling function, establishing a controlled methodological baseline against which future real-world deployments can be benchmarked.
No takes yet. Share an insight, caveat, or question.
alsadi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: