Explainable Hybrid Machine Learning Framework for Electrochemical–Physicochemical Prediction of Heavy Metal Contamination in Water Systems

Abstract

Heavy metal contamination of freshwater systems is one of the most serious environmental health challenges of the twenty-first century but existing monitoring techniques are prohibitively expensive, labor intensive and poorly matched to the spatio-temporal resolution required for modern water management. Here we propose an explainable hybrid machine learning framework that combines electrochemical sensor signals and traditional physicochemical water quality parameters to predict dissolved heavy metal concentrations with high accuracy and mechanistic transparency. An experimental dataset was compiled consisting of 300 observations of voltage (V), current (µA), temperature (K), pH and conductivity (µS/cm) and measured heavy metal content (mg/L). We extracted two engineered features, electrical power and resistance, from raw electrochemical signals, to increase the discriminative capacity of the input space. Four predictive architectures were developed and benchmarked including Random Forest (RF), Gradient Boosting (GB), Artificial Neural Network (ANN) and a Hybrid RF-ANN model. The best performing model in terms of generalization was the RF model with a coefficient of determination (R2) of 0.887, root-mean-square error (RMSE) of 0.022 mg/L and mean absolute error (MAE) of 0.018 mg/L. Feature importance analysis calculated using mean decrease impurity and permutation importance found pH (MDI weight: 0.907) and electrical conductivity (MDI weight: 0.034) to be the dominant predictors, in agreement with the known electrochemical behaviour of metal ions in aqueous solution. Framework reveals that the combination of low-cost voltammetric sensors with explainable ensemble learning can offer near real-time, decision-grade estimates of heavy metal burden, providing a scalable pathway to continuous environmental surveillance and regulatory compliance monitoring.