A data-centric framework for simultaneous Pb²⁺ and Cd²⁺ prediction from electrochemical signals

Abstract

Monitoring heavy-metal pollution remains an important analytical challenge. Conventional laboratory methods are sensitive but often constrained by cost, turnaround time, and centralized infrastructure. Electrochemical sensors remain attractive because they can deliver rapid measurements on simpler platforms and, in some cases, closer to the point of sampling. Pb and Cd are especially important targets because they are toxic, environmentally persistent, and frequently occur together in real samples. That coexistence turns concentration estimation into a simultaneous-prediction problem rather than two separate single-ion calibrations. In this study, square-wave anodic stripping voltammetry measurements recorded at an rGO/FeCo-MOF-modified glassy carbon electrode were reformulated as a supervised multi-output regression task. The dataset comprised 351 concentration conditions measured in triplicate, yielding 1,053 experimental records. Nine electrochemical descriptors were used as model inputs, while Pb²⁺ and Cd²⁺ concentrations were treated as paired prediction targets. Median imputation and z-score normalization were applied within a uniform preprocessing pipeline. The dataset was split into 60% training data and 40% test data, and five-fold cross-validation was performed on the training set. Five regression models - Ridge, SVR with an RBF kernel, Random Forest, K-Nearest Neighbors, and Gradient Boosting - were compared. Random Forest achieved the best overall performance (test RMSE_macro = 17.2611; test R²_macro = 0.82446), and the diagnostic plots indicate stable behavior across most of the concentration range. The main contribution, however, is not only model ranking. It lies in showing how an electrochemical measurement campaign can be converted into a learning-ready dataset for simultaneous prediction.