Synthetic noise injection asymmetrically affects machine-learning quantification of Cd2+ and Pb2+ in multi-scan cyclic voltammetry

Machine-learning quantification of toxic trace metals such as Cd2+ and Pb2+ from cyclic voltammetry (CV) reads concentration from peak features that a deterministic peak-finding step extracts from repeated scans, and its accuracy is limited by how stably those features can be recovered from noisy signals. The prevailing assumption treats acquisition noise as a nuisance to be removed before feature extraction. The present study tests the opposite hypothesis, namely that controlled noise injected before feature aggregation can improve quantification. A Random Forest regressor was applied to seventeen multi-scan peak features extracted from 1500 voltammograms spanning fifteen concentrations of each metal. Four noise distributions, namely Gaussian, Student’s-t, pink 1/f3, and heteroscedastic, were injected at a peak-relative signal-to-noise ratio of 30 dB and evaluated over twenty seeds with exact Wilcoxon tests, Cohen’s dz, and bootstrap confidence intervals. Noise injection reduced the mid-range mean absolute percentage error (MAPE) for Cd2+ by 55.0 to 63.0% across all four distributions, with p<10−6 and dz near −3, while the coefficient of determination remained near 0.96. Lead, whose voltammetric peaks are an order of magnitude sharper, showed no reliable benefit. Permutation feature importance, a nested-cross-validation check, a second tree-based regressor, a train-only control, and a synthetic-peak study together supported a peak-shape mechanism, in which injected noise perturbs deterministic peak finding on broad asymmetric peaks and renders across-scan dispersion features informative. Controlled acquisition noise therefore improves mid-range quantification for analytes with broad peaks, and the dissociation between a stable coefficient of determination and a large error reduction provides a diagnostic for peak-finding artefacts in related CV pipelines.