iTransKAN: A deep chemometric model for enhanced heavy metal quantification via multi-scan cyclic voltammetry

Voltammetric analysis coupled with advanced chemometrics provides a powerful framework for heavy metal quantification. However, translating complex, non-linear electrical current responses into precise concentrations remains challenging. Conventional models and standard Transformers struggle with the quadratic computational burden of sequential data and are limited in their ability to fully capture progressive signal variations across multi-scan operations. Furthermore, standard symmetric evaluation metrics assign equal penalties to all errors and critically ignore the severe environmental hazards of toxicity underestimations. To overcome these limitations, this study proposes iTransKAN, a deep chemometric framework integrating an Inverted Transformer for efficient global feature extraction with a Kolmogorov-Arnold Network (KAN) for highly flexible regression. The architecture is further refined by a Learnable Positional Encoding to capture complex sequential shifts in the voltammetric data, alongside a Scan-Level Attention Fusion to effectively aggregate informative signal patterns across scanning cycles. A novel Relative Asymmetric Toxicity Error (RATE) metric is also introduced to penalize hazardous under-predictions while maintaining evaluation fairness. Experimental results and Wilcoxon signed-rank tests confirm that iTransKAN significantly outperforms the standard iTransformer baseline ( p < 0 . 001 ). This integration reduces the Mean Absolute Error (MAE) by 72.2% for Pb and 33.8% for Cd. Consequently, the proposed model achieves a robust R 2 of 0.9997 (95% CI: 0.9993–1.0000) for Pb and 0.9814 (95% CI: 0.9570–1.0000) for Cd. These findings highlight the potential of iTransKAN to provide highly accurate, computationally efficient, and safety-aware heavy metal quantification for advanced environmental sensing.