SHAP-guided variable selection with radial basis function Kolmogorov-Arnold networks for interpretable infrared...

Infrared spectral quantification is challenged by high-dimensional inputs, overlapping absorption bands, nonlinear matrix effects, and limited interpretability. This study developed a SHapley Additive exPlanations (SHAP)-guided radial basis function Kolmogorov-Arnold network (RBF-KAN) framework for accurate, compact, and interpretable spectral regression. KAN was first benchmarked against representative nonlinear models on two public near-infrared olive datasets, demonstrating favorable generalization with a compact architecture. Nine KAN edge-function forms were then compared under identical conditions, and RBF-KAN achieved the strongest independent prediction performance, supporting localized nonlinear representation for complex spectra. Although the full-spectrum models were already accurate, SHAP revealed that predictive information was concentrated within limited local spectral regions. Based on the SHAP rankings, the Top-40 and Top-30 variables were retained to reconstruct the olive-fruit and olive-oil models, respectively, while the corresponding hyperparameters were optimized by five-fold cross-validation within the calibration set. The resulting models substantially reduced spectral dimensionality and further improved independent prediction performance. The framework was subsequently extended to whole-blood FTIR-ATR glucose quantification. The Top-23 model achieved a prediction-set Rp² of 0.9951 and an RMSEP of 0.1744 mmol/L. Bland-Altman analysis indicated low overall bias, and all 32 independent prediction samples fell within Zone A of the Clarke error grid. These results demonstrate that SHAP-guided RBF-KAN provides an accurate, compact, and interpretable strategy for infrared spectral quantification across food and biomedical matrices.