Species discrimination and VIP-stacking quantitative models for Curcumae Rhizoma utilizing multi-modal spectra combined with machine learning algorithm

Curcumae Rhizoma (Ezhu) is a multi-species herbal medicine with excellent medicinal value and development potential. However, challenges such as the difficulty in differentiating its varieties and the limitations of current methods for determining minor component content, which are time-consuming and cumbersome, necessitate improved approaches. Spectroscopic techniques combined with chemometrics offer a powerful alternative for developing qualitative and quantitative models, and the spectral data fusion has emerged as a key research hotpot. This study employed multi-modal spectroscopy including Fourier transform infrared (FT-IR), Fourier transform near-infrared (FT-NIR), and ultraviolet (UV) combined with multivariate algorithms to establish species discrimination and content prediction models for minor constituents in Ezhu. For qualitative analysis, linear discriminant analysis (LDA), k-nearest neighbor (KNN), and decision tree (DT) models based on fused UV+FT-NIR+FT-IR spectral data achieved 100 % classification accuracy. For quantitative analysis, a novel variable importance in projection (VIP)-guided stacking ensemble strategy was proposed, leveraging VIP scores derived from partial least squares regression (PLSR) to optimize base-learner combinations. This approach successfully constructed robust models for predicting the content of (3,5-dihydroxy-1-(3,4-dihydroxyphenyl)-7-(4-hydroxyphenyl)-heptane), (1,7-bis-(4-hydroxychalcone)-3,5-dihydroxy-heptane), (3S,5S)-3-acetoxy-5-hydroxy-1-(3,4-dihydroxyphenyl)-7-(4-hydroxyphenyl)-heptane, germacrone, zederone, curzerene, and curdione. Compared to conventional machine learning models and prior studies, the VIP-stacking ensemble models demonstrated superior predictive accuracy and robustness. This work highlights the efficacy of spectral data fusion in both qualitative and quantitative analyses and validates the potential of VIP-stacking ensemble strategies to enhance the performance of content prediction models. This study not only offers a more effective way to identify and quality control of Ezhu, but also provides a promising approach for species authentication and quality control in pharmaceutical, agricultural, and food science applications.