On-site quality control of Hypericum perforatum L. in the Xinjiang Region of China using a portable near-infrared spectrometer

Field quality control of Hypericum perforatum L. (HPL) was conducted using portable near-infrared spectroscopy (NIRS). Multi-component content prediction models and NIRS-HPLC signal conversion models were established to predict the contents of six marker compounds and the characteristic fingerprint of HPL, respectively. In this study, the impact of five regression algorithms and 73 spectral preprocessing methods on model performance was evaluated, and a feature importance-backward stepwise regression-recursive feature elimination (FI-BSR-RFE) algorithm was developed for selecting feature variables. For multi-component content prediction, Ridge regression models exhibited superior performance for compounds with higher concentrations. Meanwhile, the PLSR model yielded better prediction results for compounds with lower concentrations. The FI-BSR-RFE algorithm effectively reduced model complexity and significantly enhanced the predictive capability of the models. For the multivariate signal conversion model, the ensemble weighted approach effectively combined the predictive strengths of multiple models, significantly improving the overall prediction stability and accuracy. External validation results demonstrated excellent prediction for chlorogenic acid, hyperoside, rutin, isoquercitrin, and quercitrin, with relative errors all less than 10 %. For quercetin, 90 % of the samples had relative errors less than 10 %, while one batch exhibited a relative error greater than 10 % between the true and predicted values. For the multivariate signal conversion model, the similarity between the true and predicted spectra for all 10 batches of external validation samples exceeded 0.95, indicating excellent model performance.