Multi-target active component detection of Panax notoginseng based on ATR-FTIR combined with machine learning

The quality of Panax notoginseng is strongly influenced by factors such as geographical environment, harvesting period, and plant part, resulting in considerable variation in product quality. This study proposes a rapid and accurate approach for classifying different parts of Panax notoginseng (main root, rhizome, lateral root and fibrous root) and simultaneously quantifying active components (Rg1, Rb1 and R1), based on attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR) combined with machine learning. HPLC analysis revealed significant differences in Rg1, Rb1, and R1 contents across the four parts, with mean concentrations in the order: rhizome > main root > lateral root > fibrous root. In classification modeling, radial basis function neural network (RBFNN) outperformed partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM), achieving up to 97.5 % accuracy. For quantitative analysis, feature selection was conducted using competitive adaptive reweighted sampling (CARS) and uninformative variable elimination (UVE). Predictive models were subsequently established using PLS regression (PLSR), least squares SVM (LSSVM), and SVM optimized by gray wolf optimization (GWO-SVM). The optimal models for Rg1, Rb1, and R1 were CARS-PLSR, CARS-GWO-SVM, and UVE-GWO-SVM, respectively, with R2 of 0.9680, 0.9287, and 0.9255, and root mean square errors of 0.2427 %, 0.3399 %, and 0.0811 %. Importantly, the selected feature wavelengths closely matched the absorption peaks associated with the vibrational modes of saponin chemical bonds. Overall, these findings demonstrate the great potential of ATR-FTIR spectroscopy for quality control in the processing and application of Panax notoginseng.