Near infrared system coupled chemometric algorithms for the variable selection and prediction of baicalin in three different processes.

Characteristic variables are essential and necessary basis in model construction, and are related to the prediction result closely in near infrared spectroscopy (NIRS) analysis. However, the same compound usually has different characteristic variables for different analysis and it would be lower correlation between variables and structure in many researches. So, the accuracy and reliability are expected to improve by exploring characteristic variables in different spectrum analysis. In this study, competitive adaptive weighted resampling method (CARS) was applied to select characteristic variables related to baicalin from NIRS analysis data, which were applied to analysis of baicalin in three different processes including the herb, extraction process and concentration process of Scutellaria baicalensis. After application of CARS method, 70, 50 and 50 variables were selected respectively from three processes above. The selected variables were firstly analyzed by statistical methods that they were found to be consistent and correlated among three different processes after one-way analysis of variance test and Kendall's W. Partial least-squares (PLS) regression and extreme learning machine (ELM) models were constructed based on optimized data. Models after variable selection were less complicated and had better prediction results than global models. After comparison, CARS-PLS was suitable for the prediction of extraction process, while for the concentration process and herb, CARS-ELM performed better. The Rc value of the herb, extraction and concentration model were 0.9469, 0.9841 and 0.9675, respectively. The RSEP values were 4.54%, 6.96% and 8.37%, respectively. The results help to frame a theoretical basis for characteristic variables of baicalin.