红参提取物总皂苷近红外定量分析建模中的变量筛选

注射用益气复脉（冻干）是由红参、麦冬、五味子3种药材制成的新型冻干粉针制剂。红参提取物总皂苷是注射用益气复脉（冻干）生产过程的重要质控指标，传统分析方法分析结果具有滞后性，无法快速反馈生产过程质量信息。近红外光谱(NIR)作为一种快速无损的过程监控工具已经广泛应用于中药质量控制领域。中药成分复杂且近红外光谱吸收强度弱、谱区重叠严重，如何从干扰严重的复杂光谱中提取有效信息是提高测量准确度的关键。模型集群分析（MPA）通过随机采样，最大限度地提取了样本信息，打破了传统一次性建模思路，为变量筛选方法提供了新的思想。采集了55批红参提取物近红外光谱数据，运用多元散射校正（MSC）进行光谱数据预处理，并将MPA衍生的随机蛙跳法（RF）、竞争自适应重加权（CARS）、变量组合集群分析法（VCPA）、VCPA联合迭代保留信息变量（IRIV）方法与OPUS软件自带的变量筛选方法分别用于总皂苷含量偏最小二乘（PLS）定量分析模型的建立中。结果表明，OPUS软件、CARS-PLS与RF-PLS所建模型校正集相关系数（R_c）仅为0.601 3，0.565 3与0.644 0，拟合效果不理想。VCPA-PLS法所建模型的R_c为0.951 2，是几种变量筛选方法中最高的，但是其预测性能不佳，模型稳健性不理想。VCPA-IRIV-PLS模型具有最好的预测效果，R_c为0.928，RSEP%为7.99%。

Variable Selection Method in the NIR Quantitative Analysis Model of Total Saponins in Red Ginseng Extract

Yiqi Fumai Lyophilized Injectionis a new type of freeze-dried powder injection made of red ginseng,ophiopogon japonicus and schisandrachinensis.The total saponin content of red ginseng extract is an important quality control index in the production process of Yiqi Fumai lyophilized injection.The results of traditional analysis methods lag far behind,which cannot feedback the quality information of the production process timely,it is necessary to establish a rapid method for the determination of total saponin.As a process monitoring tool,near-infrared spectroscopy(NIR)has been widely used in the quality control of traditional Chinese medicine.How to extract the effective information from the spectrum with weak absorption and serious overlapping of spectral regions is the key to improve the monitoring veracity.Model population analysis(MPA)provides a new idea for variable selection method.In this study,the near-infrared spectrum data of 55 batches red ginseng extracts were collected.Themulti scattering correction(MSC)method was used to preprocess the spectrum data,the variable screening methods derived from MPA,such as random frog(RF),competitive adaptive reweighted sampling(CARS),variable combination population analysis(VCPA),VCPA combined with IRIV(iterative retaining information variables)and the variable selection method of OPUS were respectively used in the establishment of PLS quantitative analysis model.The results showed that the Rc of model established by OPUS,CARS-PLS and RF-PLS were only 0.6013,0.5653 and 0.6440,respectively.The Rc of model established by VCPA-PLS was 0.9512,which was the highest,but this model did not present good robustness.The model established by VCPA-IRIV-PLS had the best prediction effects;its Rc was 0.928,RSEP% was 7.99%.