基于多变量检测限的模型变量筛选方法研究

摘要

This study was aimed to optimize the near infrared (NIR) variable selection method based on multivariate detection limit (MDL). Using Qing-Kai-Ling (QKL) injection as object, three variable selection methods (interval par-tial least-squares, iPLS; backward interval partial least squares, BiPLS; moving window interval partial least squares, mwPLS) were used to establish the PLS models of baicalin in QKL injection, respectively. The prediction ability of different variable selection method was compared. MDL of all models were calculated in contrast to the MDL value of full spectra PLS model, to select optimal variable selection method. The results showed that different variable selec-tion methods had different prediction ability. Among them, iPLS had the best performance which determination coef-ficient of prediction (Rpre2) and the root mean square errors of prediction (SEP) were 0.996 5 and 602.3 μg·mL-1, re-spectively. All MDLs of different variable selection methods were reduced compared with the full spectra PLS model. The value of iPLS was the lowest comes to be 1.19 μg·mL-1. The results above indicated that the best variable se-lection method for baicalin in QKL injection was iPLS. MDL theory took the error of calibration and validation set and the leverage of external sample into account, which can comprehensively evaluate model detection performance compared to the classic chemical indicator parameters. This method was particularly suitable for the variable selec-tion method optimization of NIR quantitative model of low concentration sample such as Chinese herbal medicine.%目的：优选清开灵注射液中黄芩苷含量的偏最小二乘模型的变量筛选方法。方法：以近红外透射模式，采集清开灵注射液近红外光谱，分别采用间隔偏最最小二乘算法（iPLS）、向后间隔偏最小二乘算法（BiPLS）、移动窗口间隔偏最小二乘算法（mwPLS）对光谱进行变量筛选，建立清开灵中黄芩苷定量模型，与全谱PLS模型进行比较，优选出最佳变量筛选方法；计算不同变量筛选方法下的多变量检测限（MDL），以MDL值优选出最佳变量筛选方法，验证经典化学指示参数得出的结果。结果：经不同变量筛选方法处理后所建PLS定量模型的预测性能不同，其中iPLS算法所建模型预测能力最好，预测集决定系数（Rpre2）和校正均方根误差（SEP）分别为0.9965和602.3μg·mL-1；iPLS变量筛选所得MDL值最低（1.19 mg·mL-1）， MDL值计算结果与经典化学指示参数结果一致。结论：本文以MDL值作为最佳变量筛选方法评价指标，多变量检测限同时考虑了待测成分校正集样本相应误差，测试集样本相应误差以及未知样品的杠杆率，作为一个综合量化指标，能够用于中药分析体系变量筛选方法的优选。