近红外光谱技术应用于中药四类味觉分类辨识的可行性分析

Feasibility analysis of near-infrared spectroscopy technology applied to classification and identification of four kinds of taste in traditional Chinese medicine

目的 探讨近红外光谱(near infrared spectroscopy,NIRS)技术用于不同味觉中药分类辨识的可行性。方法 以分别具有苦、甜、酸、咸4种味道的35种饮片水煎液和12种常用食品类成分溶液为研究载体,获取其NIRS信息作为自变量(X),以《中国药典》2020年版一部饮片性状项下味觉描述结合口尝结果作为标杆信息(Y),比较5种光谱预处理方法,然后利用主成分分析-判别分析(principal component analysis-discriminant analysis,PCA-DA)、偏最小二乘-判别分析(partial least squares-discriminant analysis,PLS-DA)、K-近邻算法(K-nearest neighbor algorithm,KNN)分别对中药苦、甜、酸、咸4类味觉进行模型辨识探讨,并基于留一法交互验证结果的混淆矩阵(confusionmatrix,CM)和敏感性、特异性、精度等指标对模型的性能进行综合评价。结果 标准正态变量变换(standard normal variable transformation,SNV)是相对更有效的预处理方法,以预处理后的光谱数据建立的PCA-DA模型为最优辨识模型,其对苦与非苦、甜与非甜、酸与非酸、咸与非咸、四分类辨识的留一法交互验证正判率分别为89.4%、93.6%、87.2%、97.9%、87.2%。四分类辨识混淆矩阵也以PCA-DA模型性能较好,对苦、甜、酸、咸的分类正确率分别为87%、94%、73%、100%。PCA-DA模型的敏感性、特异性、精度分别平均为0.89、0.91、0.88,均极显著优于PLS-DA和KNN模型(P<0.01)。结论 基于NIRS技术初步建立了中药苦、甜、酸、咸4类味觉的分类辨识模型,可为中药五味的定性辨识研究提供新的方法参考。

Objective To explore the feasibility of near-infrared spectroscopy(NIRS) technology for classification and identification of traditional Chinese medicine with different tastes. Methods The 35 kinds of traditional Chinese medicine decoctions and 12kinds of common food ingredient solutions with four tastes of bitterness, sweetness, sourness and saltiness were taken as research object. The spectral information of samples which was obtained by near-infrared spectroscopy technology was used as independent variable(X), the taste description of Chinese medicinal decoction pieces feature in the first part of Chinese Pharmacopoeia(2020 edition) combined with the results of traditional human taste panel method were used as benchmarking information(Y). After compared five kinds of spectral pretreatment methods, three chemometric methods including principal component analysis-discriminant analysis(PCA-DA), partial least square-discriminant analysis(PLS-DA) and KNN(K-nearest neighbor) were used to establish the identification model of four kinds of traditional Chinese medicine tastes, respectively. The performance of models was evaluated synthetically by confusion matrix(CM), sensitivity, specificity and precision of leave-one-out cross validation results. Results Standard normal variate transformation(SNV) was a relatively effective pretreatment method. The PCA-DA model established with the pretreatment spectral data was the optimal identification model, and the accuracy of leave-one-out cross validation of bitterness or non-bitterness, sweetness or non-sweetness, sourness or non-sourness, saltiness and non-saltiness and four-class identification were 89.4%, 93.6%, 87.2%, 97.9% and 87.2%, respectively. The PCA-DA model performed better in confusion matrix of four-class identification, and the classification accuracy of bitterness, sweetness, sourness and saltiness were 87%, 94%, 73% and 100%,respectively. The average sensitivity, specificity and precision of PCA-DA model were 0.89, 0.91 and 0.88, respectively, which were significantly higher than PLS-DA and KNN models(P < 0.01). Conclusion In this study, the identification models to bitterness,sweetness, sourness and saltiness of traditional Chinese medicine were established by near-infrared spectroscopy technology, which provided a new method for qualitative identification of five flavors of traditional Chinese medicine.