Qu Nan, Li Xuesong, Dou Ying, Mi Hong, Guo Ye, Ren Yulin
Department of Chemistry, College of Chemistry, Jilin University, Changchun 130021, China.
Eur J Pharm Sci. 2007 Jul;31(3-4):156-64. doi: 10.1016/j.ejps.2007.03.006. Epub 2007 Mar 19.
A new assay method for the nondestructive determination of erythromycin ethylsuccinate powder drug via short-wave near-infrared spectroscopy (NIR) combined with radial basis function (RBF) neural networks is investigated. The modern near-infrared spectroscopy analysis technique is efficient, simple and nondestructive, which has been used in chemical analysis in diverse fields. Short-wave NIR is a more rapid, flexible, and cost-effective method to control product concentration in pharmaceutical industry. The RBF neural networks are local approximation networks that have superiorities in function approximation and learning speed. In addition, the structure of RBF networks is simple. Estimate and calibration of the sample concentration via short-wave NIR are made with the aid of RBF models based on conventional spectra, standard normal variate (SNV), multiplicative scatter correction (MSC) and the first-derivative spectra. Various optimum models of them are established and compared. Experiment results show that the models of SNV spectra can give better performance, and the optimized RBF neural network model after SNV treatment were given, by which the root-mean-square-errors (RMSE) for calibration set and test set were 0.3266% and 0.5244%, respectively and the correlation coefficients (R) for calibration set and test set were 0.9942 and 0.9852, respectively. The proposed RBF method based on short-wave NIR is more valuable and economical for quantitative analysis than traditional methods such as partial least squares (PLS).
研究了一种结合径向基函数（RBF）神经网络，利用短波近红外光谱（NIR）无损测定琥乙红霉素粉针剂药物的新方法。现代近红外光谱分析技术高效、简单且无损，已在多个领域的化学分析中得到应用。短波近红外光谱是制药行业控制产品浓度更快速、灵活且经济高效的方法。RBF神经网络是局部逼近网络，在函数逼近和学习速度方面具有优势。此外，RBF网络结构简单。借助基于常规光谱、标准正态变量（SNV）、多元散射校正（MSC）和一阶导数光谱的RBF模型，通过短波近红外光谱对样品浓度进行估计和校准。建立并比较了它们的各种最优模型。实验结果表明，SNV光谱模型性能更佳，并给出了经SNV处理后的优化RBF神经网络模型，其校准集和测试集的均方根误差（RMSE）分别为0.3266%和0.5244%，校准集和测试集的相关系数（R）分别为0.9942和0.9852。所提出的基于短波近红外光谱的RBF方法在定量分析方面比偏最小二乘法（PLS）等传统方法更具价值和经济性。
