为寻找药品近红外通用性定量模型在建立过程中用于确立最优模型的关键评价参数组合, 收集整理了目前各种商品化化学计量学软件及文献中的13个常用于评价近红外定量模型的统计学参数, 结合人用药品注册技术要求国际协调会对于药品定量分析方法验证基本要求, 对92个药品近红外通用性定量分析模型的这些参数进行了计算和分析。 通过对各个参数之间相互关系的研究, 确定了适合于药品近红外通用性定量分析模型评价的参数组合, 并统计出了这些参数的数值范围: 用于模型准确性评价的关键参数为交叉验证均方根误差/预测均方根误差、 平均相对偏差和相对分析误差; 大部分交叉验证均方根误差/预测均方根误差结果在3%以内, 其中交叉验证均方根误差在数值上与平均绝对偏差相当, 大部分相对分析误差值大于2, 而平均相对偏差的数值与所建模型的类型（剂型、 样品的包装形式）和待测成分含量的分布有关。 模型线性评价关键参数为决定系数; 大部分模型的决定系数在80%～100%之间。 模型耐用性关键评价参数为预测均方根误差与交叉验证均方根误差的比值, 大部分模型该参数在1.5以内。 精密度评价关键参数为重复测定结果的标准差; 该参数对于规范近红外的操作, 以及考核模型能否在不同仪器间传递具有重要的意义, 但目前药品近红外通用性定量模型对于分析精密度的关注较少, 无法估计出具体数值范围。 该研究不仅为药品近红外通用性模型的建立者和使用者提供了评价模型优劣的依据, 也为完善药品近红外光谱通用性定量分析模型的参数评价体系提供了基础数据。

In order to find out the optimum combination of the evaluation parameters for the selection of the best drug near infrared (NIR) universal quantitative model during model optimization, 13 common evaluation parameters of NIR quantitative models were collected and arranged from commercial chemometrics software or references based on the requirements of validation of quantitative analytical procedures of ICH (International Conference on Harmonization of Technical Requirements for Registration of Pharmaceuticals for Human Use). Then all these parameters of 92 drug NIR universal quantitative models were calculated and analyzed. By studying the correlation of these parameters, the optimum combination of evaluation parameters for drug NIR universal quantitative models was determined. And the value range of these parameters in the optimum combination was also obtained. Root mean square error of cross-validation(RMSECV)/root mean square error of prediction (RMSEP), average relative deviation (ARD) and ratio of (standard error of) prediction (validation) to (standard) deviation (RPD) were used as the key parameters to evaluate the model accuracy. Most of RMSECV/RMSEP was within 3%, and the value of RMSECV was roughly equivalent to the average absolute deviation of the corresponding model. Most of RPD was more than 2. The value of ARD was related to the type of universal models (such as the drug preparation and packing) and the content range which the test sample belonged to. Determination coefficient (R2) was used as the key parameter to evaluate the model linearity and most of its values were from 80% to 100%. The ratio of RMSEP to RMSECV was selected as the key evaluation parameter of model robustness and its value was usually within 1.5. The standard deviation of repeated measurement data was chosen to evaluate model precision. And it was an important parameter for standardizing operation of NIR instruments and studying the feasibility of model transfer in different instruments. However, the parameter for NIR universal quantitative models received much less attention in previous studies and it was difficult to give a value range for this parameter at present. All the results can not only provide evidence for evaluation of drug NIR universal quantitative models for the model builders or users, but also supply basic data to establish and improve the parameter evaluation system of drug NIR universal quantitative models. Near infrared spectroscopy; Universal quantitative models; Model optimization; Evaluation parameters; Accuracy; Linearity; Precision; Robustness