以注射用头孢他啶通用性定量分析模型为例，通过对新的预测样本光谱与原模型训练集光谱相似性的研究，寻找更为合理的用于判断模型更新的指标以及模型更新的一般方法。对注射用头孢他啶新的待测样品，首先使用聚类分析的方法将所有样品分成5类，依次加入每一类中代表性光谱到原模型中进行模型更新，计算模型更新前后的平均预测偏差，并以此作为评判模型更新是否有效的标准。同时以相似系数为指标探讨模型更新的一般方法。再应用注射用头孢他啶通用性定量分析模型（包括头孢他啶、水分及精氨酸的定量分析模型）以及新的预测样品对该方法进行验证。实验数据证明当新样本光谱与原模型训练集样本的平均光谱在原模型建模谱段上的相似系数（ rT ）小于96.5％时，模型需要更新。使用上述判断指标对含碳酸钠为助溶剂的样品进行模型更新后，预测头孢他啶含量的平均偏差由8.1％变为2.3％，预测水分含量的平均偏差由2.2％变为0.3％；对含精氨酸为助溶剂样品，预测头孢他啶含量的平均偏差由7.0％变为1.9％，预测水分含量的平均偏差由0.6％变为0.3％，预测精氨酸含量的平均偏差由2.3％变为0.4％。经过更新的模型能够用于分析国内市场上常见的注射用头孢他啶中头孢他啶、水分及精氨酸含量。以 rT 作为模型更新指标较为合理，必要时可以参考新预测样品光谱与原模型训练集光谱的PC A得分图进行判断，该模型更新方法具有一定的普适性， rT ＝96.5％可以初步作为判断模型是否需要更新的阈值。%To find a more reasonable index to decide whether the universal quantitative NIR model needs to be updated and to de-velop a general method to update universal quantitative NIR models ,the quantitative models for testing ceftazidime ,water and arginine contents in ceftazidime for injection were taken as example .The study was performed by analyzing the similarity be-tween new sample spectra and the training set spectra of the original models .At first ,new samples of ceftazidime for injection were divided into five groups by cluster analysis .Then representative samples of each group were selected by sample selection strategy .Spectra of those samples were used to update the original quantitative models .The prediction deviation of the new ceftazidime powder injection samples by the models before and after updating was calculated .Decreasing the prediction deviation was regarded as the standard to decide if the updating was effective .At the same time ,the correlation coefficient of new sample spectra and reference sample spectra was defined as the index to study the general method for model updating .(Reference sample refers to training set sample) Finally ,the proposed method was validated by updating universal models for testing ceftazidime , water and arginine contents in ceftazidime powder injections .Results show that the correlation coefficient of new sample spectra and training set sample spectra of the original model was calculated within modeling wavelength range .It was proved that when correlation coefficient rT < 96.5% ,the model needs to be updated .Accordingly , rT = 96.5% was set as the threshold .The quantitative models were updated by the method mentioned above .As a result ,when testing ceftazidime for injection containing sodium carbonate using newly updated models ,the average predicting deviation of ceftazidime contents decreased from 8.1% to 2.3% .And the average predicting deviation of water contents decreased from 2.2% to 0.3% .Meanwhile ,with regard to sam-ples containing arginine using the updated models ,the average predicting deviation of ceftazidime contents decreased from 7.0%to 1.9% .The average predicting deviation of water contents decreased from 0.6% to 0.3% .And that of arginine contents de-creased from 2.3% to 0.4% .Conclusion:The newly updated models can be used for testing ceftazidime ,water and arginine contens in ceftazidime for injection samples of domestic market .It is reasonable to set rT as the index to decide whether the model needs updating .Moreover ,it is necessary to take PCA scores graph of new sample spectra and training set spectra of the original model into account .The proposed method for updating models can be used as a usual approach .And rT =96.5% can be set as the threshold to determine whether the model needs to be updated .
展开▼
