Calibration standardization algorithm for partial least-squares regression: application to the determination of physiological levels of glucose by near-infrared spectroscopy.
Calibration standardization algorithm for partial least-squares regression: application to the determination of physiological levels of glucose by near-infrared spectroscopy.
复制标题
偏最小二乘回归的校准标准化算法：应用于近红外光谱法测定葡萄糖的生理水平。
DOI：
10.1021/ac020023r
复制
发表时间：
2002
影响因子：
7.4
通讯作者：
Arnold,MarkA
中科院分区：
文献类型：
--
作者：
Zhang,Lin;Small,GaryW;Arnold,MarkA
描述了近红外 (near-IR) 光谱学的校准标准化方法，用于更新部分最小二乘校准模型以考虑仪器响应的变化。引导模型重新优化 (GMR) 算法使用八个样本的传输集来表征新响应，并使用先前采集的光谱数据库来开发原始校准模型。传输集中的样本无需在旧仪器条件下进行测量，从而使算法与随时间变化的样本兼容。包含传输集的光谱用于指导迭代优化过程，该过程（1）从原始数据库中找到最佳样本子集以用于计算更新模型，以及（2）找到应用于光谱分辨率元素的最佳权重集，以便最大限度地减少仪器变化对计算模型的影响。优化依赖于交替网格搜索和逐步添加/删除步骤。通过使用组合区域近红外光谱来评估该算法，以确定磷酸盐缓冲液中含有牛血清白蛋白和三醋精的合成生物基质中葡萄糖的生理水平。本研究检验了更新校准以解释傅里叶变换光谱仪响应在四到六年内的变化的能力。 1994 年和 1996 年收集的单独光谱数据库与 2000 年收集的光谱的传输集和单独测试集一起使用。利用 1994 年数据库，标准化算法实现了 2000 年测试集的 0.69 mM 预测标准误差 (SEP)。当使用未标准化的原始 1994 年校准模型时，这与 >2 mM 的 SEP 值相比是有利的。使用 1996 年数据库 (SEP = 0.70 mM) 进行标准化后，2000 年测试集的预测性能也得到了类似的改进。
Calibration standardization methodology for near-infrared (near-IR) spectroscopy is described for updating a partial least-squares calibration model to take into account changes in instrumental response. The guided model reoptimization (GMR) algorithm uses a transfer set of eight samples to characterize the new response and a database of previously acquired spectra used to develop the original calibration model. The samples in the transfer set need not have been measured under the old instrumental conditions, making the algorithm compatible with samples that change over time. The spectra comprising the transfer set are used to guide an iterative optimization procedure that (1) finds an optimal subset of samples from the original database to use in computing the updated model and (2) finds an optimal set of weights to apply to the spectral resolution elements in order to minimize the effects of instrumental changes on the computed model. The optimization relies on an alternating grid search and stepwise addition/deletion steps. The algorithm is evaluated through the use of combination region near-IR spectra to determine physiological levels of glucose in a synthetic biological matrix containing bovine serum albumin and triacetin in phosphate buffer. The ability to update a calibration to account for changes in the response of a Fourier transform spectrometer over four to six years is examined in this study. Separate spectral databases collected in 1994 and 1996 are used with a transfer set and separate test set of spectra collected in 2000. With the 1994 database, the standardization algorithm achieves a standard error of prediction (SEP) of 0.69 mM for the 2000 test set. This compares favorably to SEP values >2 mM when the original 1994 calibration model is used without standardization. A similar improvement in the prediction performance of the 2000 test set is obtained after standardization with the 1996 database (SEP = 0.70 mM).