基于近红外光谱的SG-MSC-MC-UVE-PLS算法在全血血红蛋白浓度检测中的应用

为提高全血血红蛋白浓度预测模型的预测精度，基于近红外光谱分析，首先对原始全血透射光谱数据分别进行均值中心化、标准化、标准正态变量变换(SNV)、多元散射校正(MSC)以及Savitzky-Golay（SG）卷积平滑结合MSC的预处理操作，最终选择预处理效果最好的SG-MSC方法作为数据预处理方法，其最大相关系数达到0.944 1。对SG平滑的平滑窗口宽度进行讨论，找出平滑效果最好的窗口宽度为27。数据预处理消除了全血吸收光谱的基线失真，提高了全血吸收光谱数据的信噪比。将190个样本（190个血红蛋白浓度对应的透射光谱数据）分为具有相近血红蛋白浓度分布的校正集和测试集，其中校正集为143个样本（对应血红蛋白浓度分布为10.6～17.3 g·dL-1），测试集为47个样本（对应血红蛋白浓度分布为10.3～17.3 g·dL-1），确保建立模型的适用性。对校正集数据预处理后利用蒙特卡洛无信息变量消除（MC-UVE）方法对其进行波长变量选择，剔除含信息量少的波长点，提高含信息量多的波长占比。设置蒙特卡洛迭代次数为1 000，最终从全血吸收光谱的700个波长变量中筛选出191个波长变量用于建立全血血红蛋白浓度偏最小二乘（PLS）回归模型。对比分析原始全血透射光谱全谱PLS模型、原始全血吸收光谱全谱PLS模型、预处理全血吸收光谱全谱PLS模型、SG-MSC-MC-UVE-PLS模型以及已有二阶导数PLS模型的模型效果，表明基于SG-MSC-MC-UVE-PLS算法的全血血红蛋白浓度预测模型效果较其他模型效果更优，预测相关系数由0.676 3提高到0.979 1，预测集均方根误差由0.898 1减小到0.220 3，最大绝对误差由2.426 1减小到0.411 2。同时，利用MC-UVE方法进行波长变量选择，在保证预测精度的前提下，筛选出建模的波长个数更少，有利于提高模型计算效率。研究结果表明，SG-MSC-MC-UVE-PLS方法能够提高全血吸收光谱信号的信噪比，简化模型结构，提高模型的预测精度和计算效率，对推动血红蛋白浓度检测技术的发展具有进步意义。

Application of SG-MSC-MC-UVE-PLS Algorithm in Whole Blood Hemoglobin Concentration Detection Based on Near Infrared Spectroscopy

In order to improve the accuracy of the whole blood hemoglobin (Hb) concentration prediction model, the original whole blood transmission spectrum signals were first preprocessed by using centering, auto scaling, standard normal variate (SNV), multiplicative scatter correction (MSC), and Savitzky-Golay (SG) smoothing combined with MSC. And the best preprocessing effect was obtained with a R2 value of 0.9441 by using SG smoothing combined with MSC. The width of the SG smoothing window was discussed, and the optimal width is 27.The baseline shift of the whole blood absorbance signals was eliminated, and the signal-to-noise ratio was improved after data preprocessing. The 190 samples were divided into a calibration set (corresponding Hb concentrations from 10.6 to 17.3 g·dL-1) of 143 samples and a validation set (corresponding Hb concentrations from 10.3 to 17.3 g·dL-1) of 47 samples. The model’s applicability was ensured when two sets have a similar distribution and range of Hb concentrations. And then, the Monte Carlo uninformative variable elimination (MC-UVE) was used to select the informative wavelength, which simplified the model structure and increased the proportion of useful wavelengths. When the Monte Carlo iteration number was 1000, 191 wavelength points were selected from the 700 wavelengths of the whole blood absorbance spectrum to build the whole blood Hb concentration partial least squares (PLS) model. Finally, a comparison was performed among the model based on the original whole blood transmission spectrum, the model based on the whole blood absorbance spectrum, the SG-MSC-PLS model, the SG-MSC-MC-UVE-PLS model and an existing model. In addition to this, the number of selected wavelengths based on MC-UVE was much smaller than the total number, but the predictive effect was much better, which was beneficial to improve the calculation efficiency of the model. The results indicate that the SG-MSC-MC-UVE-PLS method effectively increases the signal-to-noise ratio of the whole blood absorption spectrum signal and simplifies the model. Besides, our procedure’s prediction accuracy and calculation efficiency of the model was improved by our procedure, which has reference significance for the development of hemoglobin concentration detection technology.