【作者】 王龙；

【导师】 骆清铭；

【作者基本信息】 华中科技大学 ， 生物医学工程， 2006， 硕士

【摘要】 近红外区域按ASTM定义是指波长在750~2500nm范围内的电磁波,可以对特定的官能团进行定量的测量,如C-H,O-H,N-H和C=O等基团。近红外无创血糖检测方法是将一束近红外光通过人体一部分血管区域,从得到的光谱信息中提取与血糖浓度相关联的变化信息。其关键在于收集具有足够高信噪比的光谱信息以辨别微弱的葡萄糖吸收以及用复杂的统计学方法建立血糖浓度的预测模型。本文主要在光谱数据处理方法和奇异样本点剔除方法两个方面展开研究。根据葡萄糖分子在近红外倍频区域的吸收,选择信号波长为1610nm,为去除水分、血色素等干扰成分的影响,选择1200nm,1350nm作为两个参考波长。选择浓度间隔为100mg/dL,范围为0-500mg/dL的葡萄糖水溶液模型实验数据建模,采用偏最小二乘回归建立预测模型,对预测集样品中的葡萄糖浓度进行预测,结果所得预测均方根误差(RMSEP)为17.08mg/dL,预测值与参考值的相关系数为0.998,而采用吸光度差值比率法所建预测模型的RMSEP为54.94 mg/dL,相应的相关系数仅为0.949。结果表明:偏最小二乘回归所建模型用于近红外三波长葡萄糖浓度的预测,所得预测误差较小,初步验证了系统的可行性。另一方面,本文讨论了有关奇异样本点剔除的方法,选择浓度间隔为20mg/dL时,范围为0-300mg/dL的葡萄糖水溶液模型实验数据建立预测模型。奇异样本剔除前,所得偏最小二乘回归预测模型的相关系数为0.569,RMSEP为75.86mg/dL,奇异样本剔除后,所得相关系数为0.959,RMSEP为23.22mg/dL。结果表明:运用合适的奇异样本点的剔除方法,并与偏最小二乘回归算法相结合,可以达到模型优化的目的。更多还原

【Abstract】 The Near Infrared Region extended from 750 to 2500nm and can be used for quantitative measurement of organic functional groups, especially C-H,O-H,N-H and C=O. The near-infrared continuous non-invasive method used a beam of NIR light through the blood vesselregion of human, and extracted the correlating information of blood glucose from the spectrum. The sticking point was how to improve the ratio of signal to noise in order to distinguish the very low glucose absorption and utilize complex statistics theory to establish the prediction model of blood glucose. In this thesis, we discussed the algorithms of processing experiment data based and the methods of outlier detection.According to the absorption peak of glucose molecule in near-infrared overtone band, 1610nm was chose as the signal wavelength while 1200nm and 1350nm were chose as the reference wavelengths in order to eliminate the interfere substance’s effect. In the thesis,the modeling methods had been discussed to determine glucose concentrations based on near-infrared three wavelengths system in aqueous glucose solutions. The model experiment that glucose concentration interval was 100mg/dL and the glucose concentration range was 0-500mg/dL was set up.The model had been established based on partial least-square algorithm. When it was used to predict the glucose contents of validation set, the root mean square error of prediction (RMSEP) equaled 17.08mg/dL and the correlation coefficient between prediction values and reference values attained 0.998; while using logarithmic ratio of absorbency difference method to set up prediction model, the RMSEP reached54.94mg/dL and the correlation coefficient equaled 0.949 merely. The results suggested that the model based on PLS algorithm has lesser RMSEP, it also validates the system’s feasibility primarilyOn the other hand, outlier detection methods was discussed in this paper.when used to process the raw experiment data before the PLS model was set up, it can make the prediction model optimize and get smaller RMSEP .The model experiment that glucose concentration interval was 20mg/dL and the glucose concentration range was 0-300mg/dL was set up. The model had been established based on partial least-square algorithm. It was used to predict the glucose contents, the R and RMSEP were 0.569 and 75.86mg/dL before the outlier samples were eliminated. After the outlier samples were eliminated, the R and RMSEP were 0.959 and 23.22mg/dL respectively. The results suggested that optimized prediction model could be obtained when PLS algorithm and outlier detection methods were combined更多还原