用近红外漫反射光谱无损检测血糖的初步研究

The Preliminary Study on Noninvasive Detection Using NIR Diffusion Reflectance Spectrum for Monitoring Blood Glucose

利用近红外漫反射光谱技术研究了无损定量监测血糖的方法。使用BRUKERVECTOR 2 2傅里叶变换光谱仪和GRACEⅡ型血糖仪对一组健康自愿受试者 (两名年轻男子和两名年轻女子 )测试后 ,获得了不同的状态条件下 2 6个样本的光谱及其血糖值 ,选择一部分有代表性的样本作为建立模型使用 ,而被预测的样本来自不参加建模的数据。选择不同的谱区、预处理方法对数据进行处理得到以下结果 :1)谱区选在 90 0 0～12 0 0 0cm- 1 选择MIN MAX归一化预处理方法建模、预测 ,预测结果差值在 36mg·dL- 1 以上 ,选择平滑和二阶导数 ,则建模结果≤ 16mg·dL- 1 ,预测结果≤ 2 5mg·dL- 1 ;另选谱区 4 0 0 0～ 5 0 0 0cm- 1 ,选择平滑和一阶导数 ,建模与预测结果的差值在 2 5mg·dL- 1 以上。 2 )选择 4 0 0 0～ 90 0 0cm- 1 谱区 ,经平滑和二阶导数处理并在此区建模并预测 ,其中建模中预测差值≤ 15mg·dL- 1 ,预测≤ 31mg·dL- 1 ;由OPU 3 0 1自动挑选的谱区6 0 0 0～ 75 0 0cm- 1 和 4 2 0 0～ 4 70 0cm- 1 ,采用平滑、一阶导数和矢量归一化在此区建模和预测 ,其中建模中预测差值≤ 11mg·dL- 1 ,预测≤ 2 2mg·dL- 1 。 3)采用个体自我建模方法 ,在 90 0 0～ 12 0 0 0cm- 1 选择平滑和二阶导数进行预处理 ,建模结果≤ 15mg?

The noninvasive detection method using near infrared diffusion reflectance spectrum for monitoring blood glucose was described. A group of healthy volunteers(2 young male and 2 young female) was tested twenty six times under changing of both the blood glucose concentration and spectrum region during various states. In order to set up a model ,the set was chosen from these samples ,the set of prediction was not belong to the set of the model. The results from different selected spectrum regions and reprocessing methods show as follows:1) In the spectrum region of 9 000- 12 000 cm -1 using MIN MAX normalization reprocessing method to set up a model and predict, the errors of prediction were more than 36 mg·dL -1 ; the model errors were under 16 mg·dL -1 and predictive errors were less than 25 mg·dL -1 by using smoothing and second derivatives in the same region; in the another region of 4 000-5 000 cm -1 with smoothing and first derivatives both error of model and prediction were more than 25 mg·dL -1 . 2)In the region of 4 000-9 000 cm -1 with smoothing and second derivatives, the errors from model were less than 15 mg·dL -1 , the errors of prediction were less than 31 mg·dL -1 ; selecting the spectrum regions of 6 100-7 500 cm -1 and 4 200-4 700 cm -1 by a quantitative software OPU VERSION 3 01 with smoothing , first derivatives and vector normalization ,the errors of model were no more than 11 mg·dL -1 , the errors of predication were no more than 22 mg·dL -1 .3)Choosing the sample from the subject itself to establish the model in 9 000- 12 000 cm -1 with smoothing and second derivatives, the errors of model and predication were less than 15 mg·dL -1 and 11 mg·dL -1 respectively. The results showed that the method was feasible to monitoring the blood glucose concentration by selecting an appropriate spectrum region and reprocessing methods with partial least square algorithm which set up a mathematical model. Excepting the number of sample, selected spectrum region and reprocessing methods, the other influencing factors such as the variant from the subject were discussed also.