基于人体血清荧光光谱和改进变量选择方法对血糖浓度测定的研究

利用人体血清240 nm激发波长荧光光谱在220～900 nm波段针对血糖浓度进行建模分析。在模拟退火算法和偏最小二乘算法的基础上进行建模波长变量筛选方法的改进。基于变量入选模型统计频率和无关变量消除法，分别进行了波长的初选和精选过程。同时，加入了主成分数自适应特性等加快收敛速度、减小计算量的措施。该模型对比了线性、三次样条函数、高斯函数作为偏最小二乘法插入基函数的情况下，分别对原始光谱、Daubechies小波分解第三层和第四层细节信号光谱建模。结果显示，该模型避免了主成分参数尝试导致的时间成本，在参数自适应的过程中很快趋于稳定并得到对应条件下的最佳建模主成分数。其波长变量筛选能力使得对独立样本的分析预测效果有了显著的提升，最佳建模预测效果达到0.25 mmol·L-1，达到了血糖检测的医用要求。加入了非线性建模条件使得模型有明显的改善，基于样条函数的效果总体最好，其次为基于高斯函数的情况。对原始光谱进行小波分解，得到的细节信号光谱建模效果更为可观。总体而言第四层细节信号光谱建模效果略优于第三层细节信号光谱。模型筛选波长变量的频率意味着在所给实验条件下的血糖浓度信息在不同波段的分布情况，这为血糖在血清成分中的物理化学特性提供了一定程度上的统计解释。

Research on the Determination of Glucose Based on Human Serum Fluorescence Spectrum and Improved Variable Selection Strategy

Human serum fluorescence spectrum within the range of 220~900 nm excited by 240 nm excitation wavelength is studied and analyzed in the modeling for the detection of human serum glucose concentration.Wavelength variable selection strategy was improved on the basis of simulated annealing algorithm and partial least square algorithm.According to the frequen-cy of wavelengths selected in modeling and uninformative variable elimination method,this paper executed a rough and hand-picked process for wavelength variable selection which speeded up the convergence rate and reduced the quantity of calculation such as introducing the self-adaptive property for the number of principle components.Basis interpolation functions for partial least square algorithm such as linear,cubic spline function and Gaussian function as well as the original spectra and the 3rd,the 4th detail signal decomposed by Daubechies wavelet were studied and compared in the modeling process.The result shows that the enhancement avoids the time cost by parameter setting attempts,the parameter gradually becomes stable in the calculation process and the best determination of the principle components is found.The prediction and analysis ability for independent sam-ples have been a significant improvement with the new strategy of wavelength variable selection.The minimum least square error for prediction is 0.25 mmol·L-1 in modeling results,which is up to most clinical standards for human glucose level detection. The model is apparently improved by adding the nonlinear condition,of which the best result is based on the spline function,and the second is on the gauss function.Original fluorescence spectra are decomposed and produce a better result for modeling.The 4th detail signal spectra are better than the 3rd detail signal on the whole.Given the experimental condition,the frequency of wavelength selection is meant for the distribution of glucose concentration information,which provides the statistical interpreta-tion for the physicochemical characteristics of glucose in human serum to some extent.