基于离散三维荧光光谱的糖尿病识别方法研究

Diabetes Recognition Method Based on Discrete Three-Dimensional Fluorescence Spectrum

基于皮肤荧光光谱技术的糖基化终产物(AGE)检测方法被广泛用于糖尿病及其并发症的检测评估,然而生物组织体内特异性的吸收、散射特性以及多种荧光成分对于检测会产生干扰。本研究通过分析人体皮肤组织常见荧光成分的三维荧光光谱,确定皮肤组织荧光光谱测量的最佳激发波长组合,搭建集成组织漫反射光谱测量模块的离散三维荧光光谱测量系统,研究基于扩散理论的组织生理参数提取方法以及基于高斯多峰拟合的离散三维荧光光谱分离方法。在此基础上,开展临床社区队列研究,结果表明糖尿病组的黑色素、脱氧血红蛋白相对浓度均值低于正常对照组,而500 nm处约化散射系数高于正常对照组,差异有统计学意义,氧合血红蛋白相对浓度无明显差异。基于高斯多峰拟合,每位受试者共计得到78个荧光特征,剔除差异不明显的特征(p>0.1),最终得到50个荧光特征。汇总有差异的组织生理参数和荧光特征,采用Logistic回归分析建立糖尿病筛查模型,受试者工作特征(ROC)曲线分析结果显示,该方法用于糖尿病识别时,预测训练集ROC曲线下面积为0.793,预测测试集的面积为0.799,而单波长皮肤荧光(Sf;)的面积为0.731,结果表明相比单波长荧光,模型具有更好的诊断价值。

The advanced glycation end(AGE) product detection based on skin fluorescence spectroscopy is widely used in the screening and evaluation of diabetes and its complications. However, this method is limited by the specific absorption and scattering characteristics in biological tissues and a variety of fluorescent components. In this study, the optimal excitation wavelength combination for measuring skin tissue fluorescence spectra was determined by analyzing the three-dimensional fluorescence spectra of common fluorescent components in human skin tissues. A discrete three-dimensional fluorescence spectrum measurement system was built that integrated the measurement module of tissue diffuse reflectance spectra. The methods of extracting tissue physiological parameters under the diffusion theory and separating discrete three-dimensional fluorescence spectra by multi-peak Gaussian fitting were studied. A clinical community cohort study was then carried out. The results showed that the mean relative concentrations of melanin and deoxyhemoglobin in the diabetic group were lower than those in the normal control group, while the reduced scattering coefficient at 500 nm was higher than that in the normal control group. These differences were statistically significant. No significant difference was found in the relative concentration of oxyhemoglobin. Under the multi-peak Gaussian fitting, 78 fluorescence features were obtained for each subject, and 50 fluorescence features remained after features with insignificant differences(p>0.1) were eliminated. Differential tissue physiological parameters and fluorescence characteristics were pooled, and a diabetes screening model was built by the logistic regression analysis. According to the analysis results of the receiver operating characteristic(ROC) curve, when the proposed method was applied to diabetes screening, the area under the ROC curve of the prediction training set was 0.793, and that of the prediction test set was 0.799. In contrast, that of the single-wavelength skin fluorescence(Sf;) was 0.731, which indicated that the proposed model had a better diagnostic value than that of the single-wavelength fluorescence.