光谱学与光谱分析, 2014, 34(5): 1327, 网络出版: 2014-05-01
Research on Noninvasive Risk Evaluation of Diabetes Mellitus Based on Neural Network Pattern Recognition
1中国科学院安徽光学精密机械研究所, 安徽 合肥230031
2皖江新兴产业技术发展中心, 安徽 铜陵244061
医用光学 神经网络 模式识别 晚期糖基化终末产物 荧光光谱 糖尿病 Medical optics Neural network Pattern recognition Advanced glycation end products Fluorescence spectrum
晚期糖基化终末产物在人体皮肤组织中的浓度与高血糖水平密切相关, 且具有自发荧光特性。 使用自行研制的光学无创检测装置对人体皮肤组织的自体荧光光谱进行测量, 建立神经网络模式识别模型对检测对象患有糖尿病的可能性进行风险评估。 利用检测装置获取荧光光谱后对光谱数据进行主成分分析, 选取前4个主成分作为光谱的特征, 建立一个具有4个输入层节点、 6个隐层节点、 1个输出节点的神经网络模式识别模型。 选取在安徽省立医院测量的487例对象数据训练该模型, 以70%数据作为训练集, 15%数据作为验证集, 15%数据作为测试集。 模型可给出测试对象罹患糖尿病的风险, 或直接给出是否糖尿病的判断。 结果显示该模型的受试者工作特性曲线的线下面积为0.81, 标准误差为0.02; 以模型输出0.5为分类界限时的敏感性为72.4%, 特异性为77.6%, 整体准确率为74.9%。 本研究首次提出使用皮肤组织自体荧光结合神经网络模式识别模型对糖尿病进行无创风险评估, 实验结果表明该方法的筛查效果优于目前常用的空腹静脉血浆血糖值法和糖化血红蛋白法。
Advanced glycation end products (AGEs) are highly associated with hyperglycemia in human skin tissue, and they also have the autofluorescence characteristic. A self-developed optical noninvasive detection device was used to measure the autofluorescence in human skin tissue, and then a neural network pattern recognition model was used to assess the risk of diabetes mellitus of the subject under survey. After the fluorescence spectra were acquired and processed with principal component analysis, four of the leading principal components were chosen to represent a whole spectrum. The established neural network pattern recognition model has 4 input nodes, 6 hidden nodes and 1 output node. A dataset consisting of 487 cases collected in Anhui Provincial Hospital was used to train the model. Seventy percent cases were used as the training set, 15% as the validation set and 15% as the test set. The model can output subject’s risk of diabetes mellitus, or a dichotomous judgment. Receiver operating characteristic curve can be drawn with the area under curve of 0.81, with standard error of 0.02. When using 0.5 as the threshold between diabetes mellitus and non-diabetes mellitus, the sensitivity and specificity of this model is 72.4% and 77.6% respectively, and the overall accuracy is 74.9%. The method using human skin autofluorescence spectrum combined with neural network pattern recognition model is proposed for the first time, and the results show that this method has a better screening effect compared with currently used fasting plasma glucose and HbA1c.