基于BP神经网络模型鉴别2型糖尿病肾病的认知图研究

Cognitive map study of type 2 diabetic nephropathy based on BP neural network model

目的以实验室指标建立诊断2型糖尿病肾病的BP神经网络模型并评价其诊断性能。方法收集重庆、贵州、四川五所三级医院2016年1月至2016年12月确诊为2型糖尿病肾病的患者，使用SPSS19．0和MATLAB2014a对患者89项信息进行单因素分析，提取有统计学意义的变量，以此分别构建logistic回归模型和BP神经网络模型，对比两种模型的诊断性能。结果477例2型糖尿病性肾病患者和449例对照组纳入模型分析，单因素分析结果显示差异有统计学意义42项信息。Logistic回归分类模型有12个变量纳入最佳回归方程。BP神经网络输入层、隐含层和输出层分别有42、15和1个节点。Logistic回归分类模型和BP神经网络模型（训练集，测试集）各自约登指数为0．76，0．89和0．83，对数据集的分类准确率分别为88．12％，94．24％和91．34％，ROC曲线下面积分别为0．95，0．98和0．96。结论本文建立的BP神经网络模型对2型糖尿病性肾病有很好的诊断辅助功能，但仍需进一步通过临床检验。

Objective A BP neural network model for diagnosing type 2 diabetic nephropathy based on laboratory tests was developed and evaluated. Methods Patients with type 2 diabetic nephropathy from 5 hospitals of Chongqing, Guizhou and Sichuan Provinces from January 2016 to December 2016 were collected in the study. Totally 89 parameters were analyzed by univariate analysis to identify significant variables by SPSS 19. 0 and MATLAB 2014a. The diagnostic performance of the two methods were compared. Results A total of 477 patients with type 2 diabetic nephropathy and 449 patients of control group were included. Univariate analysis showed that 42 variables had significant difference. Logistic regression analysis showed that 12 variables were included in the optimal regression equation. This BP neural network had 42 input layer nodes, 15 hidden layer nodes and 1 output layer nodes. The Youden index of logistic regression analysis and BP neural network （ training set and test set） were 0.76, 0.89 and 0.83. The accurately diagnosed were 88.12%, 94.24%, and 91.34%, the AUC were 0.95, 0.98, and 0.96. Conclusion A BP neural network model was developed, which has important accessory diagnostic value for diagnosis of type 2 diabetic nephropathy. But all these conclusions need further validation in clinic.