Purpose: Timely diagnosis of diabetic retinopathy (DR) can significantly improve the prognosis of patients. In this study, we established a prediction model by analyzing the relationship between diabetic retinopathy and related metabolic and biochemical indicators.
Methods: A total of 427 type 2 diabetes mellitus (T2DM) patients were selected from the datadryad website data. Logistic regression (MLR) was used to input layer variables of the model were screened. Then, Tan-Sigmoid was selected as the transfer function of the hidden layer node, and the linear function was used as the output layer function to establish the back propagation artificial neural network (BP-ANN) model. The model was applied to 183 patients with type 2 diabetes mellitus (T2DM) in our hospital to predict DR.
Results: A total of 167 patients (39.2%) with DR were obtained from the Datadryad database. Input variables were screened by MLR model, and it was concluded that the age, sex, albumin and creatinine, diabetes course were independently associated with the occurrence of DR. The above variables were used to establish BP-ANN model. The area under receiver operating characteristic curve (AUC) was significantly higher than that of MLR model (0.88 vs 0.74, P< 0.05), the probability threshold of the model was 0.3. Type 2 diabetes mellitus (T2DM) were selected in our hospital, including 92 patients with DR (50.2%). The above BP-ANN model was used to predict the incidence of DR, and the AUC area was significantly higher than that of the MLR model (0.77 vs 0.70, P< 0.05), the probability threshold was 0.7.
Conclusion: We established the BP-ANN model and applied it to diagnose DR. Taking diabetic course, age, sex, albumin and creatinine as the inputs of BP-ANN, the existence of DR could be well predicted. Meanwhile, the generalization ability of the model could be improved by selecting different probability thresholds in different ROC curves.

Keywords: diabetic retinopathy, type 2 diabetes, probability threshold, BP-ANN


中文翻译：

---

通过反向传播人工神经网络利用代谢和生化指标预测糖尿病视网膜病变

目的：及时诊断糖尿病视网膜病变（DR）可以显着改善患者的预后。在本研究中，我们通过分析糖尿病视网膜病变与相关代谢生化指标的关系，建立了预测模型。
方法：从datadryad网站数据中选取427例2型糖尿病（T2DM）患者。使用逻辑回归（MLR）对模型的输入层变量进行筛选。然后选择Tan-Sigmoid作为隐含层节点的传递函数，线性函数作为输出层函数，建立反向传播人工神经网络（BP-ANN）模型。该模型应用于我院183例2型糖尿病（T2DM）患者预测DR。
结果：从Datadryad数据库中总共获得了167名DR患者（39.2%）。通过MLR模型筛选输入变量，得出年龄、性别、白蛋白和肌酐、糖尿病病程与DR的发生独立相关。利用上述变量建立BP-ANN模型。受试者工作特征曲线下面积（AUC）显着高于MLR模型（0.88 vs 0.74，P<0.05），模型的概率阈值为0.3。选取我院2型糖尿病（T2DM）患者，其中DR患者92例（50.2%）。上述BP-ANN模型用于预测DR发生率，AUC面积显着高于MLR模型（0.77 vs 0.70，P<0.05），概率阈值为0.7。
结论：我们建立了BP-ANN模型并将其应用于DR的诊断。以糖尿病病程、年龄、性别、白蛋白和肌酐作为BP-ANN的输入，可以很好地预测DR的存在。同时，通过在不同的ROC曲线上选择不同的概率阈值，可以提高模型的泛化能力。

关键词:糖尿病视网膜病变, 2型糖尿病, 概率阈值, BP-ANN