Multivariable Logistic Regression And Back Propagation Artificial Neural Network To Predict Diabetic Retinopathy

Litong Yao; Yifan Zhong; Jingyang Wu; Guisen Zhang; Lei Chen; Peng Guan; Desheng Huang; Lei Liu

doi:10.2147/DMSO.S219842

Multivariable Logistic Regression And Back Propagation Artificial Neural Network To Predict Diabetic Retinopathy

Diabetes Metab Syndr Obes. 2019 Sep 25:12:1943-1951. doi: 10.2147/DMSO.S219842. eCollection 2019.

Authors

Litong Yao^#¹, Yifan Zhong^#², Jingyang Wu², Guisen Zhang³, Lei Chen², Peng Guan⁴, Desheng Huang^{4

5}, Lei Liu^{2

6}

Affiliations

¹ Department of Breast Surgery, The First Affiliated Hospital of China Medical University, Shenyang 110001, People's Republic of China.
² Department of Ophthalmology, The First Affiliated Hospital of China Medical University, Shenyang 110001, People's Republic of China.
³ Department of Ophthalmology, Hohhot Chao Ju Eye Hospital, Hohhot 010000, People's Republic of China.
⁴ Department of Epidemiology, School of Public Health, China Medical University, Shenyang 110122, People's Republic of China.
⁵ Department of Mathematics, School of Fundamental Sciences, China Medical University, Shenyang 110122, People's Republic of China.
⁶ Department of Public Service, The First Affiliated Hospital of China Medical University, Shenyang 110001, People's Republic of China.

^# Contributed equally.

Abstract

Background: Monitoring and prediction of diabetic retinopathy (DR) is necessary in patients with diabetes for early discovery and timely treatment of disease. We aimed to analyze the association between DR and biochemical and metabolic parameters, and develop a predictive model for DR.

Methods: A total of 530 Chinese residents including 423 with type 2 diabetes (T2D) aged 18 years or older participated in this study. The association between DR and biochemical and metabolic parameters was analyzed by the univariate and multivariable logistic regression (MLR). According to the MLR results, we developed a back propagation artificial neural network (BP-ANN) model by selecting tan-sigmoid as the transfer function of the hidden layers nodes, and pure-line of the output layer nodes, with training goal of 0.5×10^-5.

Results: There were 51 (9.6%) diabetic participants with DR. After univariate and MLR analysis, duration of diabetes, waist to hip ratio, HbA₁c and family history of diabetes were independently associated with the presence of DR (all P < 0.05). Based on these parameters, the area under the receiver operating characteristic (ROC) curve for the BP-ANN model was significantly higher than that by MLR (0.84 vs. 0.77, P < 0.001).

Conclusion: Our evaluation demonstrated the potential role of BP-ANN model to identify DR in screening practice. The presence of DR was well predictable using the proposed BP-ANN model based on four related parameters (duration of diabetes, waist to hip ratio, HbA₁c and family history of diabetes).

Keywords: BP-ANN; diabetic retinopathy; regression; type 2 diabetes.