Establishment And Validation Of Type 2 Diabetic Retinopathy Prediction Model Based On Extreme Learning Machine

Objective:A common cause of blindness in people with type 2 diabetes(T2D)is diabetic retinopathy(DR).Early fundus examination with prompt treatment has been shown to prevent vision loss,but routine eye screening for people with diabetes poses significant economic and material challenges to existing healthcare systems.The purpose of this study is to establish a DR prediction model based on Extreme Learning Machine(ELM)and compare its performance with that of DR prediction models based on Support Vector Machines(SVM),K-Nearest Neighbour(KNN),Random Forest(RF),and Artificial Neural Network(ANN).Methods:Collect electronic medical records of inpatients from January 1,2020 to November 31,2021.The independent sample t-test and x~2test were used to screen out the significant index(p<0.05),and the extreme learning machine algorithm was used to develop a prediction model based on demographic data,blood test and urine test results.We used the following metrics to evaluate the performance of the model:(1)classification accuracy;(2)sensitivity;(3)specificity;(4)precision;(5)negative predictive value(NPV);(6)training time;(7)Area under the receiver operating characteristic curve(AUC).Results:The DR prediction models based on SVM and ELM are superior to DR prediction models based on ANN,KNN and RF in terms of accuracy,sensitivity,specificity,precision,NPV and AUC.The prediction model for diabetic retinopathy based on extreme learning machine performed best in terms of accuracy,precision,specificity,training time and AUC,which were 84.45%,83.93%,93.16%,1.24s and88.34%,respectively.The SVM-based DR prediction model performed best in sensitivity and NPV,with 70.82%and 85.60%,respectively.Conclusion:In this study,the model based on extreme learning machine has excellent performance in predicting type 2 diabetic retinopathy,which provides technical support for early screening of type 2 diabetic retinopathy.