Predictive Modeling of Diabetic Kidney Disease using Random Forest Algorithm along with Features Selection

At present, the number of diabetes mellitus patients in China ranks first in the world, and diabetic kidney disease is the most common disease in complications. Therefore, it is necessary to establish a predictive model for early diagnosis of diabetic kidney disease. The model predicts the risk of diabetic kidney disease in the general Asian population, and recognizes high-risk groups, then warns the onset of diabetes. The data were obtained from the electronic medical record of patients in Beijing Pinggu Hospital. Twenty-nine initial candidate indicators including age, ALB, and A/C were selected. The random forest algorithm was used to predict diabetic kidney disease, and the classification accuracy was 89.831%. The importance weight ratio of each factor index was also given, Microalbuminuria (ALB), albumin-to-creatinine ratio (A/C), serum creatinine (SCr), Serum albumin (umALB), and blood urea nitrogen (BUN) accounted for a relatively high proportion of the weight of the characteristic variables. So these five indicators can be the primary indicators of our classification prediction, and the accuracy can reach 87.453%. Some other typical classification algorithms, liking KNN, logistic regression, and decision tree, were compared to classify and predict diabetic kidney disease, and precision recall f1-score and area AUC under ROC curve were used to evaluate these models. By experiments, random forest model was better than other algorithm models on both the classification accuracy and the evaluation indicators. The results can be applied to the screening of patients with high risk of diabetic kidney disease and the guidance of risk intervention measures. Consequently, the detection rate of undiagnosed diabetic kidney disease in the population can be improved, and the prevention effect of diabetic kidney disease can be enhanced as well.