Predicting the risk of chronic kidney disease using Machine Learning Algorithms

Abstract

Abstract Background and Objective: Chronic kidney disease (CKD) is among the most severe diseases in the modern world adversely affecting human life. Various risk factors, such as age, sex, diabetes, and hypertension, predispose to the occurrence of CKD. The aim of this study was to determine the predictors of CKD using machine learning algorithms. Materials and Methods: The present study was conducted on the data from the Ravansar Non-Communicable Disease (RaNCD) cohort. At the end of 5 years of follow-up, the number of participants was 10065 cases, 81 (0.8%) of whom were excluded after sub-processing, and 9984 (98.92%) subjects were finally included in the study. Different machine learning algorithms were used to analyze the data, and the performance of each model was evaluated by calculating accuracy, sensitivity, specificity, and area under the curve (AUC). The final model was used to identify the most important predictors of CKD. Results: The Generalized Linear Model (GLM) was selected as the final model with the highest sensitivity and accuracy (AUC =97%). According to this model, the most important predictors of CKD were identified to be SC=1.0, Wc=0.591, Rt=0.687, age=0.401, SGPT=0.334, TG=0.334, MCH=0.327, MCV=0.327, BFM=0.306, and HDLC=0.276. Also, the variables of SC, AIP, gender, and SGPT were most important in predicting CKD. Based on the final model, sodium, SGOT, and DBP were the most important predictors that contradicted with None-CKD patients. Conclusion: Based on our results, the GLM model delivered the most proficient performance in predicting CKD by correctly identifying all patients. In this model, serum creatinine level obtained the highest weight and, therefore, was the most important predictor of CKD.