Improving the Prediction of Chronic Kidney Disease Using the Least Number of Features

Abstract

High costs of dialysis during late stages of chronic kidney disease highlight the importance of CKD early detection. However, most people do not show any major signs or symptoms at early stages, so the disease is generally detected at later stages. CKD’s early detection can reduce mortality rate of the disease, control its progress during early stages, and lower the number of dialysis or transplantation patients. This paper aims to predict CKD using the least number of clinical and physiological tests. An intelligent artificial neural network (ANN) system was constructed to predict CKD, a dataset with 400 observations and 23 features was used, the ANN system accuracy is 99.5%, which agrees with the literature studies. Then the number of features was reduced in order to find the most related features from the dataset by applying genetic algorithm to the ANN, the algorithm reduced the number of features to three, while maintaining the ANN performance. In order to validate the importance of the deducted features, a k-means clustering algorithm was used, which is an unsupervised learning algorithm; the three features were able to detect the disease even without supervision, with an accuracy of 99.5%.