Comparison of different machine learning methods in the diagnosis of chronic kidney failure (CKF)

Chronic Kidney Failure (CKF) is an important health problem that negatively affects the physical condition and social life of patients, and studies in the field of machine learning provide decision makers with important results in the field of nephrology. The objective of this study was to create different prediction models with machine learning methods for the diagnosis of CKF and to compare and evaluate the performances of these prediction models. The material of this study was 22 different variables related to 9,136 patients which patient information was used, who applied to Ondokuz Mayıs University Faculty of Medicine, Department of Nephrology between the years 2015-2021. In this study, the dataset for OneR, Naive Bayes, logistic regression, decision trees, random forest, nearest k-neighbor and support vector machine algorithms is divided into training datasets ranging from 50% to 90% analyzed by WEKA software and forecasting models were created. The success rates of these models were compared using the criteria of accuracy, error rate, precision, sensitivity, specificity, F measurement and Matthews correlation coefficient. The results of the study revealed that 54.28% of patients had CKF, while 45.72% of patients did not have this disease. The best performance in prediction models created with the entire data set was obtained by random forest and k-nearest k-neighbor algorithms (97.1%) with respect to accuracy rate. Besides, the best performance in all prediction models, where the data set is divided between 50-90% as the training data set and the test data set was obtained by the logistic regression algorithm with respect to the accuracy rate. For other performance measures, random forest and support vector machine algorithms were observed to be more successful than the other algorithms. The results of the receiver operating characteristic analysis introduced that, a cut-off point of 54.5 was determined for the patient age variable. The results of different machine learning methods used in the diagnosis of CKF were evaluated by separating the data set into training and test data sets at different rates. The results highlighted that it was appropriate to use the training data in the range of 60-80% and the test dataset in the range of 20-40%, the success rates in different performance criteria varied for the algorithms applied to the datasets of different sizes.