Comparative Analysis of Machine Learning Algorithms for Chronic Kidney Disease Prediction.

In recent days, chronic kidney disease (CKD) has been recognized as one of the most significant health problems globally. The defining feature of CKD is a progressive deterioration in renal function over time. Since kidney damage develops slowly over a long period of time, early detection and appropriate treatment may be able to save the lives of many. Machine learning classifier algorithms have emerged as a reliable tool to identify the disease at its early stages, providing a means to intervene and manage it sooner than other methods. In this paper, the performance of 10 models is evaluated on the dataset of CKD collected from the UCI ML repository for the classification of CKD. The training data in this study was augmented by applying the SMOTE technique and Gaussian noise. In case of missing values handling, for numerical and categorical variables KNN imputation and mode imputation for features were utilized respectively. Combining the filter, wrapper and embedded feature selection strategies led to the identification of the most important 13 features. Extra Tree Classifier, XGBoost, Gradient Boosting and Random Forest performed better than other algorithms with an accuracy of 99.17%. When compared to the other nine methods, Extra Tree Classifier performed extremely well in case of precision, recall and F1 score. For this proposed approach, the error rate and training time were all comparatively low at 0.0083, and 0.0787 seconds respectively. This paper illustrates the performance comparison of ten different machine learning (ML) algorithms and the importance of feature selection for predicting CKD.