An Interpretable Machine Learning Technique for Chronic Kidney Disease Diagnosis Using Clinical Data

Abstract

Early diagnosis of chronic kidney disease (CKD) plays a key role in treatment and improvement of the patient's health. In this research paper, we introduce a machine learning approach for early diagnosis of CKD with the use of structured clinical data obtained from the UCI Machine Learning Repository. Several algorithms, such as logistic regression, support vector machine (SVM), random forest, XGBoost, light gradient boosting machine (LightGBM), and CatBoost, are applied in one preprocessing pipeline. The developed approach consists of data cleaning, filling in missing values, feature transformation, and randomized search for hyperparameters. Model evaluation metrics include accuracy, precision, recall, F1-score, area under the curve (ROC-AUC), and area under the curve (PR-AUC). Stratified cross-validation was used for unbiased assessment of model performance. Our experiments reveal excellent prediction performance for all models, as indicated by the F1-scores of 0.98-0.99. SVM shows the best prediction results with a mean F1-score of 0.9870 and a low standard deviation. To confirm the reliability of the proposed framework, another experiment was carried out with random permutation of class labels, which resulted in a significant reduction in performance, thus validating that the learning algorithms have learned important correlations and not any random correlations. The study shows that the classifiers have the ability to discriminate among the different groups, where the variables such as hemoglobin, packed cell volume, and specific gravity of urine exhibit very good discriminatory abilities.