Comparative Study of Chronic Kidney Disease Predictor Performance Given Insufficient Training Dataset

Abstract

This study compares the performance of Logistic Regression and Classification and Regression Tree model implementations in predicting chronic kidney disease outcomes from predictor variables, given insufficient training data. Imputation of missing data was performed using a technique based on k-nearest neighbours. The dataset was arbitrarily split into 10 % training set and 90 % test set to simulate a dearth of training data. Accuracy was mainly considered for the quantitative performance assessment together with ROC curves, area under the ROC curve values and confusion matrix pairs. Validation of the results was done using a shuffled 5-fold cross-validation procedure. Logistic regression produced an average accuracy of about 99 % compared to about 97 % the decision tree produced.