Construction data mining methods in the prediction of death in hemodialysis patients using support vector machine, neural network, logistic regression and decision tree

Abstract

Objectives: Chronic kidney disease (CKD) is one of the main causes of morbidity and mortality worldwide. Detecting survival modifiable factors could help in prioritizing the clinical care and offers a treatment decision-making for hemodialysis patients. The aim of this study was to develop the best predictive model to explain the predictors of death in Hemodialysis patients by data mining techniques.

Methods: In this study, we used a dataset included records of 857 dialysis patients. Thirty-one potential risk factors, that might be associated with death in dialysis patients, were selected. The performances of four classifiers of support vector machine, neural network, logistic regression and decision tree were compared in terms of sensitivity, specificity, total accuracy, positive likelihood ratio and negative likelihood ratio.

Results: The average total accuracy of all methods was over 61%; the greatest total accuracy belonged to logistic regression (0.71). Also, logistic regression produced the greatest specificity (0.72), sensitivity (0.69), positive likelihood ratio (2.48) and the lowest negative likelihood ratio (0.43).

Conclusions: Logistic regression had the best performance in comparison to other methods for predicting death among hemodialysis patients. According to this model female gender, increasing age at diagnosis, addiction, low Iron level, C-reactive protein positive and low urea reduction ratio (URR) were the main predictors of death in these patients.

Keywords: Data mining; Decision tree; Hemodialysis; Kidney failure; Logistic regression; Neural network; Support vector machine; Survival.