Improvement of prognostic models for ESRD mortality by the bootstrap method with random hot deck imputation

Prognostic models for end-stage renal disease (ESRD) have been researched extensively as an increasing prevalence internationally. Different machine learning and statistic algorithms for the models were proposed in studies corresponding to different medical datasets including a quantity of missing values for optimal outcomes. We approached this issue by applying stepwise logistic regression, ANN, and SVM algorithms to an ESRD dataset after case deletion and calculated areas under ROC curves of three algorithms as comparisons, resulting in 0.757, 0.664 and 0.704, respectively. The random hot deck, oversampling, and bootstrap methods were employed in data preprocessing to compensate the minor mortality. Afterward, average AUC of three algorithms approximated 0.90 (p<;0.02, unpaired t-test). As a result, the mentioned strategies dealing with bias medical data may ameliorate prognostic ESRD models in clinic.