Comparison of regression tree data mining methods for prediction of mortality in head injury

Abstract

With this research, we sought to examine the performance of six different regression tree data mining methods to predict mortality in head injury. Using a data set consisting of 1603 head injury cases, we assessed the performance of: the Classification and Regression Trees (CART) method; the Chi-squared Automatic Interaction Detector (CHAID) method; the Exhaustive CHAID (E-CHAID) method; the Quick, Unbiased, Efficient Statistical Tree (QUEST) method; the Random Forest Regression and Classification (RFRC) method; and the Boosted Tree Classifiers and Regression (BTCR) method, in each case based on sensitivity, specificity, positive/negative predictive, and accuracy rates. Next, we compared their areas under the (Receiver Operating Characteristic) ROC curves. Finally, we examined whether they could be grouped in meaningful clusters with hierarchical cluster analysis. Areas under the ROC curves of regression tree data mining methods ranged from 0.801 to 0.954 ( p < 0.001 for all). In predicting mortality in head injury under the ROC curve, the BTCR method achieved both the highest area (0.954) and accuracy rate (93.0%), while the CART method achieved both the lowest area (0.801) and accuracy rate (91.1%). All of the regression tree data mining methods were clustered in the same grouping, but the BTCR method was at the origin of the cluster while the CART and QUEST methods produced results that were least like the others. The BTCR, demonstrating a 93.0% accuracy rate and showing statistically significantly differences from the others, may be a helpful tool in medical decision-making for predicting mortality in head injury.