Comparison of Prediction Models for Mortality Related to Injuries from Road Traffic Accidents after Correcting for Undersampling

In this study, four models—logistic regression (LR), random forest (RF), linear support vector machine (SVM), and radial basis function (RBF)-SVM—were compared for their accuracy in determining mortality caused by road traffic injuries. They were tested using five years of national-level data from the Korea Disease Control and Prevention Agency’s (KDCA) National Hospital Discharge In-Depth Survey (2013 through to 2017). Model performance was measured for accuracy, precision, recall, F1 score, and Brier score metrics using classification analysis that included characteristics of patients, accidents, injuries, and illnesses. Due to the number of variables and differing units, the rates of survival and mortality related to road traffic accidents were imbalanced, so the data was corrected and standardized before the classification models’ performances were compared. Using the importance analysis, the main diagnosis, the type of injury, the site of the injury, the type of injury, the operation status, the type of accident, the role at the time of the accident, and the sex were selected as the analysis factors. The biggest contributing factor was the role in the accident, which is the driver, and the major sites of the injuries were head injuries and deep injuries. Using selected factors, comparisons of the classification performance of each model indicated RBF-SVM and RF models were superior to the others. Of the SVM models, the RBF kernel model was superior to the linear kernel model; it can be inferred that the performance of the high-dimensional transformed RBF model is superior when the dimension is complex because of the use of multiple variables. The findings suggest there are limitations to analyses involving imbalanced, multidimensional original data, such as data on road traffic mortality. Thus, analyses must be performed after imbalances are corrected.