Traffic Crash Severity: Comparing the Predictive Performance of Popular Statistical and Machine Learning Models Using the Glasgow Coma Scale

Crash severity analysis and prediction is a promising field in traffic safety. Various statistical methods have been used to model the severity of road crashes. However, machine learning algorithms have gained popularity in recent years. This study compares the predictive performance of various machine learning and statistical models, including prediction accuracy, and determines the influence of various variables on crash severity. The crash severity data were collected from a Hospital in Kashmir (India), an area with mixed topography. The crash severity levels (CSLs) were represented in the Glasgow Coma Scale (GCS). For estimations, the two statistical models, logistic regression (LR) and decision tree (DT), and four machine learning models, including random forest (RF), support vector machine (SVM), gradient boosted tree (GBT), and extreme gradient boosting (XG BOOST), have been used. The results show that the machine learning models have higher prediction accuracy than the statistical models. Among all, the GBT model has the best overall prediction accuracy, particularly in the prediction of individual CSLs while LR was found to have the least accuracy. The influence of variables on CSL was found from DT and GBT. Both models have indicated that ‘time’ as a variable was the most influencing, followed by the casualty class of pedestrians over the CSLs. The results also show that the variable influences over CSL were different from different models. Based on the influence of variables, certain policy implications are suggested, which might aid the transportation department, and other concerned departments to reduce the severity and number of road traffic crashes (RTCs). © 2023, The Institution of Engineers (India).