Abstract:
Limiting number of fatalities and reducing injury severity of car crashes is a continuing global concern. This study
investigates crash risk factors including driver, vehicle, roadway, and crash characteristics role in determining injury
severity levels encountered by drivers in fatal car crashes. Three types of supervised machine learning techniques were
used; Classification and Regression Tree (CART), Artificial Neural Networks (ANN) and Multinomial Logistic
Regression. The CART model was used to elect the most influential factors in determining drivers? injury severity levels.
The ANN model was used to predict drivers? injury severity based on crash attributes. The logistic model was used to
identify the effect of different crash attributes in distinguishing drivers? injury severity levels and for comparison purposes.
Consequently, CART model resulted in six significant factors, these factors are: airbag deployment, seatbelt use, driver
age, vehicle rollover, collision type, and vehicle model year. It was found that airbag deployment has a strong correlation
with severe injuries and even fatalities. The use of seatbelts appears to reduce injuries and fatalities. Furthermore, elderly
drivers, front to front collisions, vehicle rollover and older vehicles seem to cause more mortalities and injuries. On the
other hand, according to the logistic model, all the crash attributes were found significant in distinguishing between drivers?
injury severity levels except for the roadway functional system. However, the ANN model outperformed the CART model
in terms of accuracy and stability. Further, both models seem to outperform the logistic regression model in terms of
prediction accuracy.