Comparing Performance of Different Machine Learning Methods for Predicting Severity of Construction Work Zone Crashes

In 2020, more than 102,000 work zone crashes occurred in the United States, resulting in over 45,000 injuries and more than 850 fatalities. These numbers are higher than 2019 records, despite lower traffic volumes due to the COVID-19 pandemic. Also, population growth and the increased load on infrastructure are expected to lead to more construction work zones and, consequently, more crashes and fatalities in the future. As such, understanding the factors contributing to work zone crashes and their severity can assist the Department of Transportation (DOT) in safety management and planning. Accordingly, this study investigated the effect of freeway construction work zones on the severity of crashes by employing machine learning algorithms. Moreover, the performance of three of the most common machine learning models, decision tree, random forest (RF), and XGBoost, were evaluated based on accuracy, precision, recall, and F1-score. The model was developed using crash data and work zone information from the state of Utah between 2017 and 2021, considering an extensive set of work zone attributes, such as road factors, environmental conditions, driver attributes, and work zone features. The study results showed that RF has the best performance in severity classification, achieving an accuracy of 88.6%. Moreover, feature importance analysis reveals that roadway surface conditions, crash type, motorcycle involvement, weather conditions, and roadway junction type are significant contributors to work zone crash severity. The findings of this study provide valuable insights for the analysis of construction work zone crashes and DOT’s safety management plans.