Predicting the level of road crash severity: A comparative analysis of logit model and machine learning models

The global expansion of cities and roads has made road safety one of the most pressing issues in sustainable mobility, making road safety assessment a multifaceted challenge that requires an analysis of both anticipated and unforeseen factors. One of the most critical aspects of road safety investigation and planning is predicting the severity of road crashes. Accurate determination of crash severity holds significant importance in strategic planning and cost evaluation within the context of road safety assessment. Two main goals are emphasized in this study: building and developing hybrid machine learning models and comparing them with the binary logit model, as well as determining the importance of different parameters on the level of road crash severity (LRCS). For this purpose, a binary logistic regression (BLR) model and support vector machine in combination with Whale Optimization Algorithm (WOA), grasshopper optimization algorithm (GOA), and Multi-Verse Optimizer algorithm (MVO) were employed and developed. In this study, a substantial of real-existing data (1600 cases) from 2015 to 2018 on the road network of the Cosenza province in southern Italy was used. The results of this study clearly showed that the results obtained from the hybrid machine learning models were not significantly different compared to the BLR model. Also, all the hybrid models had good and robust performance, and no significant difference was observed in the prediction of LRCS, although the SVM-WOA algorithm was able to reach convergence earlier than the other machine learning models. Finally, according to the results of the sensitivity analysis, the number of vehicles had the greatest impact on LRCS, while the road element had the least impact.