Injury severity prediction of traffic crashes with ensemble machine learning techniques: a comparative study

Area of Science:

• Traffic Safety
• Machine Learning
• Data Science

Background:

• Understanding traffic crash injury severity risk factors is crucial for effective mitigation strategies.
• Traditional statistical models may produce biased predictions due to inherent assumptions.
• Advanced machine learning offers potential for more accurate crash analysis.

Purpose of the Study:

• To compare the performance of eXtreme Gradient Boosting (XGBoost) against traditional machine learning models for crash injury severity analysis.
• To identify key risk factors influencing crash injury severity using XGBoost.

Main Methods:

• Utilized a dataset of 13,546 motor vehicle collisions from Riyadh, KSA (2017-2019).
• Compared XGBoost with logistic regression, random forest, and decision tree algorithms.
• Employed k-fold cross-validation (k=10) for performance evaluation.

Main Results:

• XGBoost demonstrated superior predictive performance and individual class accuracies compared to other models.
• Feature importance analysis highlighted collision type, weather, road surface, damage type, lighting, and vehicle type as critical predictors.
• The XGBoost model outperformed previous studies in crash injury severity prediction.

Conclusions:

• XGBoost is a highly effective tool for analyzing traffic crash injury severity.
• Identifying sensitive risk factors can inform targeted safety interventions and policy development.
• This study provides a robust, data-driven approach to enhance road safety analysis.