Pickup truck crash severity analysis via machine learning: policy insights for developing countries

Area of Science:

• Road Safety
• Machine Learning in Transportation
• Traffic Accident Analysis

Background:

• Methodological gaps exist in analyzing pickup truck crash severity.
• Understanding distinct risk factors for single-vehicle versus multi-vehicle crashes is crucial.
• Machine learning offers potential for improved crash prediction.

Purpose of the Study:

• Evaluate machine learning models for crash severity prediction.
• Compare outcomes of single-vehicle and multi-vehicle crashes.
• Identify key risk factors influencing crash severity.

Main Methods:

• Compared Logistic Regression, Random Forest, XGBoost, and Deep Neural Networks.
• Utilized K-fold cross-validation and Bayesian Optimization for model tuning.
• Employed SHAP for model interpretability and feature analysis.

Main Results:

• XGBoost demonstrated superior performance for multiclass injury classification in both crash types.
• Random Forest and Deep Neural Networks showed strength in binary classification for single- and multi-vehicle crashes, respectively.
• Identified common factors like 4-lane roads, unlit roads, and barriers influencing both crash types.

Conclusions:

• Model selection and injury classification schemes significantly impact predictive performance.
• Distinct factors influence severity in single-vehicle (e.g., fatigue, road type) versus multi-vehicle crashes (e.g., vehicle involvement, collision type).
• Specific factors reducing severity in single-vehicle crashes do not apply to multi-vehicle crashes.