Assessing and mitigating traffic crash risks using interpretable machine learning techniques

Area of Science:

• Traffic Safety Research
• Behavioral Psychology
• Machine Learning Applications

Background:

• Road traffic crashes pose a significant global health challenge.
• Understanding driver behavior is crucial for developing effective safety interventions.
• Previous research has explored various factors contributing to road accidents.

Purpose of the Study:

• To identify factors associated with self-reported crash involvement among Pakistani drivers.
• To apply interpretable machine learning (ML) techniques for analyzing crash risk factors.
• To evaluate the predictive performance of different ML models in traffic safety.

Main Methods:

• Utilized a cross-sectional Internet-based survey of 623 drivers in Pakistan.
• Collected data on dangerous, aberrant, and positive driving behaviors, demographics, and self-reported crashes.
• Employed interpretable ML techniques: Logistic Regression, Categorical Boosting (CatBoost), and eXtreme Gradient Boosting (XGBoost).
• Applied Shapley Additive explanation (SHAP) for enhanced model interpretability.

Main Results:

• Aggressive driving, risky driving behaviors, errors, and violations were significantly associated with increased traffic crashes.
• Prior motorcycle riding experience showed a significant association with crash occurrence.
• XGBoost demonstrated superior predictive performance (0.86 accuracy) compared to CatBoost (0.80) and Logistic Regression (0.73).

Conclusions:

• Interpretable ML is valuable for traffic safety research, identifying key crash risk factors.
• Behavior-oriented interventions focusing on aggressive/risky driving and risk perception are recommended.
• Transition-oriented training for motorcycle riders is suggested to improve adaptation to different vehicles and road usage.