Road traffic accident severity prediction based on large-scale data and multi-dimensional factors: an explainable...

Area of Science:

• Road safety
• Traffic engineering
• Data science

Background:

• Road traffic accidents cause significant fatalities and economic losses globally.
• Large-scale accident severity analysis is hindered by a lack of high-resolution roadway geometric data.

Purpose of the Study:

• To construct a large-scale dataset with fine-grained road alignment characteristics for accident severity analysis.
• To identify key factors influencing accident severity using advanced machine learning techniques.
• To provide mechanism-informed insights for developing targeted road safety strategies.

Main Methods:

• Automatic extraction of fine-grained horizontal and vertical road geometry from extensive networks.
• Integration of 26 features spanning environmental, roadway, and geometric alignment dimensions.
• A soft-voting ensemble model (XGBoost, Random Forest, CatBoost, LightGBM) for severity prediction, with SHAP for explanation.
• Accumulated Local Effects (ALE) analysis to reveal nonlinear patterns and threshold effects.

Main Results:

• Environmental conditions are more associated with lower-severity accidents.
• Roadway type and geometric alignment features become more critical for higher-severity accidents.
• ALE analyses identified nonlinear relationships and threshold regions for key variables, explaining risk variations.

Conclusions:

• Detailed road geometry data significantly enhances the understanding of accident severity patterns.
• Mechanism-informed insights from this study can guide the development of more effective road safety interventions.
• The findings underscore the importance of considering both environmental and geometric factors in road safety engineering.