Interpretable machine learning for evaluating risk factors of freeway crash severity

Area of Science:

• Traffic Safety Engineering
• Data Science
• Transportation Research

Background:

• Machine learning (ML) models are increasingly used for crash severity prediction.
• Interpretability of these ML models is often overlooked, limiting practical insights.
• Understanding factors contributing to crash severity is vital for effective safety interventions.

Purpose of the Study:

• To implement interpretable machine learning techniques for visualizing factors affecting crash severity.
• To analyze five years of freeway crash data from Iran to identify key risk factors.
• To enhance decision-making processes in traffic safety through clear model interpretation.

Main Methods:

• Applied various machine learning models: Classification and Regression Trees (CART), K-nearest neighbours (KNNs), Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Machines (SVM).
• Utilized Accumulated Local Effects (ALE) plots for model interpretation.
• Evaluated model performance using accuracy, recall, F1-score, and ROC metrics, with RF showing superior results.

Main Results:

• Random Forest model demonstrated the highest performance in predicting crash severity.
• Key factors associated with severe crashes include light traffic conditions (critical values around 0.05 or 0.38).
• A higher proportion of large trucks and buses (e.g., 10% and 4%), speeds over 90 km/h, drivers under 30, rollover crashes, collisions with fixed objects/barriers, nighttime driving, and driver fatigue significantly increase severe crash likelihood.

Conclusions:

• Interpretable ML, specifically RF with ALE, effectively visualizes and quantifies the impact of various factors on crash severity.
• Findings provide actionable insights for targeted safety measures on freeways.
• The study highlights the importance of driver demographics, vehicle types, environmental conditions, and crash dynamics in determining severity.