Explainable machine learning for property-damage-only crash prediction: Assessing spatial and temporal generalizability

Abstract

Road traffic crashes are a persistent global public health concern. In Saudi Arabia, crash rates have remained high for decades. However, most road safety studies in the country have focused on fatal and injury crashes, with limited attention to property-damage-only (PDO) incidents. Rigorous analysis of PDO crashes is essential for early risk detection, insurance considerations, and preventing escalation to more severe outcomes. Many existing machine learning (ML) models for crash prediction overlook temporal dynamics, spatial heterogeneity, and shifts in predictor importance. This study presents an explainable machine learning (XML) framework for modeling PDO crashes on intercity and rural highways in Saudi Arabia using crash data from 2017 to 2022. Six ML algorithms were evaluated using nested cross-validation and hyperparameter tuning based on the Tree-structured Parzen Estimator (TPE), a Bayesian optimization algorithm that models the search space probabilistically. XGBoost achieved the best overall performance. Temporal generalizability was assessed using a sliding-window approach, showing generally stable performance with a decline around 2020. Spatial generalizability was examined across identified hotspot regions and showed generally consistent regional performance, with slightly lower performance in major metropolitan areas. SHapley Additive exPlanations (SHAP) were used to interpret the model. Key predictors of PDO crashes included accident type, accident cause, and the number of vehicles involved. Temporal SHAP analysis showed an increasing influence of heavy vehicles and pedestrian involvement after 2019.