Explainable Road Traffic Crash Severity Prediction Using Machine Learning: Cross-Country Validation and Data-Driven Road-Safety Interventions

Explainable Road Traffic Crash Severity Prediction Using Machine Learning: Cross-Country Validation and Data-Driven Road-Safety Interventions

Parisa GhahremanShahraki1

1) Computer Engineer

IOI Code : AAHE-ABKHCF
Publication : National Conference on Civil Engineering, Design and Construction Technologies, Energy, Materials and Advanced Machinery(icaeu.ir)
Abstract :
    Road traffic injuries remain a major global public-health and development challenge. This study developed an explainable machine-learning framework for three-class road traffic crash severity prediction and examined its transferability across countries. Official UK STATS19 data for 2025, French BAAC data for 2024, and Addis Ababa road traffic crash data from Ethiopia for 2016–2022 were harmonized at the crash level. The outcome consisted of slight, serious, and fatal crash classes. The main cross-country feature space included crash hour, day of week, weather conditions, lighting, road-surface condition, junction characteristics, number of involved vehicles, and aggregated indicators for the presence of motorcycles, heavy vehicles, public transport, and two-wheeled vehicles. Cost-sensitive CatBoost was selected as the final UK model. It achieved an accuracy of 0.6634 and a Macro F1 score of 0.4014 in internal UK validation. When externally evaluated on French data, accuracy was 0.5334 and Macro F1 was 0.3934, indicating relatively preserved overall class-balanced performance despite differences in class distributions and class-specific error patterns. In six directional transfer experiments across the three countries, Macro F1 ranged from 0.3133 to 0.3497. Greater feature-distribution shift was associated with lower Macro Recall, although transferability remained dependent on transfer direction and crash-severity class. SHAP analysis identified junction characteristics, number of involved vehicles, and weather conditions as the most important predictors in the UK. In Ethiopia, driver-related features accounted for approximately one-fifth of global SHAP importance; however, they produced only a marginal improvement in repeated cross-validation performance (+0.0030 in Macro F1). These findings show that combining harmonized multicountry data, external validation, explainable machine learning, and intervention prioritization can generate more actionable evidence for road safety. Nevertheless, feature importance should be interpreted as predictive rather than causal evidence. Therefore, proposed safety interventions—particularly for factors with substantial cross-country distributional differences—require local validation.

Keywords : road traffic crash severity machine learning explainable artificial intelligence shap cross country validation catboost road safety intervention prioritization