A Multidimensional Traffic Accident Causation Index for Severity Modeling Using Explainable Machine Learning

Road traffic accidents remain a major public health concern, and effective safety management requires interpretable tools that integrate multiple causal dimensions. This study proposes a Traffic Accident Causation Index (TACI) to provide a holistic representation of severity-related drivers by combining six theoretically grounded domains: Accident Infrastructure, Driver, Pedestrian, Road Condition, Emergency and Response, and Severity. Using a national police-reported dataset from Türkiye (N = 13,639), operational variables are mapped to normalized risk scores, aggregated into domain indices, and combined into a 0–100 composite TACI score. To assess the robustness and compatibility of the proposed index framework, we develop ensemble machine learning models (Random Forest, Gradient Boosting, LightGBM, XGBoost, and CatBoost) under two feature configurations: an Extended Feature Set (EFS) with the original variables and a Core Feature Set (CFS) consisting of the six domain indices. The results indicate that domain-level aggregation improves predictive stability, and the best-performing boosting models (XGBoost/CatBoost) achieve near-perfect agreement with the constructed index (test R2 > 0.99) and very high classification performance (AUC > 0.999). SHAP-based explainability highlights pedestrian exposure and vulnerability as the dominant contributors, followed by lighting/visibility conditions, road surface quality, and adverse road–environment factors, whereas emergency-response and infrastructural attributes show comparatively indirect effects. Overall, the proposed framework supports interpretable, domain-oriented evidence for prioritizing safety interventions and monitoring high-risk accident conditions.