Effects of 5D built environment and non-built-environment factors on injury crash risk: An interpretable machine learning analysis.

To identify the key determinants of traffic injury risk and clarify the relative roles of built environment factors and crash-context factors, this study develops a traffic injury risk identification framework that integrates 5D built environment variables with non-5D factors, using traffic crash data collected in Changsha, Hunan Province, China, from 2017 to 2019. After data cleaning, screening, and spatial matching, a total of 9,743 valid samples were obtained, and injury occurrence in a crash was defined as a binary dependent variable. On this basis, three feature sets were constructed, including a 5D feature set, a non-5D feature set, and a combined 5D + non-5D feature set. Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) models were then developed and compared, and the best-performing model was further interpreted using the Shapley Additive Explanations (SHAP) method. The results showed that the combined 5D + non-5D feature set consistently outperformed the models using either the 5D or non-5D feature set alone, indicating that traffic injury risk arises from the joint influence of the built environment and immediate crash-context conditions. Among the three models, CatBoost achieved the best performance under the combined feature set and produced the highest receiver operating characteristic-area under the curve (ROC-AUC) value, demonstrating superior overall discriminative ability. The SHAP results further revealed that, within the combined CatBoost model, 5D variables accounted for a larger share of the model-based contribution than non-5D variables. However, given the relatively small improvement in accuracy after adding 5D variables, this finding should be interpreted as evidence of complementary explanatory information rather than as a dominant source of predictive performance. Distance to the nearest metro station, lighting condition, road network density, point of interest (POI) mix, and distance to the nearest bus stop were identified as the most influential factors. Further dependence analysis showed that a greater distance to the nearest metro station generally increased traffic injury risk, whereas higher road network density and greater POI mix were generally associated with lower injury risk. From the perspective of the interaction between 5D built environment characteristics and non-5D factors, this study reveals the multidimensional pathways through which traffic injury risk is shaped. The findings also confirm the effectiveness of the CatBoost-SHAP framework for traffic injury risk identification and interpretation, and provide empirical support for urban traffic safety risk assessment, road environment optimization, and refined governance strategies.