Forecasting the Severity of Road Accidents in Russian Regions Using Interpretable Machine Learning Methods

Predicting the severity of road traffic accidents is an important applied task for reducing mortality and injury rates. Given the high dimensionality, heterogeneity, and class imbalance of data, traditional statistical models do not always provide adequate predictive performance and are limited in capturing complex non-linear relationships. This study aims to develop and comparatively analyze machine learning models for predicting crash severity using micro-level data from the Northwestern Federal District of Russia, as well as to interpret the contribution of key factors to the final prediction. The study tests two hypotheses: first, behavioral and spatial-infrastructure characteristics contribute more to predicting severe outcomes than meteorological and lighting conditions; second, ensemble machine learning methods provide higher predictive performance compared to linear and basic tree-based algorithms. The empirical base comprises data on 167,000 road traffic accidents that occurred between 2015 and 2024. The research procedure involved aggregating hierarchical records into an analytical table, generating the feature space, balancing classes, building and comparing models, and interpreting the results using SHAP analysis. The evaluated models included logistic regression, decision trees, gradient boosting ensembles, and fully connected neural networks with categorical embeddings. The experimental results showed that an ensemble of gradient boosting and a neural network demonstrated the best predictive performance (ROC AUC = 0.746). It was found that geographic location, vehicle characteristics, and the behavioral patterns of road users make the greatest contribution to predicting severe outcomes, whereas the contribution of weather conditions is secondary. The obtained results support both research hypotheses. The theoretical significance of this study lies in advancing the predictive approach to traffic accident analysis using explainable machine learning. The practical significance is driven by the potential to use the findings for developing intelligent transportation systems (ITS) and prioritizing road safety improvement measures.