A Review on Hybrid and Explainable AI Models for Road Traffic Accident Prediction

Road traffic accidents continue to be a critical global issue, causing substantial human suffering and economic costs. In recent years, predictive modeling has emerged as a promising direction for minimizing these risks by anticipating accidents and identifying key contributing factors. This review summarizes four notable studies in this area: an interpretable spatio–temporal multi-graph learning model (ASTMGCN), the use of CatBoost and BERT for accident type classification, a Random Forest–CNN ensemble (RFCNN) for severity prediction, and a CNN–BiLSTM–Attention model enhanced with DeepSHAP for risk assessment. These approaches are compared in terms of methodology, datasets, feature design, and interpretability. The findings indicate that hybrid and attention-based models generally outperform traditional methods, while explainability tools such as attention visualizations and SHAP values significantly improve trust and applicability in Intelligent Transportation Systems (ITS).