Accurate assessment of traffic accident severity is crucial for enhancing traffic management efficiency and emergency response capabilities. However, traditional methods are limited in modeling multidimensional feature coupling, specifically by neglecting spatial dependencies from road network topology and failing to explore implicit correlations between heterogeneous entities. To address these challenges, we propose a hybrid model, Multi-Stage Knowledge Graph Embedding Network (MSKGEN), which integrates RotatE and Relational Graph Convolutional Networks (R-GCN) to enhance the accuracy of traffic accident severity classification. First, a hierarchical traffic accident severity assessment dataset is established by integrating temporal, spatial, and multidimensional influencing factors. Entity and relational features are extracted from this dataset, and a traffic accident knowledge graph is constructed via Neo4j to formalize domain-specific relationships. On this basis, a cascaded risk prediction framework was developed. Firstly, RotatE was employed to generate high fidelity embeddings of entities and relationships, accurately modeling complex semantic relationships. Subsequently, R-GCN is adopted to encode structural dependencies among entities, effectively capturing the spatial heterogeneity of accident distributions and amplifying the model’s capacity to identify latent risk factors. Finally, multi-class severity prediction is achieved through Multi-Layer Perceptron (MLP). The experimental results show that MSKGEN outperforms traditional models in terms of evaluation metrics, verifying its superiority in practical applications. This study advances a novel framework for traffic accident severity analysis, demonstrating both significant theoretical value and promising practical applications for intelligent transportation systems.