The rapid development of artificial intelligence has achieved significant success in traffic accident detection. However, the complexity of deep neural network algorithms and the large number of parameters make these models opaque and difficult to understand, resulting in uncontrollable potential risks and limiting large-scale applications in critical areas such as traffic safety. By using SHapley Additive exPlanations and Local Interpretable Model-agnostic Explanations methods to explain the decision-making process of traffic accident prediction models and presenting the results visually, we found that both methods yielded similar outcomes across different traffic scenarios. The model identified vehicle collisions as the primary feature of traffic accident scenes and regular spacing between vehicles as the main characteristic of normal traffic flow scenes. We also discovered that the model considers multiple features in its predictions and uses secondary features as supplementary information. This approach can significantly help researchers effectively understand the decisions made by the model, identify decision biases, and make targeted corrections to enhance model performance.