Traffic accidents pose significant risks to public safety and represent a critical issue in transportation systems. The accurate prediction of accident severity is essential for implementing effective prevention and intervention measures. An ensemble learning approach, combining the advanced algorithms XGBoost and MLP, was proposed to enhance the accuracy of traffic accident severity predictions. A stacked classifier was established and its performance in traffic accident prediction was thoroughly evaluated. The experimental results demonstrate that the integrated model significantly improves prediction accuracy compared to the traditional XGBoost model, with a notable 20.41% increase in the macro-average F₁ score. The advantages and innovations of the model, including model integration and network transformation, were highlighted. Additionally, the key features affecting the prediction results were analyzed, and the model's potential value in practical applications was explored. This study provides more scientific and efficient decision support for traffic safety management and is expected to play a crucial role in fields such as traffic management and intelligent driving.