An explainable two stage framework for spatial temporal traffic accident prediction and injury severity classification

Abstract

Traffic accidents pose severe risks to public safety, yet existing injury severity prediction models often overlook spatial-temporal dynamics. This study proposes a two-stage framework integrating spatial-temporal analysis with machine learning to enhance prediction accuracy and interpretability. In the first stage, a Graph Convolution Network-Long Short-Term Memory (GCN-LSTM) model predicts daily accident counts across administrative regions, incorporating the predictions normalized as relative frequencies as features for injury severity classification. The second stage, which classifies the injury severity of accidents that have occurred, employs a stacking ensemble model with four tree-based classifiers, selecting the best-performing configuration based on performance metrics. Experimental results on a real-world dataset demonstrate the framework’s superiority over traditional methods. The GCN-LSTM model reduces mean absolute error (MAE) and mean squared error (MSE) by 18.4% and 3.6%, respectively, compared to the state-of-the-art Spectral Temporal Graph Neural Network (StemGNN). The stacking model with Random Forest as the meta-learner achieves an 87% accuracy in severity prediction. SHAP-based interpretability analysis reveals key factors influencing injury severity, providing actionable insights for traffic safety management. This study offers an interpretable approach to injury severity prediction, supporting data-driven decision-making in accident prevention and emergency response.