Urban Road Traffic Accident Severity Prediction Using a Hybrid LSTM—Random Forest Model with Hybrid Attention (LRHHA)

Predicting the severity of urban road traffic accidents is a key task in intelligent transportation systems (ITS). Accurately forecasting accident severity helps authorities implement effective interventions to reduce losses. In this paper, we propose a novel hybrid model called LRHHA (LSTM–Random Forest Hybrid with Hybrid Attention) for predicting the severity of urban road traffic accidents. Our model first uses a bidirectional LSTM autoencoder to extract deep nonlinear features from time-series accident data. We then apply a hybrid attention mechanism, combining adaptive attention and multi-head attention, to focus the model on the most relevant time steps and feature dimensions. These temporal features are fused with aggregated static features and reduced by PCA. Finally, a Random Forest classifier is trained on the enhanced feature set to perform severity classification. We address class imbalance using SMOTE oversampling and balanced random forest techniques. Experiments on real-world urban accident data demonstrate that LRHHA significantly outperforms traditional baselines, achieving improvements in macro-averaged F1 score and Cohen's Kappa. The results indicate the effectiveness of deep temporal feature extraction and attention mechanisms in this domain. Future work includes exploring spatio-temporal graph models and richer context features for further improvement.