TRIDENT-Risk: An AI-driven temporal ensemble framework for proactive high-severity accident risk forecasting in...

High-severity traffic accidents impose substantial human and economic burdens, necessitating proactive and data-driven risk forecasting mechanisms within intelligent transportation systems. This study proposed TRIDENT-Risk, a temporal risk-driven imbalance-optimized ensemble framework for high-severity accident prediction using the eXtreme Gradient Boosting (XGBoost) model. The approach integrates spatial-temporal aggregation, multi-scale lag and rolling risk features, imbalance-aware gradient boosting, and threshold optimization to enhance minority-class detection. Experimental evaluation using publicly accessible US-Accidents dataset including about 7.7 million accident records from February 2016 to March 2023 demonstrated that the proposed TRIDENT-Risk achieved a ROC-AUC of 0.8917, PR-AUC of 0.7266, precision of 0.60, recall of 0.81, F1-score of 0.69 and accuracy of 0.82. Compared with the strongest baseline (MLP), the proposed framework improved high-risk recall from 0.70 to 0.81 (15.7% relative improvement), increased ROC-AUC from 0.8776 to 0.8917 and improved PR-AUC from 0.7016 to 0.7266, demonstrating superior minority-class detection and reliable proactive accident forecasting. Temporal grid-based aggregation was used to frame high-severity accidents as a binary prediction problem and an 80:20 chronological train-test split was used to eliminate temporal leaking. The proposed framework offers a scalable and interpretable solution for proactive traffic safety management and supports intelligent decision-making in next-generation transportation systems.