Spatiotemporal Risk Modeling for Traffic Accident Prediction in Urban Road Networks

Abstract

Traffic accident prediction plays an important role in intelligent transportation management and urban road safety planning. Existing predictive models frequently rely on historical accident statistics without adequately capturing dynamic traffic conditions and spatial interactions among road segments. This study proposes a spatiotemporal risk modeling framework that integrates graph neural networks with temporal attention mechanisms to estimate accident risks across urban transportation networks. The proposed approach incorporates traffic volume, vehicle speed, weather conditions, road geometry, and historical accident records into a unified predictive architecture. The illustrative experimental setting includes approximately 45,000 road segments and 12,300 recorded accidents across five metropolitan areas over a three-year observation period. The proposed framework achieves a precision-recall AUC of 0.421, compared with 0.336 for a conventional gradient-boosting baseline. When prioritizing the 10% of road segments with the highest predicted risk, the model identifies 71% of recorded accident events, compared with 59% for the baseline. Cross-city evaluations further indicate improved generalization under heterogeneous traffic conditions. These findings demonstrate the potential of graph-based predictive modeling for data-driven traffic safety management.