Text-enhanced incident duration prediction: An explainable end-to-end framework

Accurate prediction of traffic incident durations is essential for enabling traffic operators to implement timely response strategies and provide reliable travel advisories. Although textual data hold significant predictive potential, their utilization has been limited in prior studies because of interpretation challenges. This article introduces an explainable end-to-end framework that fuses incident messages with structured data to enhance incident duration prediction. We then develop a sequential prediction algorithm suitable for real-world deployment and use SHapley Additive exPlanation (SHAP) to assess the contribution of individual variables and textual elements. The empirical benefits are demonstrated by applying our text analysis method and proposed framework to incident records from Shaanxi, China. The results demonstrate that our method reduces prediction errors by approximately 20%. The sequential prediction algorithm provides reliable predictions after 30 min and achieves a mean absolute error of 9.63 min after 180 min. Key factors influencing incident duration include vehicle type, rollover, infrastructure damage, night hours, and weather conditions (SHAP value ranging from −20 min to +20 min). Additionally, characters representing quantities, degrees, and incident postures significantly impact incident duration. The proposed framework offers a powerful tool for incident management and can be extended to other predictive tasks involving textual data.