Highway traffic incidents cause non-recurrent congestion and raise the likelihood of secondary rear-end collisions. Estimating when a blocked corridor will clear is difficult in day-to-day operations because crash severities vary widely, sensor feeds frequently contain missing values, and single-number point predictions offer no measure of forecast uncertainty. In this paper, we present a probabilistic duration forecasting framework built around a tri-modal feature extraction pipeline. The pipeline combines structured road and weather telemetry, dense text representations from a sentence transformer, and schema-constrained operational flags extracted from dispatch narratives using a local Large Language Model (LLM). These three feature sets train separate base models inside a out-of-fold Stacking Mixture-of-Experts (MoE) architecture, where gradient-boosted Accelerated Failure Time (AFT) base learners feed a stacked meta-learner paired with quantile regression heads. Rather than concatenating all features into one wide table, a quality-gating module scores tabular completeness, narrative character length, and extracted tag density to weight the expert predictions for each incident. This structure avoids tree dilution, a failure mode in which hundreds of sparse text features obscure core spatial and temporal predictors during tree split evaluation. Tested on a chronologically partitioned subset of the US-Accidents dataset containing 85,000 records, the model replaces fixed point estimates with continuous survival probability curves and calibrated quantile intervals. Empirical evaluation shows that the gated fusion stack lowers the Median Absolute Error (MedAE) to 31.10 minutes, improving upon both unimodal baselines and un-gated feature concatenation while providing reliable clearance bounds for traffic management centers.
 

Traffic Incident Duration,Large Language Models,Survival Analysis,Multimodal Fusion,Mixture of Experts,Probabilistic Forecasting