Tabular Fields and Narrative Text for Highway-Rail Grade Crossing Severity Assessment

Document Type

Conference Proceeding

Publication Date

1-1-2026

Abstract

Highway-rail grade crossings remain a critical safety concern, and reliable severity prediction can support timely reporting, investigation, and targeted safety action. This study develops integrated text-tabular classification models that combine structured fields with narrative text from incident reports. Using 21,000 FRA records from 2015 to 2024, we compare alternative text encoders and classifiers under five-fold cross-validation and a temporally realistic evaluation that trains on 2015-2021 incidents and tests on 2022-2024 data. On the temporal test set (6,459 records), the best hybrid model achieves 94.0% accuracy and a macro-F1 of 0.915, outperforming a structured-only baseline while maintaining 93.5% recall for fatal events. A modest threshold adjustment further increases fatal recall with a limited trade-off. Key predictors include train speed, gate violations, stalled-on-crossing conditions, and pedestrian or trespass contexts, with narrative cues improving discrimination between injury and no-harm outcomes. The framework enables incident-level severity assessment to support timely safety decisions.

Publication Title

International Conference on Transportation and Development 2026 Transportation Planning and Operations Selected Papers from the International Conference on Transportation and Development 2026

ISBN

[9780784487020]

Link to Full Text

Share

COinS