Leakage-aware explainable machine learning for highway–rail grade crossing safety screening

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Highway–rail grade crossing safety programs should allocate limited inspection and capital improvement resources across more than 200,000 public at-grade crossings in the United States. This paper presents a leakage-aware, explainable machine learning (ML) screening model that identifies candidate high-risk crossings. The model utilizes Grade Crossing Information System (GCIS) inventory records (Form 71) for approximately 20,000 Union Pacific (UP) crossings, merged with accident reports (Form 57) from 2015–2024. A critical methodological challenge is that accident records carry incident timestamps while non-accident crossings do not; naively augmenting both with weather covariates causes the absence of weather data to become predictive of the negative class—a missingness-driven leakage pathway we explicitly address by randomly imputing timestamps for non-accident rows prior to Meteostat weather lookup. An imbalance-aware eXtreme Gradient Boosting (XGBoost) classifier—assessed via the Area Under the Receiver Operating Characteristic curve ( AUROC = 0.887 )—is evaluated at a primary baseline threshold ( t = 0.50 , Recall = 0.860 ). Sensitivity is further analyzed against a validation-optimized threshold ( t = 0.71 , F1 = 0.613 ). Model transparency is provided through SHapley Additive exPlanations (SHAP) attribution and a low-depth surrogate decision tree. Robustness is assessed via a five-variant feature ablation ( AUROC = 0.887 full vs. 0.806 infrastructure-only), a leave-one-cluster-out geographic holdout across four UP network regions (mean AUROC = 0.840 ± 0.023 ), a threshold sensitivity analysis linking operating point to inspection budget, and a cross-railroad generalizability experiment on BNSF Railway ( AUROC = 0.930 , F1 = 0.711 at t = 0.50 ) using the identical pipeline without modification. Results confirm that weather and temporal features contribute meaningfully beyond infrastructure alone, that the model generalizes moderately well across geographically held-out regions, and that the leakage-aware pipeline transfers across Class I freight networks, supporting its use as a data-driven triage tool for crossing safety programs.