Concentration of injury and fatality harm at repeat-active highway–rail grade crossings: A multi-state permutation and machine learning analysis

Highway–rail grade crossing incidents remain a transportation safety concern, yet prior studies have focused mainly on crash occurrence, severity, and hotspot detection, leaving limited evidence on how injury-or-fatality harm concentrates at crossings with repeated incidents. Using Federal Railroad Administration (FRA) Form 57 records from six U.S. states during 2013–2025, this study quantified repeat-site harm concentration and evaluated leakage-controlled surveillance ranking. The final cohort comprised 8,286 incidents, of which 3,250 involved at least one injury or fatality. Repeat-active crossings, defined as crossings with at least two incidents, accounted for 56.3% of documented harm and exceeded the randomized benchmark (p = 0.013; false-discovery-rate-adjusted p = 0.026). The harm-count Gini coefficient across all observed incident-involved crossings was 0.630, indicating an unequal distribution of documented harm. Leakage-controlled rolling-origin histogram gradient boosting validation yielded a mean area under the precision–recall curve (AUPRC) of 0.721, while unseen-crossing evaluation yielded 0.693; separate leave-one-state-out logistic validation yielded 0.674. Conceptually, the study formalizes repeat-site harm concentration as distinct from incident frequency and separates surveillance prioritization from causal safety mechanisms. The framework supports repeat-site screening and prioritization for further inspection and engineering assessment without implying causal infrastructure deficiencies or intervention effectiveness.