Data-Driven Infrastructure Safety Screening of Crash Casualty Severity Using Temporally Validated CatBoost and Adjusted Logistic Regression

This study examined infrastructure-related casualty involvement in police-reported crashes and assessed how record quality and crash context affect interpretation. The cohort comprised 6190 Eslamshahr crashes recorded from March 2022 to June 2024, including 1805 injury or fatal crashes. Models developed on 5299 records were assessed on 891 spring-1403 records. The reference CatBoost achieved ROC-AUC 0.857, average precision 0.725 and Brier score 0.120; development-only sigmoid recalibration reduced Brier to 0.112. A 39-configuration comparison did not improve CatBoost and yielded stronger average-precision and Brier results for random forest on the same existing test. Adjusted associations for inadequate nighttime lighting (OR 2.78, 95% CI 1.66–4.64) and visibility obstacles (OR 1.89, 1.05–3.38) remained elevated across width and influence checks. Exactly 12 m accounted for 64.3% of widths and materially affected geometric estimates. Removing crash type and at-fault vehicle information reduced CatBoost ROC-AUC to 0.532. The findings support retrospective examination of recorded lighting and visibility conditions and show why documentation, predictor availability and calibration matter when interpreting severity models.