Answers are grounded in this paper and cite their sources.
Road traffic accidents represent a global public safety crisis that calls for accurate and interpretable injury severity analysis. This study evaluates high-performing ensemble machine learning models, including AdaBoost, XGBoost, LightGBM, HistGBRT, CatBoost, Gradient Boosting, NGBoost, and Random Forest, for distinguishing fatal from non-fatal injuries among persons involved in fatal crashes, using National Highway Traffic Safety Administration (NHTSA) Fatality Analysis Reporting System (FARS) records from 2018 to 2022, with models trained on 2018–2020 records and tested on 2021–2022 records. All models reached 83.4–83.9% accuracy and ROC AUC 0.901–0.916, with overlapping 95% confidence intervals, so no single algorithm was clearly superior. SHapley Additive exPlanations (SHAP) provided global and local explanations. The global analysis identified restraint use, harmful event type, vehicle body type, and age as the strongest correlates of fatality, while local waterfall plots for correctly and incorrectly classified crashes revealed how variables combine in individual incidents and where the models fail. Refitting four representative models without post-crash testing variables lowered ROC AUC by 0.015–0.023 and left the leading features unchanged. Framed as a retrospective analysis of risk factors rather than a real-time prediction tool, the results describe who dies within fatal crashes, not injury severity across all crashes.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: