Road traffic fatalities remain one of the most persistent public safety challenges, demanding predictive models that are not only accurate but also reliable and interpretable. While numerous machine-learning models have been developed for crash severity prediction, most remain limited in three critical dimensions: they simplify the unit of analysis by focusing only on the driver or the most severe occupant per crash, they emphasize predictive performance while neglecting model confidence and probability calibration, and they often function as black boxes lacking interpretability for real-world decision support. To address these gaps, this study proposes a calibrated hybrid machine-learning framework for nationwide, person-level prediction of fatal versus non-fatal outcomes among individuals involved in fatal crash events using over 745,000 records. The framework integrates gradient boosting, ensemble stacking, and post-hoc probability calibration to produce both high discrimination and trustworthy risk estimates. Results demonstrate that the Hybrid Calibrated Ensemble achieves strong and reliable performance, outperforming single classifiers in probability alignment while maintaining comparable discrimination. Subgroup analyses across gender, person type, seat position, and rural–urban context confirm the model's stability and fairness. SHAP-based explainable AI further reveals that ejection status, restraint use, harmful event type, age, and airbag deployment are dominant determinants of fatality risk, with clear nonlinear effects aligned with crash-mechanics.
Khanjar et al. (Wed,) studied this question.