Calibrated and Explainable Gradient Boosting for Road Traffic Crash Severity Prediction: SHAP Audit and Cross-Jurisdiction Transfer Evaluation

Crash severity prediction is critical for emergency response, infrastructure spending, and risk communication. Although machine learning has been widely applied to this problem, three gaps prevent practical deployment: uncalibrated probability scores, SHAP-based explanations whose faithfulness has not been verified, and models never tested outside their training jurisdiction. The proposed framework, SAE-XCrash (Safety-Aware and Explainable Crash Severity Prediction), addresses all three using two public datasets—US-Accidents (7.0 million records, 2016–2023) and UK STATS19 (approximately 1,010,000 records, 2016–2022)—with strict temporal splits throughout. Notably, the US-Accidents severity label measures traffic disruption duration, not injury outcome; results should be interpreted accordingly. Previously unknown label-schema drift led to a revised binary target with Severity 4 as the only positive class. Five classifiers are compared. Post hoc isotonic calibration reduces Expected Calibration Error by 97.3% at negligible discrimination cost. A four-step quantitative SHAP audit confirms statistically significant deletion faithfulness; however, explanation stability fails at realistic perturbation levels (54.3% low-stability fraction at sigma = 0.05), driven by spatial data sparsity in sparse geohash cells—a negative result that carries direct operational implications for deployment. A three-tier cross-dataset transfer experiment (zero-shot, recalibration, full retrain) shows that temporal features transfer robustly across jurisdictions, while spatial memorization is the primary generalization barrier. All code, split indices, and model artifacts are publicly available.