Explainable Machine Learning for Road Crash Severity Modelling under Temporal Validation, Calibration and Transferability Constraints

Reducing fatal and serious consequences of road crashes requires models that are accurate, calibrated, interpretable and reproducible under realistic deployment conditions. This study develops an explainable machine-learning workflow for road crash severity modelling using linked ۲۰۲۳ STATS۱۹ collision, vehicle and casualty records. The unit of analysis is the crash, and the main endpoint is a policy-relevant fatal or serious injury threshold rather than a full ordinal model. A temporal validation design trains models on January-October crashes and evaluates them on a November-December winter hold-out set. Class imbalance is handled through inverse-frequency weighting and assessed with precision-recall area, Matthews correlation coefficient, Brier score, expected calibration error and bootstrap uncertainty intervals. The analysis separates retrospective severity screening, which may use casualty-level descriptors recorded after a crash, from prospective infrastructure risk estimation, which cannot use such variables and requires exposure data. The selected CatBoost workflow achieved the highest precision-recall area among the main implemented models, but the study does not claim state-of-the-art superiority because TabPFN, Mambular, MambaAttention and Feature Group Tabular Transformer were not empirically implemented. The raw expected calibration error was high and may partly reflect seasonal distribution shift in the winter hold-out period; therefore, raw probabilities are not ready for safety-critical thresholds without cautious post-hoc recalibration, class-specific reliability assessment and local validation. Shapley additive explanations identify vulnerable road-user involvement as the most influential predictor, but these values are interpreted as predictive attributions rather than causal evidence. Transfer to Iran is framed as pipeline adaptation requiring registry linkage, exposure data, Highway Safety Manual calibration, domain adaptation, meta-learning and empirical validation.

کلیدواژه ها:

road safety ، crash severity ، explainable machine learning ، temporal validation ، probability calibration ، STATS۱۹ ، vulnerable road users