A Hybrid-Stratified Approach for the Identification of Pedestrian Crash Scenarios: The Effect of Demographic Vulnerability and Spatial-Temporal Shifts in the Pre- and Post-COVID-19 Period in Italy (2010–2023)

Pedestrian safety represents a critical priority for the development of sustainable urban mobility systems. This study proposes an innovative methodological framework integrating supervised and unsupervised learning techniques with econometric modeling to identify and interpret risk scenarios. Using the Italian national dataset from 2010 to 2023, an XGBoost model has been initially trained and tested. Then, SHapley Additive exPlanations (SHAPs) have been applied to highlight contributing factors. Using the resulting SHAP values, a K-Means clustering algorithm was finally employed to segment crashes into homogeneous clusters. For each cluster, a Generalized Linear Mixed Model incorporating geographic random intercepts and temporal random slopes was calibrated. Through this hybrid-stratified approach, three risk scenarios have been identified, primarily driven by demographic vulnerability. For elderly pedestrians, the involvement of heavy vehicles nearly doubles the odds of a fatal outcome. Crash dynamics varied significantly: heavy vehicles and speeding nearly double the fatality risk for elderly pedestrians; nighttime represents a severe hazard for adults (OR = 3.87) and youths (OR = 7.99), with the latter also highly penalized by unsafe road behaviors (OR = 3.12). From a spatio-temporal perspective, random effects revealed that the Islands (Sicily and Sardinia) are the most critical macro-areas (+55.2% baseline risk for adults) and the North-West the safest. Furthermore, the COVID-19 pandemic mitigated fatal risk for young pedestrians nationwide, had a neutral impact on the elderly, and for adults was protective in Southern regions but corresponded to higher odds of mortality in the North, reflecting altered traffic dynamics.