The effect of pre-crash variables on occupants’ injury risk in car impacts: assessment by feature ranking, regression, and machine learning for increasing injury scales

Abstract

The Vision Zero goals of the European Commission aim to eliminate road fatalities and severe injuries by 2050. To achieve this target, it is crucial to identify the most influential variables on injury outcomes in real accidents, determining priority axes for interventions aimed at increasing road safety. The study investigates the effect of variables related to impact, vehicles, and occupants in car-to-car collision scenarios through feature ranking and modeling methods. The outcomes range from minor injuries (MAIS 1+) to extremely severe injuries (MAIS 5+). The ranking illustrates that impact-related variables, such as closing speed at collision, eccentricity, and the area affected by intrusion, consistently remain the top three influencing factors on injury outcomes. The occupant’s gender is the most influential variable only in the case of MAIS 1+. Based on feature ranking, it is highlighted that variables related to occupants and vehicles do not contribute to the quality of the derived injury risk models. Machine learning models (particularly Support Vector Machines) provide accuracy up to 90%, slightly superior to established logistic regression models characterized by high interpretability. Overall, the evidence suggests that targeted interventions to modify collision-related variables are essential for minimizing injuries of any severity. At the vehicle level, for example, this can be achieved through interventions on braking and steering by advanced driver assistance systems, aimed at minimizing injury risk even in cases where the impact is unavoidable.