Enhancing bicyclist survival time in fatal crashes: Investigating the impact of faster crash notification time through explainable machine learning

Abstract

Bicyclists are recognized as vulnerable road users, with the escalating fatalities posing a safety concern. While fatal crashes involving bicyclists are often assumed to be similar, there is a crucial distinction between instant death and death occurring several days later, with the former being substantially more severe. This study delves into the analysis of bicyclists’ time-to-death, spanning from immediate fatalities to deaths within 30 days, using data from the Fatality Analysis Reporting System from 2015 to 2019. Employing the Haddon Matrix approach, the variables are categorized into pre-crash, during-crash, and postcrash phases. This study considers crash notification time as the key postcrash measure. An explainable XGBoost model is developed using the SHAP technique to investigate the associations between variables and bicyclist time-to-death. The results show that substantial delays in crash notification time considerably reduce bicyclists’ time-to-death and increase the likelihood of early death. Specifically, hit-and-run crashes, crashes in rural areas, and crashes during late-night hours exhibit notably longer crash notification times compared to non-hit-and-run crashes, urban areas, and other hours, respectively. In such cases, when no witnesses or survivors can notify emergency responders, on-road vehicle technologies like the advanced automatic collision notification system can promptly inform responders, reducing notification delays.