Interpretable crash severity prediction models to improve cyclist safety

This study focuses on 3 years (2016–2018) of cyclist crashes in the City of Rome, Italy. As the first step, a statistical analysis was carried out. Several Cycling Crash Models were developed by using Logistic Regression Models, with a deep dive into the most influencing variables. The two proposed models at intersections and single-lane carriageways have a McFadden score or pseudo-R2 of 0.3976809 and 0.4495008, respectively. The findings show that visibility does not play a key role in leading to a crash with a cyclist; sunny weather is positively correlated to crashes in intersections, while dry surfaces increase the chances of having crashes on single-lane carriageways, such as also the location of these roads in extra-urban environments and autumn and winter seasons. Weekdays are also related to an increase in the probability of having a crash at intersections and on single-lane carriageways. Cyclist crashes are more likely to happen in the evening and nighttime hours. Vertical and horizontal signposting decreases the probability of crashes in intersections and single-lane carriageways. High values of average daily traffic (>2000 vehicles/day) are strongly related to crashes on single-lane carriageways, and high speeds (>50 km/h) increase the probability of fatal crashes in intersections and on single-lane carriageways.