A Shallow Neural Network approach for identifying the leading causes associated to pedestrian deaths in Medellín

Abstract

IntroductionTraffic collisions are one of the main reasons of violent deaths in the world, and the pedestrians represent one of the principal casualties, being more critical in low- and mid-income countries. Moreover, the causes of these traffic collisions may vary depending on the studied region. Thus, to develop focused strategies to reduce these deaths, it is important to find the leading causes of these events in a specific region. MethodologyPublic collected data related to traffic collisions from 2009 to 2019 was acquired and used to train a Shallow Neural Network for modelling and to find the leading causes associated with the pedestrian deaths in Medellín, Colombia. The dataset used for training the model was highly imbalanced, which represents a big challenge to classify the traffic collision as injury or death. ResultsAfter improving and assessed the model using different performance metrics, it was found that larger vehicles (such as buses and trucks) involved in collisions, and poor visibility (associated to “blind spot” areas due to some events such as improper parking or the sharing of road space between pedestrians and vehicles in busy areas) are the leading causes associated with the pedestrian deaths in Medellín. Also, it was observed that the vast majority of pedestrian deaths occurred in the downtown area and in principal avenues of the city. ConclusionsThe major contribution of this study is the possibility of using the model as a tool to develop focused strategies aimed to reduce pedestrian deaths in Medellín, considering the leading causes associated to these events in this city.