Predicting child occupant crash injury severity in the United Arab Emirates using machine learning models for imbalanced dataset

Road traffic crashes have increased over the years leading to greater injury severity among children who are mostly vehicle occupants in high-income countries. This adversely affects the healthy development of children and might lead to death. However, studies in the literature have focused on predicting crash injuries among adults while children have different crash injury risks as well as crash kinematics compared to adults. To address this gap, this paper presents a new dataset for child occupant crash injury severity prediction collected over 8 years (2012 to 2019) in the United Arab Emirates (UAE). The performance of state-of-the-art machine learning algorithms was then evaluated using the proposed dataset. In addition, feature selection techniques and logistic regression model were employed to extract the most significant features for crash injury severity prediction among child occupants. Furthermore, the impact of data balancing approaches on the prediction performance was analyzed as the dataset is highly imbalanced. The experimental results showed that Adaboost, Bagging REP, ZeroR, OneR, and Decision Table algorithms predicts child occupant injury severity with the highest accuracy. Child occupant seating position, emirate, crash location, crash type and crash cause were observed as significant features that predicts injury severity by both the feature selection and logistic regression models.