Predicting road traffic accident severity in Colombo: A Random Forest approach

Abstract

Road traffic accidents continue to be a major global problem, since they cause deaths and nonfatal injuries, which lead to economic, healthcare and social burdens on a country. Urban centres like Colombo are crucial in the accident analysis in Sri Lanka due to the high traffic density and complex road networks. The main goal of the study is to create a classification model using Random Forest to predict the accident severity as either “Minor” or “Severe” using the variables such as time of day, road surface conditions, weather, lighting, number of vehicles involved, and location type. Furthermore, it focuses on identifying the most influential variable to the model through variable importance and provides data-driven policy recommendations to improve traffic safety. The dataset contains records of accidents that occurred in the Colombo Municipal Council area from 2019 to 2023. It was sourced from the City Traffic Police Station, Fort. The Random Forest model was constructed using the Random Forest package in R. It builds multiple decision trees using bootstrapped samples and selects random subsets of variables at each split, and then the final prediction is made through majority voting among the individual trees. The mean decrease in Gini impurity from the Random Forest model highlighted variables such as “Location Type”, “Day Night”, “Number of Vehicles” and “Weather Condition” as key predictors of accident severity. This suggests that the environment where the accident occurs significantly influences predicting the severity of the accident, and a greater number of vehicles may increase the likelihood of more severe outcomes, possibly due to increased collision complexity. Also, accident severity patterns vary across weekdays. The confusion matrix showed that out of the total cases, it accurately identifies 929 minor incidents and 160 severe ones. As a result, a sensitivity of 86.18% for minor cases but a worrying specificity of just 22.66% for severe ones was obtained. Essentially, while the model reliably spots minor incidents, it struggles to flag the more consequential severe events, which may be due to the imbalance between the classes. Therefore, addressing the class imbalance and exploring ensemble tuning or alternative algorithms can be applied to improve the performance of the model in future studies.