A Machine Learning Approach to Predict Traffic Accident Severity in Addis Ababa

Abstract

According to the annual report by WHO, the increase in population, urban expansion, and the rapid growth of motorized transportation have made traffic accidents a critical global problem. Currently road fatalities are 1.35 million annually. In Ethiopia, traffic accidents death rate is 26.7 per 100,000 populations. In Addis Ababa 403 fatal crashes were registered in 2022/23. This study aimed to use machine learning techniques to predict the severity of traffic accidents in Addis Ababa by developing a predictive model. Ten machinelearning algorithms were selected and tested using historical traffic accident data from Addis Ababa Traffic Management Agency. The dataset contains 9532 instances with 19 variables, from these 80% are used for training and 20% for testing. Association rule mining reveals that less severe traffic accidents are frequently associated with dry road conditions, with a confidence value of 13.56% and a lift value of 1.066. Severe accidents are commonly linked to the absence of road junctions, with a confidence value of 78.97% and a lift value of 1.004. Fatal accidents are strongly associated with poor lighting conditions, with a confidence value of 20.41% and a lift value of 1.456. This analysis highlights the significant impact of these factors on accident severity, providing valuable insights for developing targeted prevention and mitigation strategies. Model experimentation shows that the XGBoost model achieved the highest accuracy at 84%. Decision Trees had the lowest accuracy of 72% while Bagging with SVM improved to 76% accuracy. Random Forests achieved 79% accuracy, and Ada Boost and Na??ve Bayes also reached 79%. SVM, KNN, Logistic Regression, and RNN all achieved an accuracy of 80%. Despite the strong performance of several models, XGBoost remains the best performing model with the highest accuracy, underscoring its capability to make accurate predictions based on this study.