Road Accidents Injury Severity Prediction via Machine Learning Algorithms

Road accidents are a significant public health issue all over the world and, along with being one of the leading causes of global death, are accountable for lots of serious injuries, which in some cases consequences could last for years; thus, it is essential for better decision making, safety management and reducing acuteness of this injuries to explore and understand the factors concerning injury severity. This study considers ۲۷ attributes for injury severity prediction, some of which have yet to be in other related studies. To determine the attribute's relation with injury severity, three different methods are used to rate the attributes in the final dataset to see which one of them would give a more reliable score to each attribute; these methods are the Chi-squared test, ANOVA F Test and Random Forest for feature importance. After sorting them from best to worst in three different approaches, the top ۱۰, ۱۵, and ۲۰ features from each method are picked to predict the injury severity of each person after a vehicle accident using two Machine Learning algorithms, Decision Tree and Random Forest. This paper aims to clarify the role of each attribute regarding injury severity and to see which of the selected algorithms and feature selection methods can achieve the best result in predicting injury severity. The result shows that the Random Forest algorithm with the top ۲۰ features selected from Random forests for feature importance yields the best performance with an accuracy score of ۶۶.۱۲%, and the best accuracy score for Decision Tree also comes from the same method with a score of ۶۵.۰۸%. Although this result shows a slight superiority of the RF algorithm, the two ML algorithms' performance is not very different. In comparison between different methods, Random Forest for feature importance is the best method for RF, but Anova F gives better results for DT while chi score also got a high score in both tests. The result also showed that some attributes, like if the accident had happened in a rural or urban area or whether the vehicle had been towed, could be used for injury severity prediction, while some, like weather conditions, didn't affect the result significantly.