Enhancing Driver Injury Severity Prediction Using Optimized Oversampling and Feature Selection Techniques

Globally, it is widely acknowledged that driver errors are responsible for approximately 90% of traffic accidents. Identifying the potential risk factors affecting driver injuries' severity is crucial for informing better decision-making and selecting more effective countermeasures. This study uses a new and extensive dataset to predict driver injury severity. We employ the SMOTE-ENN technique to address the class imbalance accident datasets, which simultaneously applies both oversampling and undersampling to create a balanced and robust dataset. Additionally, significant features are selected using Relief-F, Extra Tree, and Recursive Feature Elimination methods. Driver injury severities are then classified using three different machine learning algorithms. The results demonstrate that the Random Forest algorithm achieved an F1 score of 93% using a reduced set of features.