Abstract
Road traffic accidents are among the top ten major causes of fatalities in the world, taking millions of lives annually. Machine-learning ensemble classifiers have been frequently used for the prediction of traffic injury severity. However, their inability to comprehend complex models due to their "black box" nature may lead to unrealistic traffic safety judgments. First, in this research, we propose three state-of-the-art Dynamic Ensemble Learning (DES) algorithms including Meta-Learning for Dynamic Ensemble Selection (META-DES), K-Nearest Oracle Elimination (KNORAE), and Dynamic Ensemble Selection Performance (DES-P), with Random Forest (RF), Adaptive Boosting (AdaBoost), Classification and Regression Tree (CART), and Binary Logistic Regression (BLR) as the base learners. The DES algorithm automatically chooses the subset of classifiers most likely to perform well for each new test instance to be classified when generating a prediction, making it more efficient and flexible. The META-DES model using RF as the base learner outperforms other models with accuracy (75%), recall (69%), precision (71%), and F1-score (72%). Afterwards, the risk factors are analyzed with SHapley Additive exPlanations (SHAP). The driver's age, month of the year, day of the week, and vehicle type influence SHAP estimation the most. Young drivers are at a heightened risk of fatal accidents. Weekends and summer months see the most fatal injuries. The proposed novel META-DES-RF algorithm with SHAP for predicting injury severity may be of interest to traffic safety researchers.