Millions of lives are lost in road accidents annually, underscoring the severity of traffic incidents. Despite numerous studies on crash severity forecasting with static ensemble machine learning, the potential of Dynamic Ensemble Selection (DES) for combining classifiers and improving performance remains relatively unexplored in this field. In contrast to previous research focused on static ensemble machine learning models, our study introduces an interpretable Bayesian Optimized Dynamic Ensemble Selection strategy, reinforced with instance hardness and region of competence considerations, for crash risk assessment. Utilizing six algorithms - Random Forest, Extra Tree, Cat Boost, Light GBM, KT Boost, and XGBoost - optimized through cross-validation, we employed bagging as a hybrid ensemble selection criterion to ensure ensemble diversity for DES models. Addressing class imbalance with cost-sensitive learning, we evaluated six DES algorithms, 'KNORA-E, KNORA-U, DES-KNN, DES-RRC, DES-KL, and META-DES' using a pool of input classifiers, based on key metrics such as Balanced Accuracy Score, Sensitivity, G-mean, and Matthews Correlation Coefficient. Our results revealed that the KNORA-E ensemble with Cat Boost achieved the highest performance, with a Balanced Accuracy (BCA) of 69.3%, Matthews Correlation Coefficient (MCC) of 0.36, and a G-mean score of 69.10%. DES-RRC demonstrated notable performance across all base classifiers, particularly in addressing imbalanced classes, and DES-RRC with LightGBM emerged as the second-best performer. SHAP analysis for the KNORAE-Cat Boost model identified key factors in crash severity as, road user gender, airbag deployment, vehicle age and road surface conditions. This approach could be a valuable tool for informed decision-making and targeted interventions.
