Predicting Urban Bus Crash Injury Severity in Imbalanced Datasets Using Optimised Machine Learning
This study analysed police-reported bus crash data in Dhaka City, Bangladesh from 2015 to 2022, highlighting the significant class imbalance between fatal and non-fatal incidents that challenges traditional machine learning approaches. To address this gap, this study evaluated six ensemble frameworks combining tree-based models (Random Forest, Gradient Boosting, and XGBoost) with baseline classifiers (Logistic Regression, Support Vector Machine, and CART). Various sampling techniques were used, including non-sampling, resampling (SMOTE, Random Oversampling), and training strategies (No Bagging, Bagging, optimised via GridSearchCV). Results show that Logistic Regression, when optimised with GridSearchCV and combined with SMOTE resampling, achieved the best performance metrics: accuracy = 0.745, AUC = 0.697, F1-Score for Non-Fatal = 0.840, and Precision for Fatal = 0.600. This model outperformed ensemble methods, such as Gradient Boosting (AUC = 0.702) and Random Forest (AUC = 0.680). While Gradient Boosting effectively handled feature interactions, Logistic Regression demonstrated a 13% improvement in accuracy for fatal cases with hyperparameter tuning. Resampling had a limited impact on tree-based models; for example, the AUC of Random Forest increased marginally from 0.643 to 0.648 with bagging. Moreover, the Support Vector Machine struggled with precision in fatal cases, with values ranging from 0.000 to 0.420, even with random oversampling. The key predictors identified included pedestrian casualties, vehicle fitness certification, driver age, seatbelt use, drunk driving incidents, and road geometry. These findings support a model for predicting bus crash injury severity in Dhaka City and similar urban areas, emphasising the need for data-driven strategies to prevent crashes that result in fatality or serious injuries.