American
International
University-
Bangladesh

Students & Supervisors

Abstract

Road traffic accidents are a very serious problem all over the world, causing huge economic losses and many deaths. Accidents result from a number of factors include road infrastructure, environment factors, traffic features and driver behavior. These are complex elements that interact in non-linear ways. Machine learning (ML) techniques is good at prediction accidents but are not used in safety-critical applications due to their limited interpretability. This study combines statistical analysis, data pre-processing, ensemble learning, cross-validation and explainable artificial intelligence (XAI) methods to provide an interpretable machine learning framework for prediction road traffic accidents. The proposed methodology framework uses statistical analyze and pre-processing techniques (missing value treatment, normalization, feature encoding, and class balancing) to improve data quality and reduce bias. The final preprocessed data set contained 1180 observations, 14 numerical predictor variables and a binary response variable on the occurrence of the accident. To avoid overfit and to ensure a robust performance evaluation, ten supervised machine learning classifiers were built and evaluated using stratified 10-fold cross-validation. Accuracy, precision, recall and f1-score were uses to evaluate the model’s performance. Experimental results prove that ensemble-based gradient boosting models outperformed the traditional baseline classifiers. Among all models, XGBoost achieved the highest prediction accuracy (94.3%), precision (93.8%), recall (94.1%), and F1-score (93.9%) and was a strong generalizer. The second-best accuracy was LightGBM with (92.7%) and lower computing complexity. To improvement the interpretability of the model, we used SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) to provide both global and local explanations for the model predictions. The explainability research shows that various important factors affect the probability of accidents such as traffic density, vehicle speed, road features, and environmental conditions. The proposed framework can help traffic authorities and legislators in designing data- driven plans to improve road safety and provide a clear and reliable answer for intelligent transportation systems.

Keywords

Publication Details

• Type of Publication:
• Conference Name: International Conference on Emerging Frontiers in Advanced Sciences and Technologies 2026
• Date of Conference: 27/06/2026 - 27/06/2026
• Venue: Pabna University of Science and Technology (PUST)
• Organizer: Pabna University of Science and Technology (PUST)