Road traffic accidents are a leading cause of fatalities worldwide, demanding an intelligent prevention strategies. This paper proposes a machine learning–driven framework for predicting accident severity by combining driver, environmental, and roadway factors. A stacked ensemble model, integrating Gradient Boosting, LightGBM, and a meta-learner, achieves superior predictive accuracy compared to individual classifiers. To enhance robustness, pre processing steps such as SMOTE-based balancing, normalization, and categorical encoding are employed. Experimental evaluation on a large-scale real-world dataset demonstrates consistently high performance across accuracy, precision, recall, and F1-score. Furthermore, explainability is incorporated through SHAP values, offering interpretable insights into critical risk factors. The system is designed to extend beyond off-line evaluation to real-time applications, allowing policymakers and traffic authorities to identify accident hotspots, implement preventive measures, and support infrastructure planning. The proposed approach contributes to the global challenge of road safety. It is scalable, interpretable, and highly performance-oriented.