Deep learning with neural network for classifying severity levels of road accidents in thailand

Abstract

Road traffic accidents remain a critical global challenge, particularly in low- and middle-income countries like Thailand, where they impose substantial human and economic costs. This study proposes a comprehensive hybrid framework that integrates advanced machine learning, deep learning, and statistical simulation to enhance the prediction and analysis of accident severity. Utilizing an extensive dataset of Thai highway accidents from 2011 to 2023, the methodology begins with robust data preprocessing, employing the Weight of Evidence technique to transform discrete features, and a combination of SMOTE, Edited Nearest Neighbor, and Tomek Link to address class imbalance. The analytical phase utilizes a multi-stage approach: first, a Virtual Geometry Group (VGG) deep learning model—optimized via Bayesian optimization—is employed to maximize classification performance, evaluated by F1-score and Area Under the Curve (AUC). Second, feature importance is distilled through XGBoost, Cramer’s V, and SHAP analysis, identifying road location, environmental light, safety equipment, and road user type as primary fatality predictors. Furthermore, eigenvector analysis and Gaussian Mixture Model (GMM) clustering are applied to partition risk strata, while a Monte Carlo simulation based on Bayes’ theorem establishes a stable fatal accident probability of 0.16. By bridging high-performance predictive modeling with interpretable risk assessment, this framework provides a powerful tool for evidence-based policymaking and the strategic optimization of road safety interventions.