Comparative Study for Optimized Deep Learning-Based Road Accidents Severity Prediction Models

Road traffic accidents remain a major cause of fatalities and injuries worldwide. Effective classification of accident type and severity is crucial for prompt post-accident protocols and the development of comprehensive road safety policies. This study explores the application of deep learning techniques for predicting crash injury severity in the Eastern Province of Saudi Arabia. Five deep learning models were trained and evaluated, including various variants of feedforward multilayer perceptron, a back-propagated artificial neural network (ANN), an ANN with radial basis function (RPF), and tabular data learning network (TabNet). The models were optimized using Bayesian optimization (BO) and employed the synthetic minority oversampling technique (SMOTE) for oversampling the training dataset. While SMOTE enhanced balanced accuracy for ANN with RBF and TabNet, it compromised precision and increased recall. The results indicated that oversampling techniques did not consistently improve model performance. Additionally, significant features were identified using least absolute shrinkage and selection operator (LASSO) regularization, feature importance, and permutation importance. The results indicated that oversampling techniques did not consistently improve model performance. While SMOTE enhanced balanced accuracy for ANN with RBF and TabNet, it compromised precision and increased recall. The study's findings emphasize the consistent significance of the 'Number of Injuries Major' feature as a vital predictor in deep learning models, regardless of the selection techniques employed. These results shed light on the pivotal role played by the count of individuals with major injuries in influencing the severity of crash injuries, highlighting its potential relevance in shaping road safety policy development.