Crash Injury Prediction Using Bidirectional Gated Recurrent Unit with Multi Head Attention Network

In this era, Crash injury prediction is a crucial area of research aimed at reducing road traffic mortalities and injuries, which remain a leading global cause of death. Previous researchers have suggested various methods but still there is need to improve the accuracy and reliability of these predictions by analyzing large, multi-dimensional datasets from various sources, road conditions, and driver data. Therefore, this paper suggests an approach to predict the severity of injury occurred during the vehicle crash. Firstly, the data is collected from the Ministry of transport and logistics services. Then, the accumulated data is processed through a sampling technique namely Adaptive Synthetic (ADASYN). The sampled data is embedded into the model combined with correlation based selection with Recursive Feature Elimination (RFE) techniques for extraction of suitable features and selection of optimal attributes based on ranking method. Then based on the selected optimal features, injury prediction is initiated by employing Bidirectional Gated Recurrent Unit (BiGRU) with Multi-head Attention Network (MAN) technique. Finally the overall interpretation of the model is developed by SHapley Additive exPlanations (SHAP). The proposed model namely BiGRU-MAN-SHAP resulted with 93.63% accuracy.