A Graph Based Hybrid Approach to Injury Severity Prediction in Road Accidents Using Deep Learning

Road traffic accidents are still one of the major problems all over the world that lead to a very high number of injuries and deaths. The accurate measurement of the severity of the injuries caused by these accidents is very essential indeed as a way of upgrading road safety, making the emergency response more efficient, and also, just to name a few, giving policy decisions that are based on evidence. Traditional statistical models, to a great extent, fail to capture all the complex nature of the relationships among different crash-related factors. In order to deal with the GNNRF (Graph Neural Network-Random Forest) is a combined machine learning model introduced by this research work. This model constitutes one of the best hypotheses of deep learning technologies for graphs combined with the strong points of ensemble learning to make predictions on crash injury severity. Essentially, by change crash data sets into graph format through k-nearest neighbor (kNN) connections and PCA-derived embeddings, the model becomes competent in finding hidden interdependencies among crash records. The experimental appreciations provide evidence that the performance of GNNRF is better than that of the traditional methods such as Random Forest and XGBoost, especially when it coherence with accuracy, and other evaluation metrics, such as precision, recall, and F1-score. Further, Explainable AI (XAI) techniques