Feature fusion for a robust traffic accident assistance forecasting model with deep learning

Abstract

Millions of people are involved in traffic accidents each year. According to the World Health Organization (WHO), approximately 1.19 million people experience fatal outcomes, while 20 to 50 million suffer non-fatal consequences that may affect them for the rest of their lives. Multiple studies suggest that one of the main causes of this high number of affected individuals could be mitigated through the rapid intervention of medical assistance services. This paper thus proposes a general GTAAF model based on deep learning that predicts the severity of traffic accidents in real-time based on the surrounding characteristics, determining whether an accident requires medical assistance, using tabular data as input. This model can be utilized by emergency services to prioritize the allocation of medical resources in any region of the world. The main difference from existing state-of-the-art methods is that these are focused on specific localities and so cannot be applied universally, because each existing method is applied in a specific region of the world due to the individual characteristics of these data. To demonstrate this generalization capability, the GTAAF model was compared with six other state-of-the-art models applied to 8 populations in 3 different countries, showing a significant quality advantage over the other models (up to a 13.8% improvement in the F1-Score metric in classifying accidents that require assistance in the best case, and up to a 6.5% improvement in accidents that do not require assistance).