Multi-Task FT-Transformer: A High-Performance and Interpretable Framework for Traffic Accident Severity Prediction

Traffic accidents are a significant concern worldwide, leading to severe injuries, fatalities, and property damage. Predicting the severity of these outcomes is essential for enhancing road safety. However, the complex interdependencies among injuries, fatalities, and property damage, coupled with the influence of various dynamic factors, pose challenges for prediction. To address this issue, This study proposes a multi-task feature token transformer model with uncertainty-weighted loss optimization to predict the severity of injuries, fatalities, and property damage in traffic accidents. The model is trained on a large-scale dataset comprising urban road accidents in China, incorporating critical factors such as driver characteristics, vehicle types, road conditions, and environmental influences. The multi-task learning framework enhances information sharing across the prediction tasks for fatalities, injuries, and property damage, resulting in improved accuracy compared to single-task frameworks. Furthermore, we apply SHapley Additive Explanations to quantify feature importance and analyze how different input factors influence accident severity predictions, enhancing model transparency and interpretability by providing a clear understanding of the decision-making process. The results show that the proposed multi-task FT-Transformer model achieves good accuracy in all severity prediction tasks, with 70.31% for injury severity, 90.38% for fatal accident severity, and 80.79% for property damage. This framework helps policymakers target high-risk areas and optimize resource allocation, improving road safety and reducing traffic incidents.