SET-DGCN: An end-to-end electroencephalography-based fatigue detection method for young drivers

Driver fatigue poses a critical threat to global road safety, particularly among young drivers. Nevertheless, policy-level interventions remain fragmented due to the lack of reliable and deployable detection technologies. Bridging this gap requires accurate, interpretable, and real-time fatigue monitoring systems capable of informing practical decision-making in transportation safety management. To address this challenge, we propose an end-to-end EEG-based fatigue detection model, Scale-Enhanced Transformer and Dynamic Graph Convolutional Network (SET-DGCN). The model captures multi-scale temporal dependencies and spatial brain-region interactions by integrating convolutional embeddings, attention mechanisms, and learnable graph structures. Extensive evaluations on both a driving simulation dataset and the publicly available SEED-VIG dataset confirm that SET-DGCN outperforms mainstream convolutional neural network (CNN)-based, graph convolutional network (GCN)-based, and Transformer-based models in terms of accuracy and F1-score, while maintaining strong cross-subject generalization. To enhance both interpretability and application relevance, a component-level attribution method (COAR) is employed to evaluate the functional contribution of model modules, while SHapley Additive exPlanations (SHAP) analysis is used to uncover brain region-specific patterns across fatigue stages. Based on these neural insights, a set of multi-level policy and design recommendations is proposed, ranging from infrastructure enhancements to adaptive in-vehicle systems and individualized interventions, to provide a comprehensive framework for mitigating fatigue among young drivers in real-world transportation contexts.