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

Area of Science:

• Neuroscience
• Transportation Safety
• Artificial Intelligence

Background:

• Driver fatigue is a significant global road safety concern, especially for young drivers.
• Current policy interventions are limited by the absence of reliable fatigue detection technology.
• Accurate, interpretable, real-time fatigue monitoring is crucial for transportation safety management.

Purpose of the Study:

• To develop an end-to-end EEG-based fatigue detection model.
• To improve the accuracy and interpretability of driver fatigue monitoring systems.
• To provide actionable insights for policy and design recommendations to mitigate driver fatigue.

Main Methods:

• Proposed an EEG-based fatigue detection model: Scale-Enhanced Transformer and Dynamic Graph Convolutional Network (SET-DGCN).
• Integrated convolutional embeddings, attention mechanisms, and learnable graph structures to capture temporal and spatial brain interactions.
• Utilized component-level attribution (COAR) and SHapley Additive exPlanations (SHAP) for model interpretability and brain pattern analysis.

Main Results:

• SET-DGCN demonstrated superior accuracy and F1-score compared to CNN, GCN, and Transformer models.
• The model showed strong cross-subject generalization capabilities.
• Identified brain region-specific patterns associated with different fatigue stages.

Conclusions:

• The SET-DGCN model offers a reliable and interpretable solution for EEG-based driver fatigue detection.
• Neural insights from the model support multi-level policy and design recommendations for young driver safety.
• The findings provide a framework for mitigating fatigue in real-world transportation contexts.