Abstract
Drowsy driving is one of the primary causes of driving fatalities. Electroencephalography (EEG), a method for detecting drowsiness directly from brain activity, has been widely used for detecting driver drowsiness in real-time. Recent studies have revealed the great potential of using brain connectivity graphs constructed based on EEG data for drowsy state predictions. However, traditional brain connectivity networks are irrelevant to the downstream prediction tasks. This article proposes a connectivity-aware graph neural network (CAGNN) using a self-attention mechanism that can generate task-relevant connectivity networks via end-to-end training. Our method achieved an accuracy of 72.6% and outperformed other convolutional neural networks (CNNs) and graph generation methods based on a drowsy driving dataset. In addition, we introduced a squeeze-and-excitation (SE) block to capture important features and demonstrated that the SE attention score can reveal the most important feature band. We compared our generated connectivity graphs in the drowsy and alert states and found drowsiness connectivity patterns, including significantly reduced occipital connectivity and interregional connectivity. Additionally, we performed a post hoc interpretability analysis and found that our method could identify drowsiness features such as alpha spindles.
Introduction
Drowsy driving remains a significant contributor to road fatalities. While computer vision-based systems are common, they often struggle with environmental factors like low light or physical obstructions such as sunglasses. Electroencephalography (EEG) provides a direct physiological window into a driver's state. However, EEG data is inherently "non-Euclidean," meaning the physical placement of electrodes does not fully represent the complex functional relationships between brain regions.
Traditional approaches often rely on static connectivity graphs, such as simple channel correlations, which are "task-blind" and do not adapt to the specific classification goal. This research introduces the Connectivity-Aware Graph Neural Network (CAGNN), a model designed to learn the optimal brain network structure for drowsiness classification through end-to-end training.
Methodology
The CAGNN architecture integrates several advanced components to process spectral, spatial, and temporal information from EEG signals.
- Feature Extraction: Raw signals are transformed into frequency spectra (delta, theta, alpha, and beta bands) using Fast Fourier Transforms (FFT).
- Squeeze-and-Excitation (SE) Block: This module acts as a "feature spotlight," calculating importance scores for different frequency features to emphasize informative bands while suppressing noise.
- Self-Attention Graph Construction: The model utilizes a self-attention mechanism to learn a symmetric adjacency matrix, ensuring the resulting graph reflects task-relevant functional connectivity.
- DCRNN Encoder: A Diffusion Convolutional Recurrent Neural Network uses diffusion convolutions to propagate information across the learned graph, capturing interregional interactions.
- Temporal Modeling: A Gated Recurrent Unit (GRU) processes the sequence of states to capture the transition from alertness to drowsiness.
Experimental Setup
The model was validated using an EEG dataset from the Brain Research Center of National Chiao Tung University (NCTU). The study involved 27 subjects in a 360-degree VR driving simulator performing a 90-minute monotonous drive. Drowsiness was labeled based on Reaction Time (RT) to random lane departures. The analysis utilized both a balanced dataset (1,880 samples) and an imbalanced dataset (2,848 samples) to test real-world robustness.
Key Results and Insights
CAGNN achieved a state-of-the-art 72.6% accuracy on the balanced dataset, significantly outperforming baselines like EEGNet (61.6%) and ShallowConvNet (69.2%). On imbalanced data, it maintained a high F1-score of 73.8%.
- Biomarker Identification: The SE block revealed that the theta band (4-8 Hz) is the most critical feature for classification. Retraining the model on the theta band alone actually improved performance.
- Connectivity Patterns: The learned graphs showed a significant decrease in occipital region connectivity during drowsiness, supporting the "cortical gate" theory where the brain tunes out visual stimuli. Furthermore, lower interregional connectivity suggested a more "segregated" and less efficient brain state during fatigue.
- Model Interpretability: Post-hoc analysis confirmed the model identifies "alpha spindles"—short bursts of alpha activity—as a primary trigger for drowsiness prediction, aligning with established neuroscientific observations.
Conclusion
CAGNN demonstrates that learning task-relevant connectivity graphs significantly improves drowsiness detection while providing valuable neuroscientific insights. While the model is robust, future work will address inter-subject variability and the impact of motion artifacts in real-world driving environments to move toward practical deployment.
Summary updated