GC-BERT: A Multi-branch Feature Fusion Framework for Alertness State Recognition with Interpretable Neural Representations

Background: Alertness state recognition is of great significance in fields such as traffic safety and medical monitoring. Traditional EEG analysis methods suffer from the problems of high subjectivity of feature extraction and insufficient spatio-temporal feature modeling. This study aims to develop a novel deep learning framework to achieve high-precision classification and interpretable analysis of EEG signals.
Methods: A multi-branch hybrid GC-BERT architecture is proposed, which innovatively integrates graph convolutional networks (GCN), convolutional neural networks (CNN), and Transformers. The GCN branch (GB) captures the spatial topological structure of brain regions through electrode adjacency matrices, while the Channel-wise CNN Branch (CB) performs channel-independent convolutions on the power spectral density (PSD). The features obtained from the dual branches are fused through cross-attention. The Transformer encoder processes the fused features and introduces a CLS token to enable global representation learning. The dataset is constructed based on the Psychomotor Vigilance Task (PVT), utilizing 19-channel EEG signals, and ensures data quality through rigorous preprocessing and segmented averaging.
Results: The model achieved an accuracy rate of 85.70% on the test set, significantly outperforming traditional machine learning methods and single-architecture models. Interpretability analysis showed that an increase in gamma band power corresponds to a high alertness state, while an increase in alpha band power corresponds to a low alertness state. The high contribution of frontal lobe activity to classification is consistent with known neural mechanisms.
Conclusions: GC-BERT not only achieves high-precision classification of EEG signals, but also reveals the physiological basis of model decisions through innovative and interpretable analysis methods, providing an effective solution for EEG signal analysis and demonstrating significant potential in real-time monitoring and clinical applications.