Hierarchical Dynamic Graph Convolutional Network With Interpretability for EEG-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Graph Convolutional Networks (GCNs) excel at analyzing electroencephalogram (EEG) channel topology for emotion recognition.
• Existing GCN methods often use single spatial patterns, neglecting local functional connectivity and raw EEG data dependencies.

Purpose of the Study:

• To propose a Hierarchical Dynamic GCN (HD-GCN) for enhanced EEG-based emotion recognition.
• To explore dynamic, multi-level spatial information and incorporate discriminative EEG signal features.

Main Methods:

• Developed a two-branch network: one for global dynamic information, another for local functional region augmentation.
• Utilized layer-wise adjacency matrices to increase GCN expressive power.
• Introduced a data-dependent Auxiliary Information Module (AIM) for multidimensional feature fusion.

Main Results:

• HD-GCN consistently outperformed state-of-the-art methods on the SEED and DREAMER datasets.
• Experiments validated the model's effectiveness in capturing dynamic and multi-level spatial-temporal EEG patterns.
• Interpretability analysis identified key brain regions and electrode pairs associated with emotional states.

Conclusions:

• The proposed HD-GCN model offers a significant advancement in EEG-based emotion recognition by effectively integrating dynamic spatial information and auxiliary features.
• The method demonstrates superior performance and provides insights into the neural correlates of emotion.
• HD-GCN presents a promising approach for developing more sophisticated brain-computer interfaces.