HRGNN: hierarchical region-aware graph neural network for interpretable EEG-Based emotion recognition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Cortical information processing involves complex interactions among brain regions.
• Existing electroencephalogram (EEG)-based emotion recognition methods often overlook functional specialization and local spatial dependencies, leading to limited feature representation.
• Conventional Graph Neural Networks (GNNs) can suffer from over-smoothing and global average pooling may ignore critical brain regions.

Purpose of the Study:

• To propose a novel Hierarchical Region-Aware Graph Neural Network (HRGNN) for enhanced EEG-based emotion recognition.
• To address limitations in current GNNs by incorporating brain region specialization and multi-scale interactions.
• To improve the granularity and discriminative power of emotion recognition from EEG signals.

Main Methods:

• Developed a Brain Region Embedding (BRE) module to map whole-brain EEG signals into structured multi-region subgraph representations based on topological priors.
• Introduced a region-aware graph encoder combining regional aggregation and hierarchical graph pooling to capture multi-scale intra- and inter-regional interactions.
• Integrated a Dynamic Routing-based Mixture-of-Experts (DMoE) module for adaptive fusion of regional subgraph features, prioritizing emotionally relevant brain regions.

Main Results:

• The proposed HRGNN model consistently outperformed state-of-the-art methods across eight publicly available EEG emotion datasets.
• Extensive experiments validated the model's effectiveness in emotion discrimination.
• Visualization analyses confirmed the model's alignment with neuroscientific principles of brain region interaction.

Conclusions:

• The HRGNN model offers a significant advancement in EEG-based emotion recognition by effectively modeling regional specialization and interactions.
• The HRGNN architecture provides more granular and discriminative feature representations compared to conventional GNNs.
• The findings suggest HRGNN's potential for more accurate and neuroscientifically grounded emotion recognition systems.