Cross-Level Topological Framework: Learning Explainable Region-Channel Representations from EEG Signals for Emotional...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) based emotion recognition commonly uses graph neural networks (GNNs) on full-channel signals.
• Existing methods often overlook detailed interactions within and between brain regions, limiting representational power.

Purpose of the Study:

• To propose an explainable cross-level topological network (ECTN) for enhanced EEG emotion recognition.
• To investigate and capture functional brain interactions from both channel-level and region-level perspectives.

Main Methods:

• Developed the ECTN framework with three modules: cross-region topological feature fusion, specific-region position-guided attention, and bidirectional gated fusion.
• Decoupled EEG functional interactions into global region interactions and local region dynamics.
• Integrated channel-level and region-level features using the bidirectional gated fusion module, considering inclusion relationships.

Main Results:

• The ECTN model demonstrated superior performance on the SEED, SEED-IV, and SEED-V datasets.
• Experimental results validated the effectiveness of exploring both channel-wise and region-wise interactions for emotion recognition.

Conclusions:

• The proposed ECTN advances EEG emotion recognition by effectively modeling cross-level brain interactions.
• Capturing both global and local brain dynamics significantly improves recognition accuracy.