Emotion recognition using spatial-temporal EEG features through convolutional graph attention network

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Human emotion recognition is crucial for advancing brain-computer interfaces and machine intelligence.
• Electroencephalogram (EEG) signals offer a rich source of data for understanding emotional states.
• Developing efficient models for EEG-based emotion recognition remains a significant challenge.

Purpose of the Study:

• To develop an efficient human emotion recognition model using multi-channel EEG signals.
• To integrate temporal and spatial information from EEG for improved emotion classification.
• To enhance the generalization ability of emotion recognition models.

Main Methods:

• A Spatio-Temporal Feature Fused Convolutional Graph Attention Network (STFCGAT) was proposed.
• Combined single-channel differential entropy (DE) and cross-channel functional connectivity (FC) features.
• Employed a convolutional graph attention network with a multi-headed attention mechanism for feature fusion and extraction.

Main Results:

• The STFCGAT model achieved high classification accuracies on the SEED and DEAP datasets (e.g., 99.11% on SEED subject-dependent).
• Demonstrated state-of-the-art performance in cross-subject emotion recognition tasks.
• Ablation studies and feature analysis confirmed the model's effectiveness and highlighted brain's spatial-temporal differences across emotions.

Conclusions:

• The STFCGAT architecture is highly effective for human emotion recognition from EEG signals.
• The fusion of DE and FC features significantly enhances model performance.
• Distinct spatial-temporal brain characteristics are associated with different emotional states.