SS-EMERGE - self-supervised enhancement for multidimension emotion recognition using GNNs for EEG

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) signals present challenges for emotion recognition due to low signal-to-noise ratio and high-frequency characteristics.
• Existing self-supervised learning (SSL) methods often underperform fully-supervised techniques in cross-subject EEG emotion recognition tasks.

Purpose of the Study:

• To introduce a novel hybrid SSL framework, SS-EMERGE, designed to enhance cross-subject EEG-based emotion recognition.
• To improve the accuracy and efficiency of emotion recognition from EEG data, particularly when labeled data is scarce.

Main Methods:

• Developed a hybrid SSL framework (SS-EMERGE) integrating Causal Convolutions for temporal features, Graph Attention Transformers (GAT) for spatial modeling, and Spectral Embedding for spectral analysis.
• Employed meiosis-based contrastive learning for pretraining, followed by fine-tuning with minimal labeled data to enrich dataset specificity.
• Utilized Leave-One-Subject-Out (LOSO) cross-validation for robust performance evaluation.

Main Results:

• SS-EMERGE achieved high LOSO accuracies of 92.35% on the SEED dataset and 81.51% on the SEED-IV dataset.
• A foundation model pre-trained on combined SEED and SEED-IV datasets demonstrated performance comparable to individual models.
• The framework effectively leverages unlabeled data to improve EEG emotion recognition.

Conclusions:

• SS-EMERGE significantly advances EEG-based emotion recognition by achieving high accuracy with limited labeled data.
• The proposed framework offers a promising direction for developing more robust and generalizable emotion recognition systems.
• The foundation model approach shows potential for broader applications in EEG analysis.