Self-supervised spatial-temporal contrastive network for EEG-based brain network classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG)-based brain network analysis is crucial for understanding brain diseases.
• Current methods face challenges in leveraging large unlabeled datasets for spatial and temporal brain connectivity analysis.
• High costs of data annotation limit the application of advanced machine learning models.

Purpose of the Study:

• To develop a novel Self-Supervised Spatial-Temporal Contrastive Network (SS-STCN) for brain network classification.
• To extract advanced feature representations from unlabeled EEG data, reducing annotation costs.
• To improve the accuracy and generalizability of brain network analysis.

Main Methods:

• A self-supervised contrastive learning framework (SS-STCN) was designed.
• Attention-driven two-stream encoders, including Spatial Graph Attention Network (SGAT) and Temporal Bi-directional Long Short-Term Memory (TBLSTM), were trained.
• Spatial-temporal feature fusion was achieved by utilizing the trained hybrid networks on labeled data.

Main Results:

• The SS-STCN framework demonstrated superior performance compared to existing supervised and unsupervised methods.
• Experiments on the CHB-MIT and Deap datasets showed high accuracy in epilepsy classification and emotion recognition.
• The method effectively captured spatial and temporal relationships in brain networks.

Conclusions:

• The SS-STCN framework offers a powerful approach for brain network classification using unlabeled EEG data.
• This method significantly enhances feature extraction and model generalizability.
• SS-STCN provides a cost-effective solution for brain disease research and related applications.