Self-Supervised Learning With Adaptive Graph Modeling for EEG-Based Epileptic Seizure Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Signal Processing

Background:

• Epileptic seizure classification from EEG signals is challenging due to complex spatial-temporal dependencies.
• Limited labeled data and severe class imbalance hinder the development of accurate EEG analysis models.
• Existing methods struggle to effectively capture the intricate dynamics within EEG data.

Purpose of the Study:

• To develop a self-supervised learning framework, ASGPF, for robust EEG-based seizure classification.
• To address challenges of limited labeled data and class imbalance in EEG seizure detection.
• To create a data-efficient framework with potential for clinical applications.

Main Methods:

• Proposed the Adaptive Spatio-Graph Pretraining Framework (ASGPF) incorporating a novel Spatio-Graph Learning Cell (SGLC).
• SGLC dynamically constructs EEG topology, extracts spatial features using Gated Graph Neural Networks, and captures temporal dependencies with Gated Recurrent Units.
• Employed self-supervised sequence-to-sequence pretraining on unlabeled EEG data for representation learning.

Main Results:

• ASGPF significantly outperformed state-of-the-art methods on the TUSZ dataset, achieving weighted F1-scores of 83.8% (4-class) and 73.5% (8-class).
• The model achieved comparable performance to baselines trained on 75% less data when using only 25% of labeled data.
• Demonstrated effectiveness in data-scarce and class-imbalanced scenarios.

Conclusions:

• ASGPF effectively learns discriminative EEG representations via adaptive spatial-temporal modeling and self-supervised pretraining.
• The framework enables accurate seizure classification with minimal labeled data, highlighting its data efficiency.
• ASGPF shows strong potential for clinical application in resource-constrained environments.