Enhancing Epileptic Seizure Identification by Exploring Swin Transformer Integration within Conformer Architecture

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Artificial Intelligence in Medicine

Background:

• Epileptic seizure identification from electroencephalogram (EEG) signals is crucial for patient management.
• Existing deep learning models like Conformer face challenges in capturing complex spatio-temporal patterns in EEG data.
• Class imbalance in EEG datasets often leads to biased model performance.

Purpose of the Study:

• To propose a novel Swin-Conformer model for enhanced epileptic seizure identification.
• To leverage the Swin Transformer's capabilities for improved local and global feature extraction from EEG signals.
• To address the class imbalance problem in EEG data using a weighted focal loss function.

Main Methods:

• The standard Vision Transformer (ViT) in the Conformer architecture was replaced with the Swin Transformer.
• The Swin Transformer's hierarchical patch merging and shifted-window self-attention were utilized.
• A weighted focal loss function was implemented to handle class imbalance during training.
• The model was evaluated on the CHB-MIT and Bonn EEG datasets using stratified 10-fold cross-validation.

Main Results:

• The Swin-Conformer model achieved high performance metrics: 99.24% accuracy, 99.55% specificity, and 98.47% sensitivity (segment-based).
• Event-based sensitivity reached 99.50%, outperforming existing state-of-the-art methods.
• Ablation studies confirmed the contribution of the Swin Transformer's multi-scale representation and attention mechanisms.

Conclusions:

• The proposed Swin-Conformer model significantly improves epileptic seizure identification accuracy and sensitivity.
• The Swin Transformer effectively captures local and long-range dependencies in EEG signals.
• This approach offers a promising advancement for automated seizure detection systems.