A Modified Transformer Network for Seizure Detection Using EEG Signals.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Seizures significantly impair epileptic patients' physical function and daily life.
• Automated seizure detection is crucial for timely clinical intervention and patient management.
• Current deep learning models face challenges in effectively extracting both local and global features from electroencephalography (EEG) signals.

Purpose of the Study:

• To propose an enhanced transformer network, Inresformer, for improved automated seizure detection.
• To leverage Inception and Residual networks within the transformer architecture for richer feature representation.
• To enhance the nonlinear representation capabilities of the model for more accurate seizure recognition.

Main Methods:

• Utilized discrete wavelet transform (DWT) for EEG signal decomposition into three sub-bands.
• Employed the Co-MixUp method to address data imbalance issues.
• Developed the Inresformer network with Inception, Residual, and modified Feedforward layers for seizure detection.
• Implemented discriminant fusion for final seizure recognition based on multi-scale EEG sub-signals.

Main Results:

• Achieved 100% accuracy on the Bonn dataset.
• Attained an average accuracy of 98.03% on the CHB-MIT dataset.
• Demonstrated high sensitivity (95.65%) and specificity (98.57%) on the CHB-MIT dataset.
• Outperformed existing deep learning networks in seizure detection performance.

Conclusions:

• The Inresformer network offers a promising approach for automated seizure detection.
• The proposed method shows significant potential for clinical research and diagnosis applications.
• The enhanced feature extraction and nonlinear representation contribute to competitive seizure recognition performance.