Feature fusion based on global-local weighted attention model for automatic epileptic seizure detection

Area of Science:

• Neurology
• Biomedical Engineering
• Signal Processing

Background:

• Epilepsy diagnosis and treatment are challenged by the complexity of electroencephalogram (EEG) signals.
• Current seizure detection methods struggle to integrate temporal, spatial, and spectral EEG features effectively.
• Limitations in accuracy and generalization hinder the clinical utility of existing EEG-based seizure detection systems.

Purpose of the Study:

• To develop an advanced model for improved seizure detection in epilepsy.
• To address the challenge of integrating diverse EEG signal features (temporal, spatial, spectral).
• To enhance the accuracy and generalizability of automated seizure detection.

Main Methods:

• Proposed the Global-Local Weighted Attention (GLWA) model.
• Integrated temporal, spatial, and spectral EEG features using a local-global attention mechanism.
• Balanced global and local feature extraction for comprehensive EEG signal analysis.

Main Results:

• Achieved 98.82% accuracy on the CHB-MIT dataset.
• Achieved 98.89% accuracy on the Siena dataset.
• Demonstrated effective integration of EEG features leading to superior detection performance.

Conclusions:

• The GLWA model shows superior performance in seizure detection.
• Model visualization provides insights into attention distribution across brain regions and frequencies.
• GLWA offers a robust, accurate, and generalizable approach for seizure identification with enhanced interpretability.