Epilepsy detection based on multi-head self-attention mechanism.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) show promise in electroencephalogram (EEG) signal detection but struggle with global perception and generalization due to individual differences.
• Existing epilepsy detection models often exhibit weak generalization capabilities, limiting their effectiveness across diverse patient populations.

Purpose of the Study:

• To develop a cross-patient epilepsy detection method that overcomes the limitations of existing models, particularly in generalization.
• To enhance the accuracy and reliability of epilepsy detection from EEG signals by integrating advanced deep learning techniques.

Main Methods:

• Utilized Short-Time Fourier Transform (STFT) to convert raw EEG signals into time-frequency features.
• Employed Convolutional Neural Networks (CNNs) to model local information within the time-frequency features.
• Integrated a multi-head self-attention mechanism from the Transformer architecture to capture global dependencies and relationships between features.
• Implemented a lightweight multi-head attention module with an alternating structure to extract multi-scale features efficiently and reduce computational costs.

Main Results:

• The proposed model achieved high performance metrics on the CHB-MIT dataset: 92.89% accuracy, 96.17% sensitivity, 92.99% specificity, 94.41% F1 score, and 96.77% AUC.
• Demonstrated superior performance and improved generalization capabilities compared to existing epilepsy detection methods.

Conclusions:

• The developed cross-patient epilepsy detection method effectively utilizes a multi-head self-attention mechanism to enhance global perception and model generalization.
• The integration of STFT, CNNs, and Transformers offers a robust approach for accurate and reliable epilepsy detection from EEG signals, outperforming previous methods.