A time-frequency cross-attention network model for epileptic seizure detection

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy diagnosis relies heavily on electroencephalography (EEG) signals.
• Current EEG analysis often underutilizes frequency domain information, limiting diagnostic accuracy.
• Integrating time and frequency domain features presents a significant challenge in epilepsy research.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced epilepsy detection.
• To effectively integrate time and frequency domain features from EEG signals.
• To overcome limitations of traditional models in comprehensive feature extraction and fusion.

Main Methods:

• Proposed a novel Time-Frequency Cross-Attention Network (TFCANet) utilizing residual attention mechanisms.
• Employed Fast Fourier Transform (FFT) to convert time-domain EEG features to frequency-domain.
• Integrated SE Residual modules for frequency domain and Residual Window Multi-head Self-Attention (ResWMSA) for time domain feature extraction.
• Utilized cross-attention for effective inter-modal feature fusion.

Main Results:

• Achieved 96.15% accuracy on the HMS-Harmful Brain Activity Classification dataset (5-category task).
• Attained 93.63% accuracy on the University of Bonn dataset (5-category task).
• Demonstrated superior performance of time-frequency feature fusion compared to single-modality approaches.

Conclusions:

• The TFCANet model successfully integrates time and frequency domain EEG features for robust epilepsy detection.
• Time-frequency feature fusion significantly enhances diagnostic accuracy in epilepsy classification.
• The proposed model shows promise for improving clinical diagnosis of epilepsy.