Power and Phase Fusion Spectrogram with Three-Dimensional Convolution and Vision Transformer for Seizure Detection.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Epileptic seizure detection using electroencephalography (EEG) is vital for diagnosis and reducing clinical workload.
• Existing methods often underutilize phase information and lack synergy between local and global pattern analysis.
• There's a need for advanced frameworks that integrate diverse EEG signal characteristics.

Purpose of the Study:

• To develop and evaluate a novel seizure detection framework leveraging both power and phase information from EEG signals.
• To enhance the synergy between local time-frequency pattern extraction and global dependency modeling in seizure detection.
• To improve the reliability and accuracy of automated epileptic seizure detection.

Main Methods:

• A framework combining Continuous Wavelet Transform (CWT) for power and phase spectrogram generation.
• A hybrid network utilizing a 3D Convolutional Neural Network (3D-CNN) and a Vision Transformer (ViT) for feature extraction and dependency modeling.
• Fusion of power and phase spectrograms into a unified volume for joint analysis.

Main Results:

• Achieved high segment-level sensitivities (98.68% CHB-MIT, 92.05% SH-SDU) and specificities (98.33% CHB-MIT, 97.53% SH-SDU).
• Demonstrated strong event-level sensitivities (99.13% CHB-MIT, 96.08% SH-SDU) with low false detection rates (0.88/h CHB-MIT, 0.69/h SH-SDU).
• Ablation studies and visualizations validated the effectiveness of the joint power-phase input and hybrid network architecture.

Conclusions:

• The proposed framework effectively utilizes complementary power and phase information for epileptic EEG analysis.
• The hybrid 3D-CNN-ViT architecture demonstrates strong performance in capturing local and global dependencies.
• The method shows robust seizure detection capabilities on both public and clinical datasets, supporting patient-specific evaluation.