Hsiam: A Generalized Siamese Network for Efficient Multi-Resolution EEG Analysis

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• High-frequency electroencephalogram (EEG) provides detailed neural insights but faces challenges like high computational cost and noise.
• Existing EEG analysis methods struggle with varying data resolutions and noise interference, limiting clinical application.

Purpose of the Study:

• To introduce the High-Frequency Siamese Network (Hsiam), an efficient architecture for cross-resolution EEG analysis.
• To enhance the practical utility of high-frequency EEG data for clinical applications like seizure detection and treatment efficacy prediction.

Main Methods:

• Hsiam employs a dual-branch Siamese network with weight-sharing to process EEG data at different resolutions.
• Key components include the High-Frequency Activity Enhanced Module (HEM) for emphasizing relevant high-frequency signals and the Frequency-Domain Dropout Transformer (FD-former) for robust temporal-spectral modeling.
• An alignment loss function enforces consistency between different frequency resolutions during training.

Main Results:

• Hsiam demonstrated strong performance in both treatment efficacy prediction and seizure detection tasks across multiple datasets.
• The network effectively reduces computational overhead during inference by utilizing a single-branch input.
• Analyses confirmed that performance gains stem from cross-resolution alignment, task-specific spectral utilization, and channel contributions.

Conclusions:

• Hsiam offers a practical and computationally efficient solution for leveraging high-frequency EEG data.
• The proposed architecture successfully addresses the limitations of traditional EEG analysis, improving diagnostic capabilities.
• Hsiam's ability to integrate cross-resolution information enhances the robustness and accuracy of EEG-based neurodiagnostics.