HCFT: A Hierarchical Convolutional Fusion Transformer for Cross-Task EEG Decoding

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) decoding faces challenges due to non-stationary neural signals and limited model generalization.
• Existing methods struggle to adapt across different tasks and subjects, hindering widespread application.
• The need for robust and generalizable EEG decoding frameworks is critical for advancing brain-computer interfaces.

Purpose of the Study:

• To introduce a novel, lightweight, and generalizable decoding framework for EEG signals.
• To enhance multi-scale EEG representation learning by integrating convolutional and Transformer architectures.
• To improve the performance and adaptability of EEG decoding systems across diverse tasks and subjects.

Main Methods:

• Developed the Hierarchical Convolutional Fusion Transformer (HCFT) framework, utilizing dual-branch convolutional encoders and hierarchical Transformer blocks.
• Implemented a cross-attention mechanism for feature alignment between convolutional branches at each stage.
• Employed a hierarchical Transformer for global dependency encoding and Dynamic Tanh normalization for stabilization.

Main Results:

• HCFT achieved 80.83% accuracy and 0.6165 Cohen's kappa on the BCI Competition IV-2b dataset.
• On the CHB-MIT dataset, HCFT demonstrated 99.10% sensitivity, 0.0236 false positives/hour, and 98.82% specificity.
• The framework consistently outperformed over ten state-of-the-art baseline methods in both event-related classification and seizure prediction.

Conclusions:

• The proposed HCFT framework offers a scalable and versatile solution for general-purpose neural decoding.
• HCFT demonstrates superior performance and cross-subject generalization compared to existing EEG decoding methods.
• The model's structural interpretability provides insights into its effective feature learning mechanisms.