Frequency-Gated Prompting for Enhancing Transformer-based EEG Decoding

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Electroencephalogram (EEG) decoding faces challenges in extracting neural patterns from complex signals.
• Transformer models excel at temporal modeling but neglect crucial frequency-domain EEG features.
• Existing methods struggle to fully leverage both spatio-temporal and frequency information in EEG data.

Purpose of the Study:

• To propose a novel, lightweight approach for enhancing EEG decoding by incorporating frequency-domain information.
• To address the limitations of Transformer models in capturing critical frequency features within EEG signals.
• To improve the performance and robustness of Transformer-based EEG decoding systems.

Main Methods:

• Introduced a frequency-gated prompted Transformer (FGPT) model.
• Developed learnable sparse frequency prompt tokens to represent global EEG rhythms.
• Employed a gated fusion mechanism to integrate frequency prompts with original EEG sequences in self-attention.
• Enabled joint modeling of spatio-temporal and frequency-domain features without disrupting sequence continuity.

Main Results:

• FGPT significantly improved the decoding performance of Transformer-based models (EEG-ViT, EEG-Conformer, EEG-Deformer).
• Enhanced robustness of EEG decoding was observed across three public datasets.
• The lightweight design of FGPT demonstrated potential for generalization.

Conclusions:

• FGPT offers an effective method for integrating frequency-domain features into EEG decoding with Transformers.
• The proposed prompt learning approach contributes to developing more efficient and accurate EEG decoding systems.
• FGPT successfully models spatio-temporal and frequency-domain characteristics of EEG signals.