Frequency-Gated Prompting for Enhancing Transformer-based EEG Decoding

The fundamental problem in electroencephalogram (EEG) decoding centers on the extraction of meaningful neural patterns from complex spatio-temporal signals. Although Transformer models have garnered significant attention in this domain due to their exceptional temporal modeling capabilities, their lack of perception for critical frequency-domain features within EEG signals constrains further performance enhancement. To address this issue, this paper proposes a lightweight approach termed frequency-gated prompted Transformer (FGPT). FGPT adaptively represents global EEG rhythms by introducing learnable sparse frequency prompt tokens. Utilizing a gated fusion mechanism, it synergistically embeds these tokens with the original EEG sequence into the Transformer's self-attention computation. This enables joint modeling of spatio-temporal and frequency-domain features without compromising sequence continuity. Experiments conducted on three public EEG datasets show that FGPT improves the decoding performance and robustness of the evaluated Transformer-based models (EEG-ViT, EEG-Conformer, and EEG-Deformer) on the selected baselines. With its lightweight design and observed potential for generalization on the datasets used, this approach explores a prompt learning method that may contribute to developing more efficient EEG decoding systems.