EEGHGNet: A Hybrid Global Dual-Branch Attention Module for a Lightweight and Accurate Motor Imagery-based Brain-Computer Interface

View at publisher Summarise with VIRAT
Authors: Jiajun Dai, Nobuaki Kobayashi
Cited by: 0
References: 13

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Classifying motor imagery (MI) obtained using electroencephalography (EEG) is essential in applications that rely on brain–computer interface (BCI) systems. However, EEG signals often exhibit low signal-to-noise ratios, strong inter-subject variability, and complex spatiotemporal dependencies, which pose a notable challenge for effective feature representation. To address these issues, we propose a lightweight EEG classification network called EEGHGNet designed to enhance temporal–channel feature modeling while maintaining a compact architecture. The network first employs a local spatial convolution to preserve neighborhood electrode relationships and subsequently introduces a hybrid global (HG) dual-branch module that generates temporal and channel attention maps through temporal and channel half-global pooling. Despite containing only 6.9k parameters, EEGHGNet achieved an average within-subject accuracy of 72.06% on the BCI Competition IV 2a dataset and thus showed competitive performance compared to other lightweight models. The results of ablation experiments further indicate that removing the HG module reduced accuracy to 69.72%, which highlights its contribution to the overall performance. In addition, the model attained an average cross-subject accuracy of 58.78% under a leave-one-subject-out evaluation. These results suggest that EEGHGNet offers a compact and effective solution for EEG classification that is suitable for real-time and resource-constrained BCI scenarios.

Fields of Study & Subjects

EEG and Brain-Computer Interfaces Muscle activation and electromyography studies Emotion and Mood Recognition

Publication details

Journal / source
International Winter Conference on Brain-Computer Interface (BCI)
Pages
1–6
Publication date
23 February 2026
Language
EN