MCTGNet: A Multi-Scale Convolution and Hybrid Attention Network for Robust Motor Imagery EEG Decoding

MCTGNet: A Multi-Scale Convolution and Hybrid Attention Network for Robust Motor Imagery EEG Decoding

Motor imagery (MI) EEG decoding is a key application in brain–computer interface (BCI) research. In cross-session scenarios, the generalization and robustness of decoding models are particularly challenging due to the complex nonlinear dynamics of MI-EEG signals in both temporal and frequency domain...

Full description

Saved in:
Bibliographic Details
Main Authors: Huangtao Zhan, Xinhui Li, Xun Song, Zhao Lv, Ping Li
Format: Article
Language:English
Published: MDPI AG 2025-07-01
Series:Bioengineering
Subjects:
motor imagery
EEG decoding
cross-session generalization
Kolmogorov–Arnold Network
Online Access:https://www.mdpi.com/2306-5354/12/7/775
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Motor imagery (MI) EEG decoding is a key application in brain–computer interface (BCI) research. In cross-session scenarios, the generalization and robustness of decoding models are particularly challenging due to the complex nonlinear dynamics of MI-EEG signals in both temporal and frequency domains, as well as distributional shifts across different recording sessions. While multi-scale feature extraction is a promising approach for generalized and robust MI decoding, conventional classifiers (e.g., multilayer perceptrons) struggle to perform accurate classification when confronted with high-order, nonstationary feature distributions, which have become a major bottleneck for improving decoding performance. To address this issue, we propose an end-to-end decoding framework, MCTGNet, whose core idea is to formulate the classification process as a high-order function approximation task that jointly models both task labels and feature structures. By introducing a group rational Kolmogorov–Arnold Network (GR-KAN), the system enhances generalization and robustness under cross-session conditions. Experiments on the BCI Competition IV 2a and 2b datasets demonstrate that MCTGNet achieves average classification accuracies of 88.93% and 91.42%, respectively, outperforming state-of-the-art methods by 3.32% and 1.83%.
ISSN:2306-5354

Similar Items