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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) electroencephalogram (EEG) decoding is crucial for brain-computer interface (BCI) development.
• Cross-session BCI scenarios face challenges in model generalization and robustness due to nonlinear dynamics and distributional shifts in MI-EEG signals.
• Conventional classifiers struggle with high-order, nonstationary feature distributions, hindering decoding performance.

Purpose of the Study:

• To develop an end-to-end decoding framework, MCTGNet, for enhanced motor imagery EEG decoding.
• To address the limitations of current classifiers in handling complex feature distributions for improved cross-session generalization.
• To formulate MI-EEG classification as a high-order function approximation task integrating task labels and feature structures.

Main Methods:

• Proposed MCTGNet, an end-to-end decoding framework for MI-EEG.
• Introduced a group rational Kolmogorov-Arnold Network (GR-KAN) within the framework.
• Formulated classification as a high-order function approximation task, jointly modeling labels and feature structures.

Main Results:

• MCTGNet achieved average classification accuracies of 88.93% on the BCI Competition IV 2a dataset.
• MCTGNet achieved average classification accuracies of 91.42% on the BCI Competition IV 2b dataset.
• Outperformed state-of-the-art methods by 3.32% and 1.83% on the respective datasets.

Conclusions:

• MCTGNet enhances generalization and robustness in cross-session motor imagery EEG decoding.
• The GR-KAN approach effectively models complex feature distributions, overcoming previous bottlenecks.
• MCTGNet represents a significant advancement in BCI research for robust and accurate decoding.