Graph Convolution Neural Network Based End-to-End Channel Selection and Classification for Motor Imagery Brain–Computer Interfaces

Biao Sun, Zhengkun Liu, Zexu Wu, Chaoxu Mu, Ting Li

Abstract

Classification of electroencephalogram-based motor imagery (MI-EEG) tasks is crucial in brain–computer interface (BCI). EEG signals require a large number of channels in the acquisition process, which hinders its application in practice. How to select the optimal channel subset without a serious impact on the classification performance is an urgent problem to be solved in the field of BCIs. This article proposes an end-to-end deep learning framework, called EEG channel active inference neural network (EEG-ARNN), which is based on graph convolutional neural networks (GCN) to fully exploit the correlation of signals in the temporal and spatial domains. Two channel selection methods, i.e., edge-selection (ES) and aggregation-selection (AS), are proposed to select a specified number of optimal channels automatically. Two publicly available BCI Competition IV 2a (BCICIV 2a) dataset and PhysioNet dataset and a self-collected dataset (TJU dataset) are used to evaluate the performance of the proposed method. Experimental results reveal that the proposed method outperforms state-of-the-art methods in terms of both classification accuracy and robustness. Using only a small number of channels, we obtain a classification performance similar to that of using all channels. Finally, the association between selected channels and activated brain areas is analyzed, which is important to reveal the working state of brain during MI.

Introduction

Motor Imagery (MI) is a cornerstone of Brain-Computer Interfaces (BCIs), allowing users to control external devices by simply imagining movement. However, the practical deployment of MI-BCIs is hampered by the need for high-density EEG caps, often requiring 64 or more electrodes. This setup is time-consuming, uncomfortable, and introduces significant data redundancy. Furthermore, subject variability means that the optimal electrodes for one individual may not be effective for another. Existing channel selection methods often rely on fixed spatial layouts or manual intervention. This research introduces EEG-ARNN, an end-to-end framework that automates both channel selection and task classification by learning subject-specific functional connectivity.

The EEG-ARNN Framework

The core of the proposed system is the EEG Channel Active Inference Neural Network (EEG-ARNN). It integrates temporal feature extraction with spatial reasoning to optimize classification.

Architecture of the proposed EEG-ARNN framework, showing the integration of TFEM and CARM modules.

The architecture consists of two primary modules:

  1. Temporal Feature Extraction Module (TFEM): This CNN-based component processes raw EEG signals to extract time-domain features, eliminating the need for manual frequency-band filtering.
  2. Channel Active Reasoning Module (CARM): A novel GCN-based component that dynamically learns an adjacency matrix representing the functional connectivity between channels. Unlike traditional GCNs that use static physical distance, CARM learns the graph structure directly from the data via backpropagation.

Automated Channel Selection

Once the model is trained, the learned adjacency matrix is used to identify the most informative channels. The authors propose two methods for this selection:

  • Edge-Selection (ES): This method identifies the strongest connections (edges) in the learned graph and retains the electrodes associated with those edges.
  • Aggregation-Selection (AS): This method ranks electrodes based on their total "importance" or degree within the learned graph, selecting those that aggregate the most information.
Illustration of (a) Edge-Selection (ES) and (b) Aggregation-Selection (AS) methods for automated channel selection.

Experimental Results

The EEG-ARNN was evaluated on three datasets: BCICIV 2a, PhysioNet, and a self-collected TJU dataset.

  • Superior Accuracy: The model consistently outperformed state-of-the-art (SOTA) baselines like FBCSP and EEGNet. On the TJU dataset, it achieved 92.3% accuracy. It also reached 92.2% on BCICIV 2a and 82.0% on the PhysioNet dataset.
  • Efficiency and the "1/6th" Rule: A key finding was that using only 10 to 20 channels (approximately 1/6th of the original set) yielded performance nearly identical to using the full 64-channel array. Specifically, using the AS method with only 10 channels maintained an accuracy of 87.9%, outperforming many SOTA methods using all 64 channels.
  • Neurophysiological Alignment: The automatically selected channels (e.g., C3, Cz, CP3) aligned with the primary motor cortex and somatosensory areas, validating that the model focuses on biologically relevant brain regions.

Conclusion

EEG-ARNN provides a robust, end-to-end solution for MI-BCI by combining dynamic graph learning with automated channel selection. By significantly reducing the required number of electrodes without sacrificing accuracy, this framework moves BCIs closer to practical, real-world application. Future work will focus on reducing computational complexity and testing the system in real-time environments to handle live artifacts.

Summary updated

Abstract

Classification of electroencephalogram-based motor imagery (MI-EEG) tasks is crucial in brain–computer interface (BCI). EEG signals require a large number of channels in the acquisition process, which hinders its application in practice. How to select the optimal channel subset without a serious impact on the classification performance is an urgent problem to be solved in the field of BCIs. This article proposes an end-to-end deep learning framework, called EEG channel active inference neural network (EEG-ARNN), which is based on graph convolutional neural networks (GCN) to fully exploit the correlation of signals in the temporal and spatial domains. Two channel selection methods, i.e., edge-selection (ES) and aggregation-selection (AS), are proposed to select a specified number of optimal channels automatically. Two publicly available BCI Competition IV 2a (BCICIV 2a) dataset and PhysioNet dataset and a self-collected dataset (TJU dataset) are used to evaluate the performance of the proposed method. Experimental results reveal that the proposed method outperforms state-of-the-art methods in terms of both classification accuracy and robustness. Using only a small number of channels, we obtain a classification performance similar to that of using all channels. Finally, the association between selected channels and activated brain areas is analyzed, which is important to reveal the working state of brain during MI.
BibTeX
@article{sun2022graph,
  author = {Biao Sun and Zhengkun Liu and Zexu Wu and Chaoxu Mu and Ting Li},
  title = {{Graph Convolution Neural Network Based End-to-End Channel Selection and Classification for Motor Imagery Brain–Computer Interfaces}},
  journal = {IEEE Transactions on Industrial Informatics},
  year = {2022},
  doi = {10.1109/tii.2022.3227736},
}
Figure