Scholarlink

<p>Electrode placement variability poses a critical challenge in EEG-based motor imagery tasks, often resulting in reduced classification robustness. We present the Adaptive Channel Mixing Layer (ACML), a plug-and-play preprocessing module that dynamically adjusts input signal weights through a learnable transformation matrix based on inter-channel correlations. By leveraging the inherent spatial structure of EEG caps, ACML effectively compensates for electrode misalignments and noise, enhancing resilience to signal distortion. Experimental validation on two motor imagery datasets with varying channel counts demonstrated consistent improvements in accuracy (up to 1.4%), kappa scores (up to 0.018), and robust performance across subjects, using five neural network architectures including a state-of-the-art model (ATCNet). Notably, ACML requires minimal computational overhead and no task-specific hyperparameter tuning, ensuring compatibility with diverse applications. This method offers a robust and efficient solution for advancing EEG-based motor imagery classification, with potential applications in real-time brain-computer interface systems and neurorehabilitation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive EEG preprocessing to mitigate electrode shift variability for robust motor imagery classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月19日

查看全文
  • Wenjing Xiong,
  • Lin Ma,
  • Haifeng Li

摘要

Electrode placement variability poses a critical challenge in EEG-based motor imagery tasks, often resulting in reduced classification robustness. We present the Adaptive Channel Mixing Layer (ACML), a plug-and-play preprocessing module that dynamically adjusts input signal weights through a learnable transformation matrix based on inter-channel correlations. By leveraging the inherent spatial structure of EEG caps, ACML effectively compensates for electrode misalignments and noise, enhancing resilience to signal distortion. Experimental validation on two motor imagery datasets with varying channel counts demonstrated consistent improvements in accuracy (up to 1.4%), kappa scores (up to 0.018), and robust performance across subjects, using five neural network architectures including a state-of-the-art model (ATCNet). Notably, ACML requires minimal computational overhead and no task-specific hyperparameter tuning, ensuring compatibility with diverse applications. This method offers a robust and efficient solution for advancing EEG-based motor imagery classification, with potential applications in real-time brain-computer interface systems and neurorehabilitation.