Scholarlink

<p>Electroencephalography (EEG) provides a non-invasive means for decoding covert linguistic processes such as inner speech. In this study, we present an improved approach for inner-speech EEG decoding using a Pyramid Graph Convolutional Network (GCN) designed to model multi-scale brain connectivity. The method classifies four command states like up, down, left, and right across three cognitive conditions: inner speech, pronounced speech, and visualized imagination. Unlike earlier studies that explored inner-speech recognition using CNNs or graph-based architectures, our work introduces a hierarchical graph-coarsening framework that progressively reduces the 128-channel EEG graph while preserving essential neurophysiological structure. The approach integrates differential-entropy-based features, Pearson-correlation-guided graph construction, and METIS-based coarsening to extract discriminative multi-resolution representations. The proposed model achieves accuracies of 79.51% for inner speech, 85.18% for pronounced speech, and 80.81% for visualized conditions, demonstrating improved decoding performance without claiming first-ever classification. These findings highlight the potential of multi-scale GCNs for decoding subtle neural patterns associated with silent verbalization and offer promising directions for future brain–computer interface research.</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

Constructing differential entropy feature graphs with prior knowledge for inner-speech EEG analysis using a pyramid GCN model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月16日

查看全文
  • Xiaojing Hao,
  • Xiaoqi Lu,
  • Dahua Yu

摘要

Electroencephalography (EEG) provides a non-invasive means for decoding covert linguistic processes such as inner speech. In this study, we present an improved approach for inner-speech EEG decoding using a Pyramid Graph Convolutional Network (GCN) designed to model multi-scale brain connectivity. The method classifies four command states like up, down, left, and right across three cognitive conditions: inner speech, pronounced speech, and visualized imagination. Unlike earlier studies that explored inner-speech recognition using CNNs or graph-based architectures, our work introduces a hierarchical graph-coarsening framework that progressively reduces the 128-channel EEG graph while preserving essential neurophysiological structure. The approach integrates differential-entropy-based features, Pearson-correlation-guided graph construction, and METIS-based coarsening to extract discriminative multi-resolution representations. The proposed model achieves accuracies of 79.51% for inner speech, 85.18% for pronounced speech, and 80.81% for visualized conditions, demonstrating improved decoding performance without claiming first-ever classification. These findings highlight the potential of multi-scale GCNs for decoding subtle neural patterns associated with silent verbalization and offer promising directions for future brain–computer interface research.