Scholarlink

<p>EEG-based BCI systems hold significant promise for decoding motor intentions, especially in assistive and neurorehabilitative contexts. In this paper, we propose a hybrid feature-CNN framework that integrates manually extracted statistical and nonlinear features with a low-dimensional convolutional neural network (CNN) to classify motor tasks using EEG signals. Utilizing the MILimbEEG database, we extract 128 features per trial?comprising autoregressive (AR) parameters, Hjorth parameters, and Shannon entropy?across 16 EEG channels, capturing both temporal and spatial patterns of motor-related brain activity. A lightweight CNN is trained to classify eight distinct motor tasks with notable effectiveness. The model achieved a training accuracy of 95.25% and a test accuracy of 94.44%, demonstrating strong class-wise sensitivity and specificity. This research emphasizes the benefits of combining interpretable feature extraction with deep learning in BCI applications. However, limitations such as dataset size, reliance on executed rather than imagined tasks, absence of real-time testing, and simplified CNN architecture suggest directions for future development. The framework presents a robust and efficient approach for EEG-based motor intention decoding.</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

Hybrid EEG signal classification using manual feature extraction and CNN for motor task decoding

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Alireza Golkarieh,
  • Kiana Kiashemshaki,
  • Amirhossein Golshantafti

摘要

EEG-based BCI systems hold significant promise for decoding motor intentions, especially in assistive and neurorehabilitative contexts. In this paper, we propose a hybrid feature-CNN framework that integrates manually extracted statistical and nonlinear features with a low-dimensional convolutional neural network (CNN) to classify motor tasks using EEG signals. Utilizing the MILimbEEG database, we extract 128 features per trial?comprising autoregressive (AR) parameters, Hjorth parameters, and Shannon entropy?across 16 EEG channels, capturing both temporal and spatial patterns of motor-related brain activity. A lightweight CNN is trained to classify eight distinct motor tasks with notable effectiveness. The model achieved a training accuracy of 95.25% and a test accuracy of 94.44%, demonstrating strong class-wise sensitivity and specificity. This research emphasizes the benefits of combining interpretable feature extraction with deep learning in BCI applications. However, limitations such as dataset size, reliance on executed rather than imagined tasks, absence of real-time testing, and simplified CNN architecture suggest directions for future development. The framework presents a robust and efficient approach for EEG-based motor intention decoding.