Scholarlink

Accurate classification of motor imagery Electroencephalogram (EEG-MI) signals is essential for advancing brain-computer interfaces (BCIs). This study investigates the influence of various stages in a deep learning pipeline for EEG-MI classification. We evaluate five deep learning models-EEGNet v1, EEGNet v4, Shallow FBCSP Net, Deep4Net, and Inception MI-across different preprocessing techniques, including electrode selection, data standardization, bandpass filtering, and exponential standardization, as well as varying EEG sampling rates. Our results indicate minimal preprocessing, specifically electrode selection and data standardization, significantly enhances model performance by effectively managing EEG signal variability. On the benchmark BCI IV2a dataset, EEGNet v4 achieved the highest within-subject mean accuracy of 52.10% and the highest cross-subject mean accuracy of 51.10%. InceptionMI showed minimal subject variability, as indicated by a standard deviation of 7.26 across subjects. Additionally, EEGNet v4 demonstrated the highest cross-subject mean accuracy of 74.56% on the Cho2017 dataset. We also observed that sampling rates above 200 Hz degrade accuracy due to increased noise. These findings underscore the importance of optimizing preprocessing and sampling strategies, with EEGNet v4 emerging as a leading model for motor imagery classification. This research offers valuable insights for enhancing the application of deep learning in BCI systems.

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

Deep Learning Pipeline for EEG Classification: Evaluating Models, Preprocessing and Subject Generalizability

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Deepak Mewada,
  • Mohammed Fayiz Parappan,
  • Debasis Samanta

摘要

Accurate classification of motor imagery Electroencephalogram (EEG-MI) signals is essential for advancing brain-computer interfaces (BCIs). This study investigates the influence of various stages in a deep learning pipeline for EEG-MI classification. We evaluate five deep learning models-EEGNet v1, EEGNet v4, Shallow FBCSP Net, Deep4Net, and Inception MI-across different preprocessing techniques, including electrode selection, data standardization, bandpass filtering, and exponential standardization, as well as varying EEG sampling rates. Our results indicate minimal preprocessing, specifically electrode selection and data standardization, significantly enhances model performance by effectively managing EEG signal variability. On the benchmark BCI IV2a dataset, EEGNet v4 achieved the highest within-subject mean accuracy of 52.10% and the highest cross-subject mean accuracy of 51.10%. InceptionMI showed minimal subject variability, as indicated by a standard deviation of 7.26 across subjects. Additionally, EEGNet v4 demonstrated the highest cross-subject mean accuracy of 74.56% on the Cho2017 dataset. We also observed that sampling rates above 200 Hz degrade accuracy due to increased noise. These findings underscore the importance of optimizing preprocessing and sampling strategies, with EEGNet v4 emerging as a leading model for motor imagery classification. This research offers valuable insights for enhancing the application of deep learning in BCI systems.