Scholarlink

Brain-Computer Interfaces (BCIs) offer a purely cognition-based modality for human interaction with electronic devices (HCI), which is particularly lucrative for people with challenged motor skills. The Steady-State Visually Evoked Potentials (SSVEP) paradigm is easy to master for BCI users, but generally disregards their individual variability in reactions to the photostimuli at different frequencies. In our paper, we demonstrate how the individual frequency selection algorithm that we previously proposed for SSVEP-BCIs can improve the accuracy for convolutional neural network models (CNNs) tasked with the classification of the commands. For this end, we employ two open datasets containing electroencephalography (EEG) data and eight selected CNN models: ATCNet, DeepConvNet, EEGNet, EEGNeX, EEG-TCNet, MBEEG_SENet, ShallowConvNet, and TCNet_Fusion. Our results suggest that for most models the individual frequency selection indeed led to statistically significant improvements. The leaders were ATCNet and EEG-TCNet with 95.8% and 94.6% classification accuracy respectively; and the improvement for the latter model was about +10% compared to the no frequency selection condition. The inference times were under 0.3 s, which suggests the possibility of using the proposed approach in real-time BCIs for classifying the commands.

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

Benchmarking Individual Frequency Selection for BCI SSVEP Paradigm with State-of-the-Art CNN Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Alexey Kozin,
  • Anton Gerasimov,
  • Maxim Bakaev

摘要

Brain-Computer Interfaces (BCIs) offer a purely cognition-based modality for human interaction with electronic devices (HCI), which is particularly lucrative for people with challenged motor skills. The Steady-State Visually Evoked Potentials (SSVEP) paradigm is easy to master for BCI users, but generally disregards their individual variability in reactions to the photostimuli at different frequencies. In our paper, we demonstrate how the individual frequency selection algorithm that we previously proposed for SSVEP-BCIs can improve the accuracy for convolutional neural network models (CNNs) tasked with the classification of the commands. For this end, we employ two open datasets containing electroencephalography (EEG) data and eight selected CNN models: ATCNet, DeepConvNet, EEGNet, EEGNeX, EEG-TCNet, MBEEG_SENet, ShallowConvNet, and TCNet_Fusion. Our results suggest that for most models the individual frequency selection indeed led to statistically significant improvements. The leaders were ATCNet and EEG-TCNet with 95.8% and 94.6% classification accuracy respectively; and the improvement for the latter model was about +10% compared to the no frequency selection condition. The inference times were under 0.3 s, which suggests the possibility of using the proposed approach in real-time BCIs for classifying the commands.