J. J. Liao, J. J. Luo, T. Yang, R. Q. Y. So, and M. C. H. Chua, “Effects of local and global spatial patterns in EEG motor-imagery classification using convolutional neural network,” null, pp. 1–10, Aug. 2020, doi: 10.1080/2326263X.2020.1801112.
Abstract:
Objective: An emerging idea that has gained traction in electroencephalogram motor-imagery (EEG-MI) classification is the “EEG-as-image” approach. The main idea is to preserve the spatial relationships of a set of EEG channels and to apply spatial filters which capture the local dynamics of the EEG signals. We hypothesize that due to the inherently global nature of EEG modulations coming from multiple dipolar sources, a better approach is to apply global unmixing filters to all relevant EEG electrodes instead.
Methods: Experiments were conducted to verify such hypothesis. Three deep learning models are proposed: (1) a model which applies multiple local spatial convolutions; (2) one which applies a global spatial convolution; and (3) a parallel architecture which combines the output feature maps of both (1) and (2). We tested our models on the BCI competition IV dataset 2a.
Results: The results of experiment showed that the global model achieved an overall classification accuracy of 75.0% and outperformed the local and parallel architectures by 3.3% and 2.3% respectively. It also outperformed the next best recorded result by 0.5%.
Conclusion: By exploring the impact of local and global spatial filters on EEG-MI classification, this paper helps to advance the study of EEG feature representation within a deep learning framework. The global spatial filters outperformed the local spatial filters for deep learning models in our study.
License type:
Funding Info:
This work is partially supported by BMRC-EDB JCO DP grants IAF311022 from the Agency for Science, Technology and Research Singapore.
Description:
This is an Accepted Manuscript of an article published by Taylor & Francis in Brain-Computer Interfaces on 13 August 2020, available online: http://www.tandfonline.com/10.1080/2326263X.2020.1801112.