Separated channel convolutional neural network to realize the training free motor imagery BCI systems

• Published in: Biomedical signal processing and control Vol. 49; pp. 396 - 403
• Main Authors: Zhu, Xuyang, Li, Peiyang, Li, Cunbo, Yao, Dezhong, Zhang, Rui, Xu, Peng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2019
• Subjects: Brain-computer interface; Common space pattern; Deep learning; Electroencephalography; Training free; Brain-computer interface; Deep learning; Electroencephalography; Common space pattern; Training free
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2018.12.027

[Display omitted] •The end-to-end deep learning framework to realize the training free motor imagery BCI is proposed.•Instead of log-energy from CSP filter, the multi-channels series in CSP data extracted from EEG are adopted as input.•A separated channel convolutional network, called SCCN, is proposed to encode the multi-channels EEG. In the recent context of Brain-computer interface (BCI), it has been widely known that transferring the knowledge of existing subjects to a new subject can effectively alleviate the extra training burden of BCI users. In this paper, we introduce an end-to-end deep learning framework to realize the training free motor imagery (MI) BCI systems. Specifically, we employ the common space pattern (CSP) extracted from electroencephalography (EEG) as the handcrafted feature. Instead of log-energy, we use the multi-channel series in CSP space to retain the temporal information. Then we propose a separated channel convolutional network, here termed SCCN, to encode the multi-channel data. Finally, the encoded features are concatenated and fed into a recognition network to perform the final MI task recognition. We compared the results of the deep model with classical machine learning algorithms, such as k-nearest neighbors (KNN), logistics regression (LR), linear discriminant analysis (LDA), and support vector machine (SVM). Moreover, the quantitative analysis was evaluated on our dataset and the BCI competition IV-2b dataset. The results have shown that our proposed model can improve the accuracy of EEG based MI classification (2–13% improvement for our dataset and 2–15% improvement for BCI competition IV-2b dataset) in comparison with traditional methods under the training free condition.