Enhancing EEG Motor Imagery Decoding Performance via Deep Temporal-domain Information Extraction

• Published in: Data Driven Control and Learning Systems Conference (Online) pp. 420 - 424
• Main Authors: Yang, Qihong, Yang, Mingzhao, Liu, Ke, Deng, Xin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 03.08.2022
• Subjects: Brain modeling; Brain-computer Interface; Brain-computer interfaces; Convolution; Convolutional codes; Data models; Decoding; EEGNet; Electroencephalography; Motor Imagery; Temporal Convolutional Network
• ISSN: 2767-9861
• DOI: 10.1109/DDCLS55054.2022.9858575

Electroencephalography (EEG) based motor imagery Brain-Computer Interface (MI-BCI) has been widely applied in constructing a pathway between human brains and external machines. However, decoding of MI-EEG signals is challenging as EEG is severely affected by non-stationarity and high variability in signal patterns. In this work, to fully extract the spatial and temporal information of EEG for MI decoding, we designed an end-to-end compact deep convolutional neural network model, combining EEGNet and a temporal convolutional network. The proposed model requires few data pre-processing and a small number of trainable parameters, achieving significant performance improvement on MI-EEG decoding tasks. Experimental results under the subject-dependent manner show that our method achieves 56.73%, 73.9%, and 75.4% classification accuracy on the 2020 BCIC, Same Limb, and OpenBMI datasets, respectively, which outperforms the state-of-the-art (STOA) convolutional neural network (CNN). Under the subject-independent with the 2020 BCIC dataset, the proposed approach achieves an accuracy improvement of 2.8% compared to the STOA CNN. The code of the proposed method is available at https://github.com/ingod/DDCLS-MI-EEG-BCI.