Subject-Independent Motor Imagery EEG Classification Based on Graph Convolutional Network

• Published in: Pattern Recognition Vol. 13189; pp. 268 - 281
• Main Authors: Lee, Juho, Choi, Jin Woo, Jo, Sungho
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Brain-computer interfaces (BCIs); Deep learning (DL); Electroencephalography (EEG); Graph convolutional network (GCN); Motor imagery (MI); Subject-independent
• ISBN: 9783031024436; 3031024435
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-02444-3_20

Electroencephalogram (EEG) motor imagery (MI) has attracted much attention in brain-computer interfaces (BCIs) as it directly encodes human intentions. However, the variability of EEG-based brain signals between individuals requires current BCI systems to undergo calibration procedures before its usage. In this paper, we propose a model that targets minimizing such procedures by improving inter-subject classification performance. The purpose of our proposed method is to extract features using previously studied convolution-based deep learning structures while utilizing a graph structure to analyze inter-subject relationships with multiple subjects. By utilizing not only features but also the relationship between subject-specific features, it becomes possible to make predictions focusing on subjects with high similarity. Therefore, even new users not seen during the training process are predicted relatively efficiently. To validate our method, we evaluated our model with the public dataset BCI Competition IV IIa. The results in our study suggest that our proposed method improved the cross-subject classification accuracy by combining it with the previous deep learning model and induced a balanced prediction for the classes. Our study has shown the potential to develop MI-based BCI applications that do not require user calibration by training the model with pre-existing datasets.