Data augmentation for cross-subject EEG features using Siamese neural network

• Published in: Biomedical signal processing and control Vol. 75; p. 103614
• Main Authors: Fu, Rongrong, Wang, Yaodong, Jia, Chengcheng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.05.2022
• Subjects: Complex motor recognition; Data augmentation; Siamese neural network; Similarity measurement; Transfer learning; Data augmentation; Siamese neural network; Similarity measurement; Transfer learning; Complex motor recognition
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2022.103614

•A novel data augmentation method via transfer learning and cross subject Siamese neural network was proposed.•A new experimental paradigm was applied to help researchers evaluate the performance of data augmentation methods.•The experimental results show that the proposed method obtains better performance on various machine learning methods. Electroencephalography (EEG) motor intention recognition has been extensively used in robot control, brain rehabilitation and other health care fields. Recently, some algorithms have been proposed based on generative adversarial neural network (GAN) to enhance EEG signal, and have achieved high recognition performance. However, these methods utilize the convolutional kernel method of the GAN, while the optimal convolutional scale of CNN varies from subject to subject. This may lead to the data generated by GAN to lack authenticity and produce data that does not match the ideal situation. Particularly, the performance of data augmentation degrades when the original calibrated EEG is insufficient. To address these issues, we proposed a novel cross-subject Siamese Neural Network (SNN) approach to enhance EEG feature data. Specifically, we used our proposed SNN to construct highly similar extended EEG features of different subjects and successfully improved the performance of motor intention recognition. Then, we design an accurate boundary avoidance task to evaluate the effectiveness of the proposed method. Compared with the traditional experimental paradigm, the coding process of this experiment is more complex, which makes the results more reliable when using the SNN. The extended EEG features display significantly better performance than any other common classifiers in the case of small data size, and it demonstrates that this proposed method can effectively address these issues of existing EEG motor intention recognition methods based on data augmentation and improve the classification performance.