A novel dual-step transfer framework based on domain selection and feature alignment for motor imagery decoding

• Published in: Cognitive neurodynamics Vol. 18; no. 6; pp. 3549 - 3563
• Main Authors: Bai, Guanglian, Jin, Jing, Xu, Ren, Wang, Xingyu, Cichocki, Andrzej
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.12.2024; Springer Nature B.V
• Subjects: Accuracy; Adaptation; Algorithms; Alignment; Artificial Intelligence; Biochemistry; Biomedical and Life Sciences; Biomedicine; Calibration; Classification; Cognitive Psychology; Computer Science; Datasets; Electroencephalography; Filter banks; Human-computer interface; Imagery; Machine learning; Mental task performance; Neurosciences; Regularization; Research Article; Strategy; Support vector machines; Transfer learning; Brain-computer interface; Motor imagery; Transfer learning; Domain selection; Feature alignment
• ISSN: 1871-4080; 1871-4099
• DOI: 10.1007/s11571-023-10053-1

In brain-computer interfaces (BCIs) based on motor imagery (MI), reducing calibration time is gradually becoming an urgent issue in practical applications. Recently, transfer learning (TL) has demonstrated its effectiveness in reducing calibration time in MI-BCI. However, the different data distribution of subjects greatly affects the application effect of TL in MI-BCI. Therefore, this paper combines data alignment, source domain selection, and feature alignment into the MI-TL. We propose a novel dual-step transfer framework based on source domain selection and feature alignment. First, the source and target domains are aligned using a pre-calibration strategy (PS), and then a sequential reverse selection method is proposed to match the optimal source domain for each target domain with the designed dual model selection strategy. We use filter bank regularization common space pattern (FBRCSP) to obtain more features and introduce manifold embedded distribution alignment (MEDA) to correct the prediction results of the support vector machine (SVM). The experimental results on two competition public datasets (BCI competition IV Dataset 1 and Dataset 2a) and our dataset show that the average classification accuracy of the proposed framework is higher than the baseline method (no domain selection and no feature alignment), which reaches 84.12%, 79.91%, and 78.45%, respectively. And the computational cost is reduced by half compared with the baseline method.