Manifold Embedded Domain Adaptation for Motor Imagery EEG Decoding

• Published in: IAENG international journal of computer science Vol. 51; no. 8; p. 985
• Main Authors: Jiang, Qin, Zhang, Yi, Wang, Wei, Huang, Qian
• Format: Journal Article
• Language: English
• Published: Hong Kong International Association of Engineers 01.08.2024
• Subjects: Adaptation; Algorithms; Calibration; Datasets; Decoding; Embedding; Human-computer interface; Imagery; Manifolds (mathematics); Riemann manifold; Spatial filtering
• ISSN: 1819-656X; 1819-9224

Training a robust classifier for a motor imagery-based brain-computer interface (MI-BCI) system typically requires a substantial amount of time to collect calibration data, which can be a burdensome task for participants. To enhance the classifier performance while reducing the effort during the training phase, this paper proposes a domain adaptation algorithm based on manifold embedding (eSPDA) by combining the domain adaptation approach with a dimensionality reduction framework derived from the common spatial patterns (CSP). Specifically, the CSP spatial filtering theory is construed as the dimensionality reduction in Riemannian manifold with maximum intra-class variance or maximum inter-class distance. Based on the principle of maximum inter-class distance, the labeled source data is embedded into a more discriminative submanifold, where the principal characteristics of the unlabeled target subject are preserved by the rule of maximum intra-class variance. Meanwhile, the joint distribution alignment is integrated into the framework to minimize the distribution divergences across subjects. The results on two datasets demonstrate that eSPDA outperforms several state-of-the-art domain adaptation methods, with the average accuracies 70.35% and 80.67% on BCI Competition IV dataset IIa and BCI Competition IV dataset IVa, respectively. This research indicates that eSPDA has potentials to reduce the labeling effort, resulting in calibration time and effort savings.