Unsupervised multi-source variational domain adaptation for inter-subject SSVEP-based BCIs

• Published in: Expert systems with applications Vol. 238; p. 122155
• Main Authors: Zhang, Shubin, An, Dong, Liu, Jincun, Wei, Yaoguang, Sun, Fuchun
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.03.2024
• Subjects: Deep variational inference; Inter-subject brain-computer interface; Multi-source domain adaptation; Steady-state visual evoked potential; Steady-state visual evoked potential; Deep variational inference; Multi-source domain adaptation; Inter-subject brain-computer interface
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2023.122155

Steady-State Visual Evoked Potential-based Brain-Computer Interfaces (SSVEP-based BCIs) are widely used to decode electroencephalography (EEG) data, often employing subject-specific models. However, these methods demand a substantial volume of labeled data, which tends to induce visual fatigue in subjects. To address this challenge, an efficient approach is to transfer known subject knowledge through inter-subject methods using unlabeled data. However, the previous state-of-the-art unsupervised inter-subject model is limited by the feature extraction capability of spatial filter methods. To overcome this limitation, we propose a novel Multi-source Variational Domain Adaptation (MVDA) method, which incorporates a Filter-bank Dynamic Graph Convolutional Network (FBDGCN). The FBDGCN aims to extract distinctive graph features under different harmonics from offline labeled source and unlabeled target EEG data. Subsequently, the source and target graph features are jointly trained using MVDA for domain adaptation. With proposed varitional inference mechanism, MVDA can constrain the distribution manifold and minimize the domain discrepancy. Since the performance of knowledge transfer heavily relies on the similarity between source and target domains, we also propose an unsupervised subject correlation analysis method to select the source subjects from the dataset. Our experimental results demonstrate that the MVDA method yields a higher information transfer rate (ITR) (189.87 ± 24.98 bits/min) and mean accuracy (92.11 ± 2.85%). The proposed model can be calibrated only by unlabeled target subject data to obtain a higher recognition accuracy, which is more applicable to real-world scenarios. •A novel Multi-source Domain adaptation method for inter-subject SSVEP BCIs.•Filter-bank Dynamic Graph Convolutional Network.•Deep variational inference mechanism for graph feature space constraint.•An efficient unsupervised subjects correlation analysis method.