Cross-subject EEG feature matrix classification method and its application in brain-computer interface

• Published in: Multimedia tools and applications Vol. 83; no. 33; pp. 79627 - 79646
• Main Author: Luo, Tian-jian
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.10.2024; Springer Nature B.V
• Subjects: Ablation; Classification; Cognitive tasks; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Electroencephalography; Feature extraction; Human-computer interface; Machine learning; Matrices (mathematics); Multimedia; Multimedia Information Systems; Paradigms; Pattern recognition; Signal classification; Special Purpose and Application-Based Systems; State-of-the-art reviews; Time series; Track 3: Biometrics and HCI; Brain-computer interface; Sparse support matrix machine; Covariance alignment; Cross-subject; EEG feature matrix
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-18648-4

EEG signals are widely utilized in brain-computer interface (BCI) applications. However, the non-linear and non-stationary nature of EEG signals poses a challenge when dealing with variations across subjects and sessions, leading to the covariate shift problem in recognition tasks. Conventional approaches often extract vector-form features for classifying EEG signals on a per-subject and per-session basis, resulting in the loss of discriminative features and decreased recognition performance. To address this issue, this paper presents a novel cross-subject EEG feature matrix classification method that leverages the feature matrix encompassing all subjects to recognize EEG signals. The proposed method begins by aligning the EEG covariances of each subject to an identity distribution, followed by extracting a feature matrix from the aligned EEG signals. To recognize EEG signals associated with specific mental tasks, a sparse support matrix machine is employed to select discriminative features from the feature matrix and perform classification based on these selected features. To evaluate the proposed method, two publicly available benchmark datasets containing motor imagery and event-related potentials were used in experiments. Comparative analyses with state-of-the-art methods demonstrated improved recognition performance with the proposed method. Furthermore, additional ablation studies were conducted to explore the potential application prospects of the proposed method in BCI researches.