Nonlinear difference subspace method of motor imagery EEG classification in brain-computer interface

• Published in: Digital signal processing 2024-12, Vol.155, p.104720, Article 104720
• Main Authors: Reddy, C Sivananda, Reddy, M Ramasubba
• Format: Article
• Language: English
• Subjects: Brain-computer interface (BCI); Difference subspace method (DSM); Isometric feature mapping (Isomap); Kernel principal component analysis (KPCA); Motor imagery (MI)
• ISSN: 1051-2004
• DOI: 10.1016/j.dsp.2024.104720

Noninvasive Motor Imagery (MI) electroencephalography (EEG) based brain-computer interface (BCI) systems are in high demand in both the medical and consumer sectors. However, the low signal-to-noise ratio of EEG and its nonlinear dynamics present challenges in processing. The effectiveness of information retrieval is directly related to the complexity of the algorithms, making BCI impractical for real-time applications. To enhance performance, this study introduced classification frameworks using nonlinear subspace and difference subspace methods, employing manifold learning techniques such as Kernel principal component analysis (KPCA) and Isometric feature mapping (Isomap). The optimal dimensions for the subspaces were selected using overlapping criteria in conjunction with the breadth-first-search (BFS) and depth-first-search (DFS) graph search methods. The proposed methods were evaluated on the BCI competition III and IV datasets using metrics including classification accuracy, Cohen's kappa coefficient, and F1-score. By exploiting the nonlinear characteristics of EEG and employing the geodesic distance between all pairs of trials, the modified Isomap called supervised discriminant Isomap projection (SD-ISoP) outperformed the KPCA-based methods and other recently proposed state-of-the-art methods. The results demonstrssate that the SD-ISoP-based difference subspace method achieved an average classification accuracy of 86.25% and 85.44% for BCI competition III and IV datasets, respectively.