SKSP: Selected Kernel Spatial Pattern Analysis in Euclidean and Hilbert Space for Decoding Motor Imagery

Brain-computer Interface (BCI) has promising prospects as the most active research direction in the fields of brain science and artificial intelligence. The Common Spatial Pattern (CSP) algorithm has attracted extensive attention and has been demonstrated to be effective for Motor Imagery (MI) tasks...

Full description

Saved in:
Bibliographic Details
Published inSmart World Congress (SWC), IEEE pp. 1249 - 1256
Main Authors Zhibin, Zhang, Niu, Jianwei, Ren, Lu, Ouyang, Zhenchao, Mo, Shasha
Format Conference Proceeding
LanguageEnglish
Published IEEE 02.12.2024
Subjects
Accuracy
Common Spatial Pattern
Electroencephalography
Hilbert space
Kernel
Motor Imagery
Motors
Nonlinear
Pattern analysis
Pattern recognition
Time-domain analysis
Time-frequency analysis
Wavelet transforms
Online AccessGet full text

Cover

More Information
Summary:Brain-computer Interface (BCI) has promising prospects as the most active research direction in the fields of brain science and artificial intelligence. The Common Spatial Pattern (CSP) algorithm has attracted extensive attention and has been demonstrated to be effective for Motor Imagery (MI) tasks. However, CSP is limited to capturing complex nonlinear characteristics under its linear assumption. Many variants of CSP cannot deal with multiple relationships simultaneously. This paper proposes a Selected Kernel Spatial Pattern (SKSP) algorithm for human behavior recognition. Electroencephalogram (EEG) signals are projected onto a high-dimensional Reproducing Kernel Hilbert Space (RKHS) by the selected kernel function, where the nonlinear relationship is transformed into a linear one. This approach exploits wavelet transform and CSP patterns to compute the energy spectrum and differential entropy of data in RKHS to represent EEG patterns of brain activities. In addition, a multi-view module was proposed to strengthen pattern representations, which employs mutual information maximization to select the optimal sliding window in both frequency and time domains. The experiments are conducted on the BCI Competition IV-2a dataset. The four classes' classification accuracy reaches 92% in subject seven. The proposed method outperforms the best state-of-the-art classification method by 4% in average subject accuracy.
ISSN:2993-396X
DOI:10.1109/SWC62898.2024.00196