Kernel-Based Embedded Feature Selection for Motor Imagery Based BCI

• Published in: Iranian Conference on Electrical Engineering pp. 144 - 148
• Main Author: Kamandar, Mehdi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09.05.2023
• Subjects: Brain-computer interface (BCI); Electroencephalogram (EEG); Electroencephalography; Embedded feature selection; Feature extraction; Kernel-target alignment (KTA); l1 regularization; minibatch proximal gradient ascent; Motor imagery (Ml); Redundancy; Scalp; Support vector machine (SVM); Support vector machine classification; Support vector machines; Training
• ISSN: 2642-9527
• DOI: 10.1109/ICEE59167.2023.10334784

Brain-computer interface (BCI) based on motor imagery (MI) classification using scalp recorded multi-channel EEG signal play a major role in the control of artificial limbs and machines by people with severe disabilities. The most popular features are band power of EEG signals in frequency sub-bands in a relatively wide frequency band (e. g., 8-30 Hz) in different time windows of imagination period. The effectiveness of these features is highly dependent on the sub-bands and time windows adopted, since the optimal sub-bands and time windows is generally subject-specific. The nonlinear support vector machine (SVM) with Gaussian kernel is an excellent classifier for MI classification. In this paper, the SVM with an anisotropic Gaussian kernel with a scaling coefficient for each feature is used, instead of one scale factor for all features. The scaling coefficients are tuned by maximizing the kernel-target alignment criterion with l_{l} regularization. Some of the scaling coefficients will be zero after maximization due to l_{l} regularization, that is equivalent to removing corresponding irrelevant features from classification processes. Therefore, an embedded feature selection is also done to remove the destructive effect of irrelevant and redundant features. The average accuracy of the SVM classifier with anisotropic Gaussian kernel for four subjects in BCI IV-1 dataset is about 10% more than the SVM with isotropic Gaussian kernel.