Relevance of Common Spatial Patterns Ranked by Kernel PCA in Motor Imagery Classification

Abstract

Motor Imagery handles the brain activity patterns of motor action without explicit movements. For extracting the discriminating features, Common Spatial Patterns are the most widely used algorithm that is very sensitive to artifacts and prone to overfitting. Here, we develop a metric to assess the relevance of Common Spatial Patterns using a mapping through Kernel Principal Component Analysis with the benefit of improved interpretation that allows evaluating the zones, which contribute the most to the motor imagery classification accuracy. Validation is carried out on a real-world database, appraising two labels of Motor Imagery activity. From the obtained results, we prove that the developed approach allows the performance enhancement, at the time, the relevant set decreases the number of channels to feed the classifier, and thus reducing the computational cost.