Although various kinds of motor imageries have been used for BCI applications, imaginary lower limb stepping movement has not been studied yet. The purpose of this study is to investigate the possibilities of using electroencephalography (EEG) signal to classify imaginary lower limb stepping movements and to design a robust motor imagery classifier based on support vector machine (SVM). A cue-based experimental paradigm is designed to record nine-channel EEG associated with imaginary left leg stepping (L-stepping) and right leg stepping (R-stepping) movements from eight healthy subjects. Features including band powers (BPs), common spatial pattern (CSP), and a filter-bank CSP (FB-CSP) were extracted from the recorded EEG. Fuzzy SVM (FSVM) is introduced to this study to classify L-stepping and R-stepping imageries. We propose a novel kernel-induced membership function to address the issue of data relative importance assignment. The FSVM with the membership function suggested in the original work of FSVM (Type-I FSVM) and the FSVM with the one we proposed (Type-II FSVM) is compared. Results indicated that the classification accuracies based on BP features are near the chance level (~50 %). Both alpha-band CSP (71.25 %) and FB-CSP (75.63 %) gave acceptable results as a simple k-NN classifier is performed. Results show that both types of FSVM performed better than the conventional SVM. Also, Type-II FSVM outperforms Type-I FSVM, especially when the alpha-CSP feature is employed, where the improvement in error reduction rate is over 15 %. The highest average L-stepping versus R-stepping classification accuracy over the eight subjects is achieved (86.25 % in single-trial analysis) by FB-CSP and FSVM-II. The high classification result suggests the feasibility of using lower limb stepping imagery to develop a BCI that can control devices or might be able to serve as a neurofeedback tool for users who need lower limb stepping imagery training for gait function improvement.
ASJC Scopus subject areas