Sub-Band Cascaded CSP-based Deep Transfer Learning for Cross-Subject Lower Limb Motor Imagery Classification

Abstract

Lower limb motor imagery (MI) classification is a challenging research topic in brain-computer interface (BCI) due to excessively close physiological representation of left and right lower limb movements in the human brain. Moreover, MI signals have severely subject-specific characteristics. The classification schemes designed for a specific subject in previous studies could not meet the requirements of cross-subject classification in a generic BCI system. Therefore, this study aimed to establish a cross-subject lower limb MI classification scheme. Three novel sub-band cascaded common spatial pattern (SBCCSP) algorithms were proposed to extract representative features with low redundancy. The validations had been conducted based on the lower limb stepping-based MI signals collected from subjects performing MI tasks in experiments. The proposed schemes with three SBCCSP algorithms have been validated with better accuracy and running time performances than other common spatial pattern (CSP) variants with the best average accuracy of 98.78%. This study provides the first investigation of a cross-subject MI classification scheme based on experimental stepping-based MI signals. The proposed scheme will make an essential contribution to developing generic BCI systems for lower limb auxiliary and rehabilitation applications.