Objective: The aim of this study is to propose a recognition system of pedaling motor imagery for lower-limb rehabilitation, which uses unsupervised methods to improve the feature extraction, and consequently the class discrimination of EEG patterns.
Approach: After applying a spectrogram based on short-time Fourier transform (SSTFT), both sparseness constraints and total power are used on the time-frequency representation to automatically locate the subject-specific bands that pack the highest power during pedaling motor imagery. The output frequency bands are employed in the recognition system to automatically adjust the cut-off frequency of a low-pass filter (Butterworth, 2nd order). Riemannian geometry is also used to extract spatial features, which are further analyzed through a fast version of neighborhood component analysis to increase the class separability.
Main results: For ten healthy subjects, our recognition system based on subject-specific bands achieved mean accuracy of [Formula: see text] and mean Kappa of [Formula: see text].
Significance: Our approach can be used to obtain a low-cost robotic rehabilitation system based on motorized pedal, as pedaling exercises have shown great potential for improving the muscular performance of post-stroke survivors.