Classification of kinetics of movement for lower limb using covariate shift method for brain computer interface

Detecting movement intentions from Electroencephalography (EEG) signals and extracting intended kinetic information such as force and speed may have implications for rehabilitation with assistive technologies by casually linking afferent feedback from the assistive device with the cortical generated movement potentials. However, extraction and classification of kinetics from the `movement intention´ (before task onset) on a single-trial basis have only been performed with limited performance due to low signal-to-noise ratio and large trial-to-trial variability. The aim of this study was to investigate a covariate shift method to address the basic challenge of non-stationarity (changes from session to session and trial-to-trial variability) for decoding different levels of speed and force. We tested this method using cross-validation procedures and a linear support vector machine to classify temporal features associated with two levels of force and speed in 9 subjects. The classification accuracy obtained across different class pairs across subjects was 73.1 ±6.8 % and 70.0± 3.6 % with and without the covariate shift method, respectively. The classification accuracy was significantly higher (p <; 0.03) using the covariate shift method.