Classification of Stroke Patients’ Motor Imagery EEG with Autoencoders in BCI-FES Rehabilitation Training System

Abstract

Motor imagery based Brain Computer Interface (BCI) system is a promising strategy for the rehabilitation of stroke patients. Common Spatial Pattern (CSP) is frequently used in feature extraction of motor imagery EEG signals and its performance depends heavily on the choice of frequency component. Moreover, EEG of stroke patients, which is full of noise, makes it hard for traditional CSP to extract discriminative patterns for classification. In order to deal with the subject-specific band selection, in this paper, we adopt denoising autoencoders and contractive autoencoders to extract and compose robust features from CSP features filtered in multiple frequency bands. We compare our method with traditional methods on data collected from two months clinical rehabilitation. The results not only demonstrate its superior recognition performance but also evidence the effectiveness of our BCI-FES rehabilitation training system.