A Robust Multi-Channel EEG Signals Preprocessing Method for Enhanced Upper Extremity Motor Imagery Decoding

Brain computer interface (BCI) based on non-invasive electroencephalography (EEG) signals has become a promising and alternative method for electromyography (EMG) signal in the rehabilitation process of amputees with high level of amputation. This is because their residual muscles cannot provide sufficient myoelectric signals to accurately recognize limb movement intentions. One of the major challenges of BCI based EEG methods is the inevitable artifacts contained in multi-channels EEG recordings that would affect accurate characterization and decoding of limb movement intents from brain signals. Previous studies have applied different methods based on regression, blind source separation (BSS) and filtering techniques, though with limited efficacy in fully isolating artifacts from the entire EEG signals. Also, some of the existing methods are targeted towards the removal of one or two particular types of artifacts. In this study, a linear algorithm based on generalized eigenvalue decomposition and multi-channel Wiener filter (GEVD-MWF) was proposed to simultaneously remove different type of artifacts inherent in EEG signals, recorded from four transhumeral amputee subjects. Experimental results showed that the proposed method outperformed the commonly used approach which achieved an improved average classification accuracy of 22.03% across all the classes of motor imagery (MI) tasks and subjects. For real-time applications, a modified channel selection algorithm-based on sequential forward floating selection (mSFFS) was adopted to remove redundant channels and an average accuracy of 96.87% was achieved with 10 electrode channels located at the pre-motor cortex, primary motor cortex and somatosensory cortex of the brain. The overall performance of the proposed method suggested that multiclass MI tasks can be reliably and accurately characterized using artifact-free EEG signals from a few number of channels. Thus, the proposed methods can be a potential control method for rehabilitation devices for individuals with high level amputation.