Variability in Grasp Type Distinction for Myoelectric Prosthesis Control Using a Non-Invasive Brain-Machine Interface

TL;DR

This study explores EEG-based classification of grasp movements for prosthesis control, revealing high accuracy in movement detection, variability in EEG patterns, and the impact of electrode reduction, informing wearable brain-machine interface development.

Contribution

It introduces a comprehensive analysis of EEG decoding for grasp movements across different conditions and electrode setups, highlighting variability and robustness in classification performance.

Findings

100% accuracy in movement vs. no movement detection

70-90% accuracy in classifying different grasp movements

Electrode reduction causes only a 2% accuracy decrease

Abstract

Decoding multiple movements from the same limb using electroencephalographic (EEG) activity is a key challenge with applications for controlling prostheses in upper-limb amputees. This study investigates the classification of four hand movements to control a modified Myobock prosthesis via EEG signals. We report results from three EEG recording sessions involving four amputees and twenty able-bodied subjects performing four grasp movements under three conditions: Motor Execution (ME), Motor Imagery (MI), and Motor Observation (MO). EEG preprocessing was followed by feature extraction using Common Spatial Patterns (CSP), Wavelet Decomposition (WD), and Riemannian Geometry. Various classification algorithms were applied to decode EEG signals, and a metric assessed pattern separability. We evaluated system performance across different electrode combinations and compared it to the original setup. Our results show that distinguishing movement from no movement achieved 100% accuracy, while classification between movements reached 70-90%. No significant differences were found between recording conditions in classification performance. Able-bodied participants outperformed amputees, but there were no significant differences in Motor Imagery. Performance did not improve across the sessions, and there was considerable variability in EEG pattern distinction. Reducing the number of electrodes by half led to only a 2% average accuracy drop. These results provide insights into developing wearable brain-machine interfaces, particularly for electrode optimization and training in grasp movement classification.