Characterization of error-related potentials during the command of a lower-limb exoskeleton based on deep learning

Abstract

Background Brain-Machine Interfaces (BMI) based on motor imagery (MI) are promising assistive neurotechnology tools for gait rehabilitation that allow users to control exoskeletons by imagining motor actions. Literature has proven the influence of BMIs over neuroplasticity mechanisms. However, the accuracy of MI-BMIs is often limited by the weak brain signals associated with lower-limb movements. To enhance system reliability and safety, Error-Related Potentials (ErrP), which are exogenous potentials evoked by erroneous system actions, can be integrated to correct commands. This study characterizes and detects ErrP during the use of a lower-limb exoskeleton, making use of a deep learning approach to improve accuracy and robustness over traditional classifiers.

Methods ErrP detection is performed using the EEG-Inception neural network, a convolutional deep learning model, and applying data augmentation techniques to the imbalanced dataset. The methodology is tested first for the characterization of ErrP during the start of gait with static data and, after confirming its improvement, regarding previous developments, it is also applied to motion data during the stop of the exoskeleton. With this objective, an experimental protocol is designed to evoke ErrP and NoErrP during motion, using tactile stimuli. ErrP is elicited when the exoskeleton stops erroneously in a gait region, while NoErrP is generated when it stops correctly in a stop region.

Results The proposed approach achieves a True Positive Rate (TPR) of approximately 95% and a False Positive Rate (FPR) below 20% in both static and motion conditions, significantly outperforming traditional ensemble classifiers. In terms of MI-BMI performance, these results indicate that most erroneous commands are successfully canceled, while only a small number of correct commands are wrongly canceled. In addition, statistical analysis revealed no significant differences between the detection of ErrP in static and motion scenarios, nor between sessions or subjects just in static. However, significant differences are observed between subjects in motion and also the outcomes of ErrP and NoErrP classes in both scenarios.

Conclusion The EEG-Inception neural network provides a robust and accurate method for ErrP detection. Future research will focus on integrating ErrP detection with MI classifiers and validating the system with SCI patients for improved gait rehabilitation therapies.