A systematic review of state of the art integration of neural biosignals during robotic assisted gait training (RAGT) after stroke has been completed and a checklist developed to guide more uniform data collection. Open access publication:
https://pure.ulster.ac.uk/ws/portalfiles/portal/79628365/DCoyle_OA_Frontiers.pdf(öffnet in neuem Fenster)Early laboratory work examined EEG data (N=10 healthy individuals) during stepping for 3D motion trajectory prediction. Accuracy predicting movement in the forward plane was established; testing of the GTec 32 channel GNautilus wireless device observed low artefact.
Clinical work in acute stroke exoskeleton training (N=10 individuals). Differences in EMG signals in 4 lower-limb muscle groups during different RAGT modes were evident; adaptive assistance showed better muscle activity in comparison to a passive mode.
Qualitative data (N=8 ) explored experiences of RAGT in acute/subacute stroke where RAGT was considered acceptable; participants noted they were anxious starting RAGT but there was nothing to fear and in many cases allowed them to walk when they could not otherwise. Negative features included the device weight and the time to measure and first set up.
Data (N=3 healthy; N=3 stroke) co-registered 32 channel EEG with EMG during RAGT walking in different modes and during overground walking without an exoskeleton. Preliminary analysis identifies that classifying stepping by EEG is better during robotic walking that overground walking.
ERSP and Cortico-Muscular Coherence (CMC) neurophysiological biomarkers showed modulation in healthy individuals during RAGT represented in high θ-desynchronization, undetected in stroke during RAGT. Reduced swing time observed during exoskeleton walking in stroke was not correlated with an increment of θ-desynchronization and/or β-CMC.
Co-registered EMG and EEG collected during RAGT (transparent, adaptive and full assistive modes) and non-assisted gait in healthy (N=10) and neurological patients (N=3)identified that irrespective of the exoskeleton mode, stronger modulation of central mid-line mu (8–13 Hz) and low-beta (14–20 Hz) rhythms compare with free overground walking, with no significant differences by assistance levels during RAGT observed. Implementation of gait decoders, based on deep neural networks trained on the EEG data during the different walking conditions, demonstrated all decoders achieved an average accuracy of 84% in predicting swing and stance phases. In addition, the decoder trained on transparent mode exoskeleton data predicted gait phases during adaptive and full modes with 78% accuracy, while the free overground walking decoder failed to decode gait during exoskeleton walking (accuracy 59%). This work validated the new 64 channel GNautilus wireless device.
A closed loop BCI robotic gait device was explored with partners EKSObionics, centred on a data portal allowing external commands to the operating system, in addition to a new hardware and software integration platform using Matlab Simulink for real time EEG processing (GTec neurotechnology), translation of neural data into exoskeleton actuate step commands through sensorimotor modulation alone. Trained in a healthy individual, a BCI classification system was deployed in real time RAGT.
Dissemination activities have included: three publications, 2 preprint publications, and one paper under peer review at project completion; conference presentations, outreach activities, show case events and an online research symposium.
