From: Brain–computer interface robotics for hand rehabilitation after stroke: a systematic review
Authors | Participants | Study design | Task design | BCI-Hand robot | Main outcomes |
---|---|---|---|---|---|
Studies involving stroke patients | |||||
Ang et al. 2014 [75] | N = 27 (7F:14 M) Moderate to severe impairment of UE function Mean age: 54.2y Mean stroke duration: 385.1 days | 3-armed RCT of motor function with MI-BCI-device as intervention Control groups: device only (Haptic Knob), SAT | Photo manipulation: hand opening and closing, pronation and supination Cue: visual (photo) Feedback: visual (photo) and kinaesthetic | EEG: 27 channels to classify ERD/ERS and coupled with EMG to confirm MI Device: Haptic Knob, 2-DOF for hand grasping and knob manipulation Actuation: DC brushed motors with linear belt drive Control: trigger | Clinical outcome measure: FMMA Distal, improvement in weeks 3, 6, 12, 24 BCI-device group = 2.5 ± 2.4, 3.3 ± 2.3, 3.2 ± 2.7, 4.2 ± 3.1 Device only group = 1.6 ± 2.5, 2.9 ± 3.0, 2.5 ± 2.6, 2.5 ± 3.0 SAT group = 0.4 ± 1.1, 1.9 ± 1.9, 1.0 ± 1.3, 0.3 ± 2.1 |
Barsotti et al. 2015 [76] | N = 3 (1F:2 M) Chronic stroke survivors with right arm hemiparesis Mean age: 62 ± 12y | Probing MI classification by BCI training, time–frequency analysis and robot trajectories Uncontrolled | Reaching-grasping-releasing Cue: visual Feedback: kinaesthetic Minimum time required to perform MI = 2 s | EEG: 13 channels to classify ERD Device: BRAVO 2-DOF hand orthosis attached to full UE exoskeleton Actuation: DC motors with rigid links Control: trigger | Mean classification accuracy during BCI training = 82.51 ± 2.04% Average delay from visual cue to robot initiation = 3.45 ± 1.6 s Average delay due to patient’s ability to start MI = 1.45 s |
Bundy et al. 2017 [77] | N = 10 Chronic hemiparetic stroke with moderate to severe UE hemiparesis Mean age: 58.6 ± 10.3y | Motor function evaluation before and after intervention by MI-BCI from unaffected hemisphere Uncontrolled | Opening of affected hand Cue: visual Feedback: visual and kinaesthetic | EEG: 8 channels to classify ERD Device: 3-pinch grip, 1-DOF hand exoskeleton Control: continuous depending on spectral power | Clinical outcome measure: ARAT Score, improvement from baseline to completion (12 weeks) Mean ± SD = 6.20 ± 3.81 Note: 5.7 ARAT Score is the minimal clinically important difference in chronic stroke survivors |
Carino-Escobar et al. 2019 [85] | N = 9 (4F:5 M) Subacute ischaemic stroke Mean age: 59.9 ± 2.8y Mean stroke duration: 158(± 74)-185(± 73) days | Determine longitudinal ERD/ERS patters and functional recovery with BCI-robot Uncontrolled | Extension-flexion of hand fingers Cue: visual (Graz MI) Feedback: visual and kinaesthetic | EEG: 11 channels to classify ERD/ERS Device: hand finger orthosis Actuation: DC motor with screw system for linear displacement, flexible links Control: trigger | FMA-UE: N = 3 reported equal or higher than 3 score gains, N = 3 no score gains, Mean longitudinal ERD/ERS: beta bands have higher association with time since stroke onset than alpha, and strong association with UL motor recovery |
Chowdhury et al. 2018-b [78] | N = 20 10 healthy and 10 hemiplegic stroke patients Mean age (healthy, stroke): 41 ± 9.21y, 47.5 ± 14.23y | Probe non-adaptive classifier (NAC) vs. Covariate Shift adaptive classifier (CSAC) of MI in EEG Control group: healthy participants | Extension-flexion of hand fingers Cue: visual Feedback: visual and kinaesthetic | EEG: 12 channels with EMG to classify ERD/ERS Device: EMOHEX 3-finger, 3-DOF each, exoskeleton (thumb, index, middle) Actuation: servomotors with rigid links Control: trigger | Mean classification accuracies during BCI training: Healthy group: calibration = 78.50 ± 9.01%, NAC = 75.25 ± 5.46%, CSAC = 81.50 ± 4.89% Patient group: calibration = 79.63 ± 13.11%, NAC = 70.25 ± 3.43%, CSAC = 75.75 ± 3.92% |
Chowdhury et al., 2018-c [79] | N = 4 (2F:2 M) Chronic hemiplegic stroke patients, right-handed, left hand impaired Mean age: 44.75 ± 15.69y Mean stroke duration: 7 ± 1.15mo | Motor function evaluation by using active physical practice followed by MI-BCI-controlled device intervention Uncontrolled | Extension-flexion of hand fingers Cue: visual Feedback: visual and kinaesthetic | EEG: 12 channels with force sensors to classify ERD/ERS Device: EMOHEX 3-finger, 3-DOF each, exoskeleton (thumb, index, middle) Actuation: servomotors with rigid links Control: trigger | Classification accuracies of 4 participants: P01 = 81.45 ± 8.12%, P02 = 70.21 ± 4.43%, P03 = 76.88 ± 4.49%, P04 = 74.55 ± 4.35% Clinical outcome measures: GS and ARAT Scores, improvement from baseline to completion (6 weeks) GS scores: group mean difference = + 6.38 kg, p = 0.06 ARAT scores: group mean difference = + 5.66, p < 0.05 |
Frolov et al., 2017 [80] | N = 74 (26F:48 M) BCI 55: Control 19 Subacute or chronic stroke with mild to hemiplegic hand paresis, right-handed | Multi-centre RCT of MI-BCI-controlled hand exoskeleton Control group: SHAM | 3 Tasks: (1) motor relaxation, (2) imagery of left-hand opening, (3) imagery of right-hand opening Cue: visual Feedback: visual and kinaesthetic | EEG: 30 channels to classify the three mental tasks by Bayesian classifier based on covariance matrices Device: hand exoskeleton by Neurobotics, Russia Actuation: pneumatic motors with spring flexors Control: trigger | Mean classification accuracy during BCI training = 40.6% Clinical outcome measures: FMMA Distal and ARAT Scores, improvement in 10 days of training FMMA Distal = 2.0, p < 0.01 (BCI) and 1.0, p = 0.046 (control) ARAT Grasp = 3.0, p < 0.01 (BCI) and 1.0, p = 0.0394 (control) ARAT Grip = 1.0, p < 0.01 (BCI) and 1.0, p = 0.045 (control) ARAT Pinch = 1.0, p < 0.01 (BCI) and 0.0, p = 0.675 (control) |
Norman et al., 2018 [82] | N = 8 (All male) Chronic cortical and subcortical single haemorrhagic or ischaemic stroke (at least 6 months) Mean age: 59.5 ± 11.8y | Implementation of sensorimotor rhythm (SMR) control on robot-assistive movement Uncontrolled | Extension of hand finger Cue: visual Feedback: visual and kinaesthetic | EEG: 16 channels mapping SMR changes Device: FINGER robot Actuation: Linear servo-tube actuator with rigid links Control: Visual—continuous (colour change respective to SMR), Robot—trigger | Mean classification accuracies: 8 participants: 83.1%, 76.3%, 73.3%, 68.2%, 74.5%, 86.5%, 47.9%, 40.0% Box and blocks test (BBT): At screening: mean score = 14.3 ± 10.0, mean change after therapy = 4.3 ± 4.5 (range 0–12). Higher score changes in participants who demonstrated SMR control but not significant (p = 0.199) |
Ono et al., 2016-a [81] | N = 21 (9F:12 M) Chronic stroke patients with hemiplegic hands Mean age: 57.9 ± 2.4y | Probe congruent vs. incongruent MI feedback strategies Control groups: congruent (synchronous proprioceptive and visual feedback) and incongruent (proprioceptive feedback given 1 s after visual) | Grasping of a tennis ball with a hand Cue: visual (video of hand performing action) Feedback: visual and kinaesthetic | EEG: 9 channels to classify ERD Device: Power Assist Hand—Team ATOM, Atsugi, Japan Actuation: pneumatic motors with rigid links Control: trigger | Mean classification accuracies: Congruent feedback = 56.8 ± 5.2%, chance level = 36.4 ± 4.5% Incongruent feedback = 40.0 ± 3.5%, chance level 35.4 ± 4.5% |
Tsuchimoto et al. 2019 [84] | N = 18 (3F:14 M) Chronic haemorrhagic or ischaemic stroke (from 2mo onwards) Mean age: 58 ± 10y | Implementation of MI-controlled robotic orthosis as neurofeedback Control: SHAM | Extension of hand finger Cue: unspecified Feedback: kinaesthetic and electrical stimulation | EEG: 5 channels to classify MI Device: robotic finger orthosis Actuation: servo motors with rigid links Control: trigger | Significant time-intervention interaction in the ipsilesional sensorimotor cortex. Higher coactivation of sensory and motor cortices for neurofeedback group in the ipsilesional sensorimotor cortices as compared to SHAM |
Wang et al. 2018 [83] | N = 24 (4F:20 M) Chronic stroke patients with paralysed hands Mean age: 54 ± 9y | Implementation of action observation and motor imagery (AO + MI) with kinaesthetic feedback Control: SHAM | Hand grasping Cue: visual (video of hand action / textual cues in SHAM group) Feedback: visual and kinaesthetic | EEG: 16 channels to classify ERD Device: robot hand Control: Trigger | AO + MI with kinaesthetic feedback group showed significant improvements in FMA-UE across longitudinal evaluation [χ2(2) = 7.659, p = 0.022], no significant difference in SHAM group [χ2(2) = 4.537, p = 0.103] |
Studies involving healthy participants | |||||
Bauer et al. 2015 [97] | N = 20 (11F:9 M) Right-handed Mean age: 28.5 ± 10.5y | Study on MI as compared to motor execution (ME) using BCI-device | Opening of left hand Cue: auditory Feedback: kinaesthetic | EEG: 31 channels to detect ERD, with EMG to classify MI from execution and account for tonic contraction Device: Amadeo, Tyromotion, Austria Control: discontinuation of ERD stops finger extension | Principal component analyses (between MI and execution) generated coefficients for the visual (VIS) and kinaesthetic (KIS) imagery scale, BCI-robot performance (BRI), tonic contraction task (MOC) and visuomotor integration task (VMI). VIS and KIS yielded high coefficients on MI while MOC and VMI yield high coefficients on ME. BRI show high coefficient yields on both MI and ME |
Cantillo-Negrete et al. 2015 [86] | N = 1 | Design and implementation of a MI-controlled hand orthosis | Extension-flexion of right-hand finger Cue: visual (modified Graz) Feedback: kinaesthetic | EEG: 11 channels to detect MI Device: 1-DOF hand finger orthosis Actuation: DC motor with screw system for linear displacement, flexible links Control: trigger | Classification accuracy = 78% |
Chowdhury et al., 2015-a [87] | N = 6 Age range: 20-30y | Study of cortico-muscular coupling in robotic finger exoskeleton control | Extension-flexion of hand fingers Cue: visual Feedback: kinaesthetic | EEG: 10 channels with EMG to classify MI Device: 3-finger, 3-DOF each, exoskeleton (thumb, index, middle) Actuation: servomotors with rigid links Control: trigger | Mean classification accuracies: passive execution = 69.17%, hand execution = 71.25%, MI = 67.92% |
Coffey et al. 2014 [92] | N = 3 (All male) Right-handed Age range: 24-28y | Design and implementation of a MI-controlled hand orthosis | Hand digit and wrist contraction and extension Cue: visual (Graz MI) Feedback: kinaesthetic | EEG: 27 channels to classify MI Device: hand glove controlled by Arduino Actuation: pneumatic Control: trigger | Glove inflation-deflation cycle = 22 s Classification accuracies of 3 participants: A = 92.5%, B = 90.0%, C = 80.0% |
Diab et al. 2016 [103] | N = 5 | Design and implementation of EEG-triggered wrist orthosis with accuracy improvement | Hand opening and closing Cue: verbal instruction Feedback: kinaesthetic | EEG: 14 channels to detect hand movement-related EEG Device: actuated Talon wrist orthosis Actuation: linear Control: trigger | Mean classification accuracies: simulation studies = 95%, online BCI training = 86% |
Fok et al. 2011 [102] | N = 4 | Design and implementation of a MI-controlled hand orthosis | Hand opening and closing Cue: unspecified Feedback: visual (cursor movement) and kinaesthetic | EEG: 14 channels to detect MI-related ERD Device: actuated Talon wrist orthosis Actuation: linear actuator Control: trigger | EEG signals from imagined hand movement was correlated with the contralesional hemisphere and utilised to trigger the actuation of orthosis ERD was detected from 12 Hz bin power of EEG during move condition |
Li et al. 2019 [88] | N = 14 (4F:10 M) Mean age: 23.8 ± 0.89y | Design and implementation of an attention-controlled hand exoskeleton with rigid-soft mechanism | Hand grasping Cue: visual (video of hand action) Feedback: kinaesthetic | EEG: 3 channels to map signals relative to attention Device: hand exoskeleton Actuation: linear actuator with rigid-soft mechanism Control: Trigger | Mean classification accuracy: 95.54% actuation success rate against the attention threshold |
Holmes et al. 2012 [93] | N = 6 (All male, young adults) | Design and implementation of a MI-controlled hand orthosis | Hand opening and closing Cue: textual Feedback: kinaesthetic | EEG: 14 channels to detect hand movement-related EEG Device: ExoFlex Hand Exoskeleton controlled by Arduino Actuation: linear actuator connected to chained links that flex Control: trigger | Classification accuracies of 6 participants: T001 = 95%, T002 = 98%, D001 = 91%, U001 = 93%, E001 = 87%, E002 = 86% |
King et al. 2011 [104] | N = 1 (Female) 24y | Contralateral control of hand orthosis using EEG-based BCI | Right hand idling and grasping Cue: textual Feedback: visual and kinaesthetic | EEG: 63 channels to control contralateral hand movement Device: hand orthosis Actuation: servomotors attached to Bowden cables as tendons Control: trigger | Offline classification accuracy = 95.3 ± 0.6%, p < 3.0866 × 10−25 Average lag from voluntary contractions to BCI-robot control = 2.24 ± 0.19 s (after 5 sessions) |
Naros et al. 2016 [98] | N = 32 (16F:16 M) Mean age: 25.9 ± 0.5y | 2 × 2 factorial design with parameters: adaptive classifier threshold and non-adaptive classifier threshold, contingent feedback and non-contingent feedback | Opening of right hand Cue: auditory Feedback: kinaesthetic | EEG: 32 channels to detect ERD, with EMG to classify MI (FC3, C3, CP3 used) Device: Amadeo, Tyromotion, Austria Control: trigger | Significant enhancement in group 1 (adaptive classifier + contingent feedback), p = 0.0078 Significant reduction in group 4 (non-adaptive classifier + non-contingent feedback), p = 0.0391 Motor performance improvement over baseline from first and last tasks, significant results: Group 1 (adaptive classifier + contingent feedback), p = 0.0313 Group 4 = (non-adaptive classifier + non-contingent feedback), p = 0.0411 |
Ono et al. 2018-b [100] | N = 28 Right-handed except 1 | Implementation of an action observation strategy with visual and proprioceptive, or auditory feedback to MI Control group: SHAM | Grasping of a tennis ball with a hand Cue: visual (video of hand performing action) Feedback: visual, kinaesthetic and auditory | EEG: 9 channels to classify ERD Device: Power Assist Hand—Team ATOM, Atsugi, Japan Actuation: pneumatic motors with rigid links Control: trigger | AO + MI + proprioceptive and visual feedback: Mean MI-ERD powers of correct feedback vs SHAM provide significant interaction, F1,17 = 6.618, p = 0.020 (6 days) Statistically significant increase in MI-ERD power in correct feedback group over baseline, p = 0.012 (6 days) |
Stan et al. 2015 [94] | N = 9 | Trigger a hand orthosis using a P300 speller BCI | Spell E (enable), A (hand opening) and B (hand closing) in P300 speller BCI to perform hang grasping, moving and releasing objects Cue: textual (spelling) Feedback: visual (textual) and kinaesthetic | EEG: 8 channels focusing on visual cortex Device: hand orthosis Actuation: 2 servomotors and current feedback circuitry Control: trigger | Mean classification accuracies: 100% (on 6th letter flash during calibration) |
Ramos-Murguialday et al. 2012 [95] | N = 23 Mean age (contingent positive, contingent negative, SHAM): 26.6 ± 4y, 26.5 ± 5y, 26.2 ± 2y | Probing MI with proprioceptive feedback Experimental groups: contingent positive, contingent negative feedback Control group: SHAM | 5 tasks: MI without direct control, MI with direct control, passive, active, rest Cue: auditory Feedback: visual and kinaesthetic | EEG: 61 channels with EMG to classify ERD/ERS Device: hand orthosis Actuation: DC motor M-28 with a worm gearhead and Bowden cables for each finger Control: trigger | Contingent positive feedback provided higher BCI performance during MI without feedback than contingent negative and SHAM; and higher during MI with or without feedback as compared to rest |
Ramos-Murguialday and Birbaumer 2015 [96] | N = 9 Right-handed Mean age: 26.6 ± 4y | Detect oscillatory signatures of motor tasks during EEG | 5 tasks: MI without direct control, MI with direct control, passive, active, rest Cue: auditory Feedback: visual and kinaesthetic | EEG: 61 channels with EMG to classify ERD/ERS Device: hand orthosis Actuation: DC motor M-28 with a worm gearhead and Bowden cables for each finger Control: trigger | Significant change in power in all frequency ranges during MI with direct control before trial initiation Kinaesthetic feedback increased significant changes in alpha and beta power; therefore, increasing BCI performance |
Randazzo et al. 2018 [90] | N = 9 (2F:7 M) Mean age: 23 ± 5y | Design and implementation of a hand orthosis with testing of kinaesthetic effects in EEG | 4 tasks: rest (REST), exoskeleton-induced hand motions (EXO), MI of right hand (MI), exoskeleton-induced hand motions plus MI (MIEXO) Cue: visual Feedback: kinaesthetic | EEG: 16 channels to detect MI Device: mano hand exoskeleton Actuation: linear servomotors attached to Bowden cables as tendons Control: passive (exoskeleton not dependent on MI to move during MIEXO task) | Mean classification accuracies among groups: (vs REST) MI = 63.02 ± 5.91%, EXO = 69.64 ± 5.74%, MIEXO = 72.19 ± 6.57% MIEXO vs EXO = 69.91 ± 9.86% Chance level at 95% confidence = 58% (N = 50 trials) |
Tacchino et al. 2017 [91] | N = 8 (7F:1 M) Right-handed Mean age: 26.3 ± 1.9y | 2 × 2 factorial design with parameters: glove, no glove, active movement, passive movement | Opening and closing of hand, 4 tasks: (A) glove with active movement, (B) glove with passive movement, (C) no glove with active movement, (D) no glove and no movement Cue: auditory Feedback: kinaesthetic | EEG: 19 channels with EMG to detect ERD/ERS (C3, F3, Cz used) Device: Gloreha hand rehabilitation glove Actuation: electric actuators with Bowden cables on each finger Control: passive (glove not dependent on brain-state during tasks) | Statistically significant ERD changes in beta and mu bands were observed to initiate earlier in tasks A and C (involves active movement) Stronger and longer ERD was observed in tasks A and B (involves robotic assistance) suggesting reinforced afferent kinaesthetic feedback |
Vukelic and Gharabaghi 2015 [99] | N = 11 (4F:7 M) Right-handed Mean age: 25.83 ± 3.1y | Assessment sensorimotor activity during MI with either visual or kinaesthetic feedback | Right hand opening Cue: visual (coloured cursor ball) Feedback: visual and kinaesthetic (separated by experimental groups) | EEG: 128 channels to detect ERD/ERS during MI (F3, CP3, C3 used) Device: Amadeo, Tyromotion, Austria Control: trigger | MI + kinaesthetic feedback group resulted in higher beta ERS (p = 0.02) during rest and higher beta ERD (p = 0.04) during MI Kinaesthetic feedback provides higher stability and sustained beta ERD activity than visual feedback |
Witkowski et al. 2014 [101] | N = 12 (4F:8 M) Right-handed Mean age: 28.1 ± 3.63y | Assessment performance and safety of EEG-EOG hybrid BCI | Right hand grasping Cue: visual (coloured squares and arrows) Feedback: kinaesthetic | EEG: 5 channels with EOG and EMG to detect ERD during MI Device: HX hand exoskeleton Actuation: DC motors with Bowden cables for thumb and index fingers Control: trigger | Mean classification accuracies: EEG only = 63.59 ± 10.81% EEG/EOG hybrid = 60.77 ± 9.42% Mean safety criterion violations during rest: EEG only = 45.91 ± 26.8% EEG/EOG hybrid = 10.14 ± 0.3% |
Zhang et al. 2019 [89] | N = 6 (2F:4 M) Right-handed Age range: 23-26y | Implementation of a multimodal system using EEG, EMG and EOG to control a soft-robotic hand | Graz visualisation and auditory instructions, eye movements and physical practice (hand gestures) Cue: visual (Graz MI), auditory Feedback: visual and kinaesthetic | EEG with EMG and EOG: 40 channels to analyse ERD/ERS patterns Device: Soft pneumatic finger Actuation: pneumatic actuator with soft structures Control: trigger | Mean classification accuracies: EOG = 94.23% EEG = 31.46% EMG = 36.38% Multimodal = 93.83 ± 0.02% |