Archives of Physical Medicine and Rehabilitation

journal homepage: www.archives-pmr.org

Archives of Physical Medicine and Rehabilitation 2024;105: 2336−49

REVIEW ARTICLE (META-ANALYSIS)

Neural Interface-Based Motor Neuroprosthesis in

Poststroke Upper Limb Neurorehabilitation: An
Individual Patient Data Meta-analysis
Yu Tung Lo, MBBS,a,b,* Mervyn Jun Rui Lim, MBBS, MPH,c,d,* Chun Yen Kok,a
Shilin Wang, MBBS,a Sebastiaan Zhiyong Blok,d Ting Yao Ang, MBBS,a
Vincent Yew Poh Ng, MBBS,a Jai Prashanth Rao, MBBS,a,b Karen Sui Geok Chua, MBBSd,e,f
From the aDepartment of Neurosurgery, National Neuroscience Institute; bDuke-NUS Medical School; cDepartment of Neurosurgery, National
University Hospital; dNational University of Singapore, Yong Loo Lin School of Medicine; eInstitute of Rehabilitation Excellence, Tan Tock Seng
Hospital Rehabilitation Centre; and fLee Kong Chian School of Medicine, Nanyang Technological University, Singapore.

Abstract
Objective: To determine the efficacy of neural interface−based neurorehabilitation, including brain-computer interface, through conventional and
individual patient data (IPD) meta-analysis and to assess clinical parameters associated with positive response to neural interface−based
neurorehabilitation.
Data Sources: PubMed, EMBASE, and Cochrane Library databases up to February 2022 were reviewed.
Study Selection: Studies using neural interface−controlled physical effectors (functional electrical stimulation and/or powered exoskeletons) and
reported Fugl-Meyer Assessment−upper-extremity (FMA-UE) scores were identified. This meta-analysis was prospectively registered on PROS-
PERO (#CRD42022312428). PRISMA guidelines were followed.
Data Extraction: Changes in FMA-UE scores were pooled to estimate the mean effect size. Subgroup analyses were performed on clinical param-
eters and neural interface parameters with both study-level variables and IPD.
Data Synthesis: Forty-six studies containing 617 patients were included. Twenty-nine studies involving 214 patients reported IPD. FMA-UE
scores increased by a mean of 5.23 (95% confidence interval [CI]: 3.85-6.61). Systems that used motor attempt resulted in greater FMA-UE gain
than motor imagery, as did training lasting >4 vs ≤4 weeks. On IPD analysis, the mean time-to-improvement above minimal clinically important
difference (MCID) was 12 weeks (95% CI: 7 to not reached). At 6 months, 58% improved above MCID (95% CI: 41%-70%). Patients with severe
impairment (P=.042) and age >50 years (P=.0022) correlated with the failure to improve above the MCID on univariate log-rank tests. However,
these factors were only borderline significant on multivariate Cox analysis (hazard ratio [HR] 0.15, P=.08 and HR 0.47, P=.06, respectively).
Conclusion: Neural interface−based motor rehabilitation resulted in significant, although modest, reductions in poststroke impairment and should
be considered for wider applications in stroke neurorehabilitation.
Archives of Physical Medicine and Rehabilitation 2024;105:2336−49
Ó 2024 by the American Congress of Rehabilitation Medicine. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Stroke has an enormous disease burden, with a 12.2 million global Some therapeutic options include conventional therapy (directed
incidence and a 101 million prevalence in 2019.1 It is the leading by physiotherapists),2 robot-assisted therapy (involving the use of
cause of long-term disabilities worldwide, and the third leading robotic exoskeletons to supplement patients’ own movement),3
cause of combined death and disability.1 Stroke rehabilitation is and constraint-induced movement therapy (CIMT, in which the
essential to improve function and reduce this burden of care. neurologically intact limb is restrained to induce the use of the
weaker arm).4 However, it often is difficult or impossible for the
stroke-affected limb to be used in high-intensity exercises in treat-
Clinical Trial Registration No.: #CRD42022312428
Disclosures: none.
ment such as CIMT.5,6 Neural interfaces, which include brain-
* Lo and Lim contributed equally as co−first authors. computer interfaces (BCIs) and myoelectric interfaces, allow the

0003-9993/$36 - see front matter Ó 2024 by the American Congress of Rehabilitation Medicine. Published by Elsevier Inc. This is an open access article
under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
https://doi.org/10.1016/j.apmr.2024.04.001
Motor BCI for stroke rehabilitation 2337

control of neuroprosthesis or functional electrical stimulation shed light on how much impairment gain could be expected. We
(FES) and enable patients with little or no residual limb functions further qualified this finding by tracking the time-to-improvement
to mobilize the paralyzed limb through imagined or attempted through Kaplan-Meier analysis of IPD extracted from the study,
movements. Importantly, neural interfaces also allow closed-loop, and we investigated whether any phenotypic variables correlated
functionally contingent proprioceptive feedback. with impairment gain. Unlike previous reviews that provided gen-
Neural interfaces can be linked to various end effectors to form a eral overview of the neural interface systems used in neurorehabi-
closed system. These systems include virtual reality−based sys- litation,18-25,27 we further reduced heterogeneity by confining our
tems,7 visual feedback systems (eg, through an avatar limb shown analysis to only BCI-driven physical neuroprosthesis systems
on a screen),8 or other nonanthropomorphic signals (eg, moving a because such systems, in principle, provide the most natural and
screen cursor sideways).9,10 These systems augment stroke recovery anthropomorphic feedback (both proprioceptive and visual).
by facilitating use-dependent neuroplastic changes.11,12 They can
also be further combined with gamification to improve user engage-
ment and motivation.13 Several studies have demonstrated that the
addition of motor imagery (MI) per se to robotic therapy has addi-
Methods
tional benefits due to its effects on strengthening the sensorimotor
feedback loop.6 In addition, unlike virtual feedback systems, motor Search strategy
commands actuated either via an external exoskeleton or through
FES of the patient’s own limbs provide physical feedback. Literature databases PubMed, Embase, and Cochrane Library
To date, there have been several studies with different study were searched through inception to February 19, 2022. The proto-
designs and patient characteristics that tested the effectiveness of col was registered on the international prospective register of sys-
various BCI-assisted rehabilitation systems.13-16 Previous meta- tematic reviews (PROSPERO, CRD42022312428) (supplemental
analyses, such as a recent one by Bai et al,17 compared study-level materials). The PRISMA (Preferred Reporting Items for System-
characteristics and concluded that BCI has more beneficial effects atic Reviews and Meta-Analyses) checklist was followed. Key
compared to control (standardized mean difference [SMD] of 0.42). terms used included “brain-computer interface,” “neural inter-
Similarly, two similar meta-analyses reported an SMD of 0.48 face,” “neuroprosthesis,” “stroke,” “intracerebral hemorrhage,”
(Nojima et al)18 and 0.62 (Xie et al),19 both in favor of BCI over and their respective synonyms (refer to supplemental materials for
their respective controls. Qu et al found no difference between BCI the exact search terms). In addition, a Google search was con-
and a robotic therapy−only group (SMD of 1.09).20 However, these ducted using these terms, and relevant articles were also identified
meta-analyses pooled various outcome measures (for instance, through the references of included studies.
Fugl-Meyer Assessment [FMA] score, Manual Function Test, and
Jebsen-Taylor Hand Function Test), and each of the included stud- Scope
ies defined “control group” differently (some compared to standard
arm therapy, others to robotic therapy); it is not immediately appar- Inclusion criteria: We specifically evaluated studies that utilized
ent what the SMD actually quantifies. There have been further sys- physical effectors—that is, exoskeletons, prostheses, or FES—
tematic or narrative reviews about BCI in general in stroke controlled by neural signals (we included myoelectric signals in
rehabilitation,21-25 as well as reviews focusing on neuroprostheses this study). These systems all physically move the patients’ own
(including ones not controlled by neural or myoneural signals).26,27 limbs and generate proprioceptive feedback. Such closed-loop
Still, while the pooled findings were in favor of BCI, a substantial neural control systems must necessarily consist of 3 components
number of studies reported in these reviews did not manage to show (figure 1): (1) a signal source (eg, electroencephalogram [EEG],
significant differences. Furthermore, an analysis of more granular electrocorticography, electromyography [EMG], functional neuro-
individual patient data (IPD) to determine patient-level characteris- imaging); (2) a decoder; and (3) physical end effector (either a
tics that could explain these differences in outcome has not, to our powered orthosis or FES). The decoder could be machine-learning
knowledge, been undertaken. algorithms, such as linear discriminant analysis; the simple detec-
We, hence, sought to quantify the absolute impairment gain tion of an event-related potential like event-related desynchroniza-
reported in the literature, rather than to compare SMD, to better tion; or a simple threshold (eg, a specific amplitude for EMG-
triggered orthosis). The physical end-effector could either be
external (eg, orthosis or prosthesis) or internal (FES), but it must
List of abbreviations:
be controlled by neural signals rather than preprogrammed move-
ARAT action research arm test ments. All 3 components must have been present for the study to
BCI brain-computer interface
be included. We defined outcome measures as FMA scores or
CIMT constraint-induced movement therapy
CI confidence interval
Action Research Arm Test (ARAT) scores, which are 2 of the
EEG electroencephalogram most commonly and consistently used scores,28,29 and included
EMG electromyogram only studies reporting these scores.
FES functional electrical stimulation
FMA Fugl-Meyer
FMA-UE Fugl-Meyer Assessment upper extremity Study selection
HR hazard ratio Five coauthors (YTL, CY, SLW, TYA, ML) independently partici-
IPD individual patient data
pated in the screening process, facilitated by the Rayyan platform,a
MCID minimal clinically important difference
MI motor imagery an online collaborative systematic review and meta-analysis plat-
NIH National Institute of Health form.30 Inclusion and exclusion decisions were made by at least 2
SMD standardized mean difference of the independent, blinded reviewers. Conflicts were resolved
through mutual discussion with the other reviewer(s) on the

www.archives-pmr.org
2338 Y.T. Lo et al

Fig 1 The type of motor BCI system included in this review, which consists of (1) a signal source (eg, EEG, EMG, magnetoencephalography), (2)
a decoder/transformer that transforms neural signals to motor outputs, and (3) an effector (eg, an exoskeleton/powered orthosis, or FES).

Rayyan platform. Any remaining discrepancies were resolved by grasp (6 items), grip (4 items), pinch (6 items), and gross move-
the lead author, YTL. ment (3 items).29 Higher scores indicate less impairment for both
To avoid double-counting,31 for studies with substantially sim- FMA-UE and ARAT.
ilar patient cohorts, only the study with the most complete cohort
was chosen.
IPD analysis
Variables collected IPD were extracted from the included studies. Studies with mixed
cohorts (eg, stroke patients and normal subjects) from which IPD
We collected clinical parameters, including age, time since stroke
of stroke patients could not be isolated were excluded. We utilized
(in mo), type of stroke (hemorrhagic or ischemic), location (corti-
a 1-stage method for IPD meta-analysis followed by the Kaplan-
cal vs subcortical or both) and the presence or absence of voli-
Meier method to construct a cumulative incidence curve.34 The
tional hand (wrist or finger) actions. Additionally, we collected
event of interest was defined as FMA-UE gain above the MCID of
neural interface−related variables, namely the type of control sig-
5.25,33 and the time-to-event was measured in months. The analy-
nal (eg, EEG, magnetoencephalography, functional near-infrared
sis followed the PRISMA-IPD guidelines,35 and attempts were
spectroscopy, EMG), the type of joint movements controlled
made to contact the authors for access to IPD. We further
(upper and/or lower extremities, proximal and/or distal), and the
accounted for within-study clustering of participants using the
end effector utilized (ie, FES or powered exoskeleton/orthosis).
gamma frailty model, which represents an unobserved random
For cortical signals, the laterality of signals (ipsilesional, contrale-
effect shared by patients within each study.36
sional, or both), task performed (ie, MI or motor attempt) were
also collected. For cases in which a study cited another study, the
referenced paper was reviewed to gather these variables. Supple- Risk of bias assessment
mental materials of the respective articles were searched to clarify
any information. The National Institute of Health (NIH) quality assessment tool for
For the FMA and ARAT scores, we collected the means and pre-post studies was utilized in this analysis, treating all included
SDs of pre- and post-BCI intervention scores. The longest follow- cohorts as pre-post cohorts. Two independent reviewers (YTL and
up outcome was selected for each included study for study-level SB) evaluated the studies. The assessment criteria included the
analyses. The motor upper-extremity subscore (Fugl-Meyer definition of the study question, eligibility assessment, representa-
Assessment upper-extremity [FMA-UE], maximum score=66) tiveness of the cohort, enrollment process, adequate sample size,
was reported.28 The FMA-UE scores were categorized into 3 consistency of the intervention, consistency of the outcome meas-
severity groups based on Woytowicz et al’s cutoff values32: severe ures, appropriate statistical analysis, and utilization of multiple
(0-28), moderate (29-42), and mild (43-66). In addition to report- outcome measures.37 The overall quality of each study was also
ing median differences in FMA-UE scores, we defined the mini- assessed. Publication bias was examined using a funnel plot and
mal clinically important difference (MCID) as 5.25 points.33 The was further quantified using Egger’s regression and Begg’s rank
ARAT score is assessed on a scale of 0-57 and includes 4 subtests: correlation test.38,39

www.archives-pmr.org
Motor BCI for stroke rehabilitation 2339

Fig 2 PRISMA flow diagram for study selection.

Statistical analysis in table 1. Comparing all studies with IPD (the latter being a sub-
set of the former), some baseline characteristics were significantly
Continuous variables were reported as means and SDs, while cate- different from each other (table 1). Additional qualitative informa-
gorical data were presented as counts and percentages. The statis- tion about the included studies can be found in supplemental
tical software R (version 4.2.3)b was utilized for conducting both table 1.
the meta-analysis and Kaplan-Meier analysis. Effect sizes were Most studies (n=31) involved distal effector (either powered
pooled using the DerSimonian-Laird method. Heterogeneity was orthosis or FES, for control of wrist and/or finger joints),6,8,14-
assessed using the I2 test (>40% considered high) and tau-squared 16,40-65
6 studies involved proximal effector (controlling shoulder
statistics. Prediction intervals were estimated. The log-rank test and/or elbow joints),5,13,66-69 and 5 studies involved both distal
was utilized to compare event rates between groups. and proximal effectors.70-74 Two involved lower limb
effectors.75,76

Results Functional outcomes

Study selection Using a random-effects model, the pooled improvement in FMA-

UE scores was 5.23 points (95% confidence interval [CI]: 3.85-
A total of 1181 articles were identified through an electronic data- 6.61). The effect sizes were highly homogeneous across studies,
base search and an additional 7 articles through a citation search with I2 of 0% (95% CI: 0%-44.6%, P=.51), with t 2 of 0 (95% CI:
and web search. After a title and abstract screening and then a 0-7.9) (figure 3).
full-text search, 46 articles with unique patient cohorts were
included in the meta-analysis. There were 38 interventional stud- Study-level subgroup analyses
ies, 7 case reports, and 1 abstract (supplemental table 1). Out of
the included studies, 29 contained IPD, with a total of 214 unique We performed subgroup analyses for study-level variables: dura-
patients. The PRISMA flowchart for study selection is shown in tion of BCI use (≤4 wk vs >4 wk), effector type (powered orthosis
figure 2. vs FES), BCI task (MI vs motor attempt), and extent of wrist/fin-
ger extension (no to minimal movement vs mixed severity/not
Study characteristics specified; see supplemental table 3 for the specific patterns of defi-
cits inferred from the study). Only the motor attempt subgroup, on
A total of 617 stroke patients from 46 studies with distinct cohorts aggregate, demonstrated significant improvement above the
were included in the analysis. The demographics are summarized MCID of 5.25 (table 2). None of these were significant after

www.archives-pmr.org
2340 Y.T. Lo et al

Table 1 Baseline characteristics of the pooled population

Parameter All Included Studies Studies With IPD P Value
Number of studies 46 29 -
Number of subjects 617 214 -
Age, mean § SD 55.8 (12.1) 55.4 (12.7) .681
Time since stroke, mo, mean § SD 34.2 (40.3) 41.1 (47.3) .040
Stroke type, n (%) .025
Ischemic 297 (48.1) 119 (55.6)
Hemorrhagic 167 (27.1) 57 (26.6)
Not specified 153 (24.8) 34 (15.9)
Signal type, n patients (%) .001
EEG only 512 (83.0) 198 (92.8)
EMG only 56 (9.1) 4 (1.9)
EEG+EMG 43 (7.0) 10 (4.7)
fNIRS 4 (0.6) 0 (0)
MEG+EEG 1 (0.2) 1 (0.5)
Brain implant 1 (0.2) 1 (0.5)
Effector type, n patients (%) .002
Powered exoskeleton 393 (63.7) 110 (51.4)
FES 170 (27.6) 86 (40.2)
Combined 54 (8.8) 18 (8.4)
Duration of BCI use, n patients (%) .906
≤4 wk 346 (56.1) 121 (56.5)
>4 wk 271 (43.9) 93 (43.5)
Baseline FMA-UE score* .016
Number of studies 33 22
Number of patients 381 136
Mean § SD 21.9 (11.1) 24.7 (13.0)
Mild (43-66), n (%) - 14 (10.3)
Moderate (29-42), n (%) - 29 (21.3)
Severe (0-28), n (%) - 93 (68.4)
Baseline ARAT score .257
Number of studies 15 8
Number of patients 206 60
Mean § SD 15.6 (13.2) 18.1 (20.0)
NOTE. The Student t test was performed for continuous variables and the Chi-square test for categorical variables. All P values were 2 tailed.
Abbreviation: fNIRS, functional near-infrared spectroscopy; MEG, magnetoencephalography.
* Only for studies reporting the full FMA-UE score (ie, total score of 66).

correction for multiple comparisons. The corresponding forest respectively, log-rank test), with severe FMA-UE at baseline and
plots are shown in supplemental figures S1-S4. age >50 years associated with lower event rates, while time-since-
stroke and stroke type were nonsignificant (P=.79 and 0.18,
IPD analysis respectively) (figure 5). The median time-to-improvement above
MCID was 4, 6, and 26 weeks for those with mild, moderate, and
A total of 29 studies provided individually reported data on 214 severe stroke severity, respectively (as measured by FMA-UE),
patients, but the full FMA-UE score was reported in only 136 and 6 and 43 weeks for ≤50 years and >50 years, respectively
patients. Only 1 study provided ARAT scores without FMA-UE (figure 5). However, on further testing with multivariate Cox
scores. Since ARAT and FMA scores are highly correlated,77 sep- regression, only severe impairment (hazard ratio [HR] 0.15,
arate Kaplan-Meier analyses for ARAT changes were not P=.08) and age >50 years (HR 0.47, P=.06) remained borderline
repeated. The functional scores are described in table 1. These significant, albeit with moderate effect sizes (ie, HRs). The gamma
patients were regarded as a single cohort, and a multi-study frailty term was not significant (P=.94), indicating that there was
Kaplan-Meier time-to-event analysis was conducted (figure 4). negligible unobserved heterogeneity between studies (table 3).
The median time-to-improvement above MCID was 12 weeks
(95% CI: 7 to not reached), and at 26 weeks (ie, 6 mo) 58%
improved above MCID (95% CI: 41%-70%). Four studies Risk of bias assessment
reported outcome beyond 6 months.14,45,65,74 Study-level variables
described above were not re-analyzed (since all patients within We assessed the risk of bias using the NIH Quality Assessment
each study would have the same parameter). tool with respect to the FMA-UE scores (figure 6). Excluding the
Baseline FMA-UE and age resulted in significantly different 5 studies with a high risk of bias, we observed an insignificant
time-to-improvement above MCID (P=.042 and .0022, change in the pooled effect size in this sensitivity analysis (5.34

www.archives-pmr.org
Motor BCI for stroke rehabilitation 2341

Fig 3 Forest plot of the pooled improvement in FMA-UE scores after BCI neurorehabilitation. Case studies and abstracts were excluded from the
meta-analysis.

increase in FMA-UE, 95% CI: 3.6-7.0, n=20; P=.86 compared to neural interface−based neurorehabilitation. This improvement
the 5 studies with a high risk of bias). was statistically significant, although not significantly higher than
the MCID of 5.25.33 This pooled estimate complements other
Publication bias recent meta-analyses of similar studies that demonstrated superior-
ity of BCI-based rehabilitation over its alternatives.17-19 Severe
We found no evidence of publication bias based on the funnel plot impairment (HR 0.15, P=.08) and age >50 years (HR 0.47, P=.06)
(figure 7), and neither Egger’s test nor Begg’s rank correlation test were borderline significant in their associations with the failure to
indicated any significant small-study effect (P=.084 and .384, improve above MCID on multivariate analysis.
respectively).
Identifying optimal patient subgroups
The effects of BCI could be heterogeneous across different patient
Discussion phenotypes, but previously, it was not clear how best to identify
At the study level, we observed a modest average improvement of patients most likely to respond to BCI neurorehabilitation or iden-
5.23 points in the FMA-UE score (95% CI: 3.85-6.61) following tify if clinical parameters were associated with therapy response.

Table 2 Subgroup analyses based on study-level variables

Parameter Subgroups, Effect Size (95% CI, I2 [%], n) P Value
Duration of BCI use ≤4 wk: 4.1 >4 wk: 7.3 .047
(2.4-5.9, I2: 0%, n: 14) (4.7-9.8, I2: 9.6%, n: 10)
Effector type Powered orthosis: 4.3 FES: 7.1 .059
(2.6-6.0, I2: 0%, n: 14) (4.7-9.5, I2: 0%, n: 10)
BCI task MI: 4.2 Motor attempt: 9.0 .021
(2.6-5.8, I2: 0%, n: 18) (6.0-12.0, I2: 0%, n: 5)
Wrist or finger extension Absent or minimal: 6.0 Mixed or not specified: 4.7 .362
(3.8-8.3, I2: 13.5%, n: 8) (2.8-6.6, I2: 0%, n: 16)
NOTE. The corresponding forest plots are shown in supplemental figures.

www.archives-pmr.org
2342 Y.T. Lo et al

Fig 4 Overall pooled Kaplan-Meier curve for all the patients with reported IPD, with events being defined as those with an FMA-UE score
increase of MCID ≥5.25.

We found on univariate IPD analysis that patients with severe the anterior deltoids, triceps, and wrist extensors and assisted the
stroke impairment and older patients improved less after BCI ther- patients’ own attempted movement, rather than completely taking
apy, although multivariate analysis failed to show statistical sig- over the movement (the authors reported that FES alone was not
nificance. Nonetheless, the comparisons may be underpowered, able to generate enough power to lift the arm). They observed that
and the pooled effect sizes were moderate; future studies should there was improved upper limb kinematics during reaching move-
further evaluate these parameters. In principle, closed-loop senso- ments. Zhang et al provided another example of a subject who
rimotor activation may be more beneficial for those with no resid- used a hybrid FES-orthosis setup to perform a series of 3 progres-
ual limb functions and those who must otherwise rely on sively challenging tasks.69 An EEG-driven powered orthosis was
therapists or robotic systems to move the paralyzed limb. It is also used for elbow extension (ie, wiping a plate carried in the other
possible that those with milder impairment may also benefit from hand) and flexion (ie, picking up a bucket from in front of them)
an assistive form of BCI that supports their own attempted move- for the first and second tasks, respectively; for the third, more diffi-
ments to allow the limb to complete a wider range of movement. cult task (ie, picking up a ball and then releasing it), FES was used
For patients with severe impairment (68.4% of those with to assist hand opening. There was drastic improvement in reaching
IPD), the commonest group of patients indicated for BCI rehabili- ability and coordination of proximal muscles (eg, hand-to-box and
tation, the BCI system completely takes over the function of the forearm-to-box tasks on the Wolf Motor score), although there
plegic limb. However, this scenario need not be the sole indica- was no FMA score improvement. Bhagat et al included chronic
tion. Several of the included studies also included patients with stroke patients who have plateaued in FMA-UE score recovery
moderate impairment (14 of the 88 patients, or 16%, from studies over the preceding 3 months, including some with relatively unim-
that reported FMA-UE change). Many of these patients still paired hand or wrist movements (by FMA and ARAT component
retained some residual limb movement,48,52,69,71 and the BCI acts scores).66 After using a BCI-controlled exoskeleton that primarily
more in an assistive capacity to support the movement, resulting targeted the elbow, there was, interestingly, improvements in vari-
in qualitative improvements in the upper limb kinematics. One ous FMA-UE and ARAT domains relating to distal finger and
good example was Ibanez et al’s study,71 in which they included wrist movements, even for patients with high initial FMA-UE
chronic stroke patients (time since stroke between 3 and 5y) spe- scores—1 patient, for instance, had a high initial FMA-UE score
cifically with retained ability to manipulate most objects with their of 51 and an ARAT score of 43, and another had an FMA-UE
paretic limb (FMA-UE between 12 and 31). FES was rendered to score of 49 and an ARAT score of 42. The authors argued that

www.archives-pmr.org
Motor BCI for stroke rehabilitation 2343

(A) (B)

(C) (D)

Fig 5 Kaplan-Meier plots of the time-to-event for clinically significant improvement for the prespecified subgroups, with corresponding log-
rank P values stratified by (A) severity based on FMA-UE scores (severe 0-28, moderate 29-42, mild 43-66), (B) age (>50y vs ≤50y), (C) time since
stroke (<6, 6-12 and ≥12mo), and (D) type of stroke (hemorrhagic vs ischemic).

BCI-based neurorehabilitation engaged users better than alterna-

tive therapy, and the improved functions of proximal joints could
Table 3 Multivariate Cox proportional hazard regression, with have also led to greater functional use of the hand and wrist. Vice
gamma frailty adjustment for clustering effect of participants versa was also true because several studies also reported that train-
within studies ing of wrist and fingers also led to improvement in more proximal
joints.6,14 In this group of patients with residual limb functions
DFMA-UE Score Above MCID (which have plateaued with other rehabilitation therapies), BCI
Adjusted HR 95% CI P Value training could bring about more subtle but, nonetheless, clinically
Baseline FMA-UE groups meaningful benefits in other aspects of upper-limb functions,
Mild (43-66) Reference beyond just improvements in motor impairment per se.
Moderate (29-42) 0.29 0.03-2.56 .27 With regard to stroke chronicity, time since stroke appeared to
Severe (0-28) 0.15 0.02-1.28 .08 have no particular effect on treatment outcome. Interestingly, the
Age groups most remarkable FMA-UE score gains occurred in patients
≤50 Reference reported by Li et al, who specifically recruited patients in the acute
>50 0.47 0.21-1.03 .06 phase of stroke (ie, 1-6mo post stroke).54 For example, a patient 1
Time since stroke month post stroke with a severely impaired limb (FMA-UE score
<6 mo Reference of 9, ARAT score of 0) improved to an FMA-UE score of 40
6-12 mo 0.68 0.22-2.13 .51 (ARAT score of 28) after 8 weeks of BCI training. This finding
≥12 mo 1.02 0.40-2.60 .97 was contrasted against a control group that received only conven-
Stroke types tional physiotherapy and demonstrated poorer outcomes. While it
Ischemic Reference may not be possible to entirely rule out the effects of the natural
Hemorrhagic 1.64 0.69-3.91 .26 history of stroke recovery in a limited patient sample, it could, per-
Gamma frailty .94 haps, be considered that patients with acutely hemiplegic limbs
with little alternative means of mobilizing the affected limb could
Abbreviations: HR, hazard ratio. also benefit from early BCI-assisted rehabilitation as they recover

www.archives-pmr.org
2344 Y.T. Lo et al

Fig 6 The NIH quality assessment tool is used to quantify the risk of bias for the effect size estimates, for full studies reporting FMA-UE scores
only. Case studies and abstracts were excluded. Detailed justifications are described in supplemental table 3.

naturally, which prevents the onset of learned disuse.78 This bene- a stroke 15 years ago.14 Bundy et al also reported marked
fit appeared to be greater when BCI and robot-assisted therapy improvements in the grasp and grip ARAT subscores and grip
were used than when robot-assisted therapy was used alone.79 strength in chronic stroke patients with their IpsiHand system.41
Conversely, chronic stroke patients who had reached a functional Gaining the ability to merely extend fingers or the wrist (eg, from
plateau demonstrated additional improvement with highly inten- a score of 0 to 2) could also profoundly expand the range of activi-
sive interventions, potentially due to the reversal of learned non- ties of daily living that the patients can perform and could enable
use/disuse.80 Overall, the chronicity of stroke should probably not further rehabilitative training, even if the gain falls below the
be used to exclude patients from such neural interface−based MCID. Whether this impairment gain is sustained months or years
rehabilitation. Various other groups also came to similar conclu- after the completion of BCI training remains to be verified in
sions, albeit based on study-level comparisons. Xie, for instance, future studies.
described no difference in outcome between subacute and chronic Movement dyssynergy and disorganized coactivations of ago-
strokes (SMD of 1.11 and 0.68, respectively; P=.38).19 nistic, antagonistic, and synergistic muscles are hallmarks of post-
For those with severe impairment, those with complete hand stroke movement derangement (which are themselves measured
paralysis would stand to benefit from BCI training because hand by the FMA-UE score).81 Powered exoskeleton and, to a lesser
paralysis, arguably, has the greatest effect on activities of daily degree, FES would produce controlled kinematics, which would
living.14 BCI rehabilitation had been shown to restore hand func- not be possible for alternative therapies like CIMT, even in
tions even in those with very chronic strokes. For instance, in the patients with residual limb functions. The further addition of a
series by Biasiucci et al, several patients improved in wrist exten- closed-loop system would generate appropriate sensory feedback
sion strength from Medical Research Council grade 1 or 2 to grade to drive sensorimotor integration and relearning in the brain.6
4, including 2 with complete hand paralysis who regained wrist Beyond clinical parameters, imaging or electrophysiological
extension and some finger extension—one of whom suffered from biomarkers could potentially also predict BCI response. Stroke

www.archives-pmr.org
Motor BCI for stroke rehabilitation 2345

Fig 7 Contour-enhanced funnel plot for evaluation of small-study effect. There was no asymmetry suggestive of significant publication bias
based on small-study effect. Individual studies’ statistical significance (against null hypothesis that the effect size is zero) is indicated by the
shaded regions with their corresponding P values.

recovery in general, for instance, has been shown to be dependent attempt−based BCI systems showed a higher FMA-UE
on the structural integrity of the cortical circuitries as measured on improvement of 9.0 points (95% CI: 6.0-12.0) compared to MI
diffusion tensor imaging.82 EEG biomarkers such as the Brain (P=.021), although this finding was nonsignificant after correct-
Symmetry Index69,83 or metrics of EEG connectivity changes14 ing for multiple comparisons. This finding provides some evi-
were found to correlate with impairment gains. Future reviews dence that motor attempts may lead to better improvement in
could further elaborate on the role of these biomarkers in the con- motor impairment compared to MI and, where feasible, should
text of BCI. be considered in future BCI studies.
Unfortunately, the evidence for the above could only be
Identifying optimal treatment paradigms derived from study-level comparisons, as each included study
only evaluated a single type of neuroprosthesis. It is also not clear
We found that FES was marginally better than powered orthosis in if these modalities can be compared fairly between studies, as the
improving FMA-UE scores, although this finding did not reach operating principles and setup are very different.
statistical significance (7.1 vs 4.3, P=.059). This finding is consis-
tent with previous meta-analyses of controlled trials (albeit not Neuroplasticity in BCI-driven neurorehabilitation
head-to-head comparisons).17,19 On the one hand, FES directly
results in movement of the muscle when stimulated, as opposed to A closed-loop system has been postulated to facilitate cortical
passive movement by an orthosis. This type of movement acti- reorganization, and this phenomenon appears to involve a system
vates the Golgi apparatus and muscle spindles more strongly than of parallel networks in both hemispheres. Hebbian plasticity, a
orthosis. On the other hand, a robotic orthosis generates more con- process in which synaptic strength is strengthened when presynap-
strained kinematics and produces greater strength than FES.14 tic (ie, cortical) and postsynaptic neurons (ie, spinal cord) fire in
Whether these differences are necessarily important factors that synchrony, may underlie this process.11,12 Motor recovery has
affect use-dependent plasticity is not yet clear. been shown to be correlated with the activation and improved
Two common strategies were used to produce control sig- functional connectivity of ipsilesional somatosensory cortex, dor-
nals: MI or motor attempts. MI involves imagining limb move- sal premotor, and supplemental motor cortices,13,14,84-87 as well as
ment without physically moving the limbs and has been shown inhibition of the competing contralesional hemisphere.66,88 In
to modulate sensorimotor rhythms in the alpha (mu) and beta addition, improved connectivity between both hemispheres has
frequency bands.14,41,44,47 All EMG or myoelectric signals, on been shown to be important.13,89 The extent of recovery might
the other hand, are motor attempts. We found that motor also be dependent on the integrity of structures such as the

www.archives-pmr.org
2346 Y.T. Lo et al

parietofrontal cortex,90 and the corticospinal tract.91 There is also rehabilitation, to proceed to more permanent invasive brain-
evidence of involvement of large-scale functional networks like machine interface systems.
the dorsal attention network.92 Often, though, the cortical network
still retains some residual function after stroke and is not
completely injured, as seen on diffusion tensor imaging studies,43 Conclusion
and can relay sufficient information to control the neural interfa-
ces.93 About one-third of the studies included in our analysis used The use of BCIs and motor neuroprostheses in stroke neurorehabi-
signals from the ipsilesional cortex as the control signal. litation shows promise in restoring lost function, although chal-
lenges remain. Stroke patients with severe impairment make up
Study limitations the typical group of patients who can benefit from BCI-based
treatment, but we also noted that those with moderate impairment
Firstly, the gains in motor impairment could be partially attributed showed remarkable functional gains. There was no robust evi-
to the natural history of stroke recovery rather than the use of the dence that clinical parameters correlated with the extent of
neural interface system, especially for studies that recruited impairment reduction, and other biomarkers should be further
patients earlier in their stroke recovery. Many studies did attempt examined.
to control for this by checking for the plateau of motor recovery
prior to recruitment. Our findings should nevertheless be inter-
preted as the upper limit of functional recovery with such systems. Suppliers
Furthermore, analysis of controlled studies performed by other
reviews also demonstrated that BCI is generally more efficacious a. Rayyan platform; Rayyan.
than comparative treatments.17,19 Secondly, the MCID of 5.25 b. Statistical software R, version 4.2.3; R Core Team.
was determined in chronic stroke patients with minimal to moder-
ate FMA-UE scores (28-50, inclusive),33 and even improvement
below the MCID could still be clinically significant, particularly if Keywords
it involves recovery of hand function. Conversely, those with
more severe deficits could require a larger impairment reduction BCI; BMI; Brain-computer interface; Brain-machine interface;
to be perceived as a meaningful gain in function.94,95 Thirdly, Neuroprosthesis; Neurorehabilitation; Rehabilitation; Stroke
there is also a lack of long-term data to understand if the gains
from such BCI neurorehabilitation can be maintained over time.
Also, as different studies measured outcomes at different time Corresponding author
points, interval censoring remains a problem.96 For instance,
Yu Tung Lo, MBBS, Department of Neurosurgery, National Neu-
patients in 1 study that reported outcome at an earlier time point
roscience Institute, 11 Jalan Tan Tock Seng, Level 3, Singapore
might have manifested as an earlier improvement compared to
308433 E-mail address: [email protected].
another study that reported the outcomes later (eg, due to longer
duration of BCI use), even though in the latter case, the gains
might have occurred before the conclusion of the BCI training
cycle. However, analyzing at the individual patient level reduced References
the effect of this scenario, as patients are compared within each 1. Feigin VL, Stark BA, Johnson CO, et al. Global, regional, and
study (ie, with the same measurement time points). national burden of stroke and its risk factors, 1990-2019: a systematic
analysis for the Global Burden of Disease Study 2019. Lancet Neurol
Future directions 2021;20:795–820.
2. Howle JM. Neuro-developmental treatment approach: theoretical
Despite recent advances in motor neuroprosthesis−based stroke foundations and principles of clinical practice. California, United
rehabilitation, there is still a sizable proportion of patient who had States: Neuro-Developmental Treatment Association; 2002.
no response to BCI-based treatment or who have significant resid- 3. Chang WH, Kim Y-H. Robot-assisted therapy in stroke rehabilita-
ual deficits after BCI training. Our results show that with the cur- tion. J Stroke 2013;15:174–81.
4. Kwakkel G, Veerbeek JM, van Wegen EEH, Wolf SL. Constraint-
rent BCI systems, although they can reduce impairment even in
induced movement therapy after stroke. Lancet Neurol 2015;14:
chronic stroke patients (many of whom have had functional recov- 224–34.
ery plateau with conventional therapy options), the impairment 5. Ambrosini E, Gasperini G, Zajc J, et al. A robotic system with EMG-
gain is still incomplete in many cases. Furthermore, combinatorial triggered functional electrical stimulation for restoring arm functions
effects with other rehabilitation therapy modalities remain poorly in stroke survivors. Neurorehabil Neural Repair 2021;35:334–45.
characterized. Current paradigms in BCI-driven neurorehabilita- 6. Ang KK, Guan C, Phua KS, et al. Brain-computer interface-based
tion may pave the way for more long-term BCI options,97,98 with robotic end effector system for wrist and hand rehabilitation: results
lessons learned from these studies potentially translating to next- of a three-armed randomized controlled trial for chronic stroke. Front
generation systems, including implantable ones that provide long- Neuroeng 2014;7:30.
term movement assistance. This scenario will also overcome the 7. Wen D, Fan Y, Hsu S-H, et al. Combining brain-computer interface
and virtual reality for rehabilitation in neurological diseases: a narra-
current hurdles in noninvasive BCI, which mostly involve wet
tive review. Ann Phys Rehabil Med 2021;64:101404.
EEG electrodes that are time intensive to apply and calibrate. 8. Miao Y, Chen S, Zhang X, et al. BCI-based rehabilitation on the
However, concerns such as surgical risks and implant longevity stroke in sequela stage. Neural Plast 2020;2020:8882764.
need to be addressed and balanced against potential benefits.99-101 9. Schalk G, McFarland DJ, Hinterberger T, Birbaumer N, Wolpaw JR.
A potential future strategy may involve patients who have BCI2000: a general-purpose brain-computer interface (BCI) system.
exhausted all rehabilitative treatment, including BCI-assisted IEEE Trans Biomed Eng 2004;51:1034–43.

www.archives-pmr.org
Motor BCI for stroke rehabilitation 2347

10. Tam W-K, Tong K, Meng F, Gao S. A minimal set of electrodes for 30. Ouzzani M, Hammady H, Fedorowicz Z, Elmagarmid A.
motor imagery BCI to control an assistive device in chronic stroke Rayyan-a web and mobile app for systematic reviews. Syst Rev
subjects: a multi-session study. IEEE Trans Neural Syst Rehabil Eng 2016;5:210.
2011;19:617–27. 31. Senn SJ. Overstating the evidence - double counting in meta-analysis
11. Grosse-Wentrup M, Mattia D, Oweiss K. Using brain-computer and related problems. BMC Med Res Methodol 2009;9:10.
interfaces to induce neural plasticity and restore function. J Neural 32. Woytowicz EJ, Rietschel JC, Goodman RN, et al. Determining levels
Eng 2011;8:025004. of upper extremity movement impairment by applying a cluster anal-
12. Soekadar SR, Birbaumer N, Slutzky MW, Cohen LG. Brain-machine ysis to the Fugl-Meyer Assessment of the upper extremity in chronic
interfaces in neurorehabilitation of stroke. Neurobiol Dis stroke. Arch Phys Med Rehabil 2017;98:456–62.
2015;83:172–9. 33. Page SJ, Fulk GD, Boyne P. Clinically important differences for the
13. Ang KK, Chua KSG, Phua KS, et al. A randomized controlled trial of upper-extremity fugl-meyer scale in people with minimal to moder-
EEG-based motor imagery brain-computer interface robotic rehabili- ate impairment due to chronic stroke. Phys Ther 2012;92:791–8.
tation for stroke. Clin EEG Neurosci 2015;46:310–20. 34. Messori A. Synthetizing published evidence on survival by recon-
14. Biasiucci A, Leeb R, Iturrate I, et al. Brain-actuated functional elec- struction of patient-level data and generation of a multi-trial Kaplan-
trical stimulation elicits lasting arm motor recovery after stroke. Nat Meier curve. Cureus 2021;13:e19422.
Commun 2018;9:2421. 35. Stewart LA, Clarke M, Rovers M, et al. Preferred reporting items for
15. Cheng N, Phua KS, Lai HS, et al. Brain-computer interface-based systematic review and meta-analyses of individual participant data:
soft robotic glove rehabilitation for stroke. IEEE Trans Biomed Eng the PRISMA-IPD statement. JAMA 2015;313:1657–65.
2020;67:3339–51. 36. de Jong VMT, Moons KGM, Riley RD, et al. Individual participant
16. Kasashima-Shindo Y, Fujiwara T, Ushiba J, et al. Brain-computer data meta-analysis of intervention studies with time-to-event out-
interface training combined with transcranial direct current stimula- comes: a review of the methodology and an applied example. Res
tion in patients with chronic severe hemiparesis: proof of concept Synth Methods 2020;11:148–68.
study. J Rehabil Med 2015;47:318–24. 37. National Heart, Lung, and Blood Institute. Study quality assessment
17. Bai Z, Fong KNK, Zhang JJ, Chan J, Ting KH. Immediate and long- tools. Available at: https://www.nhlbi.nih.gov/health-topics/study-
term effects of BCI-based rehabilitation of the upper extremity after quality-assessment-tools
stroke: a systematic review and meta-analysis. J NeuroEngineering 38. Egger M, Davey Smith G, Schneider M, Minder C. Bias in meta-
Rehabil 2020;17:57. analysis detected by a simple, graphical test. BMJ 1997;315:629–34.
18. Nojima I, Sugata H, Takeuchi H, Mima T. Brain-computer interface 39. Begg CB, Mazumdar M. Operating characteristics of a rank correla-
training based on brain activity can induce motor recovery in patients tion test for publication bias. Biometrics 1994;50:1088–101.
with stroke: a meta-analysis. Neurorehabil Neural Repair 40. Belardinelli P, Laer L, Ortiz E, Braun C, Gharabaghi A. Plasticity of
2022;36:83–96. premotor cortico-muscular coherence in severely impaired stroke
19. Xie Y-L, Yang Y-X, Jiang H, Duan X-Y, Gu L-J, Qing W, et al. patients with hand paralysis. NeuroImage Clin 2017;14:726–33.
Brain-machine interface-based training for improving upper extrem- 41. Bundy DT, Souders L, Baranyai K, et al. Contralesional brain−com-
ity function after stroke: a meta-analysis of randomized controlled puter interface control of a powered exoskeleton for motor recovery
trials. Front Neurosci 2022;16:949575. in chronic stroke survivors. Stroke 2017;48:1908–15.
20. Qu H, Zeng F, Tang Y, et al. The clinical effects of brain-computer 42. YuV Bushkova, GE Ivanova, Stakhovskaya LV, Frolov AA. Brain-
interface with robot on upper-limb function for post-stroke rehabili- computer-interface technology with multisensory feedback for con-
tation: a meta-analysis and systematic review. Disabil Rehabil Assist trolled ideomotor training in the rehabilitation of stroke patients.
Technol 2022: 1–12. BRSMU 2019: 27–32.
21. Colucci A, Vermehren M, Cavallo A, et al. Brain-computer interface 43. Caria A, Weber C, Br€otz D, et al. Chronic stroke recovery after com-
−controlled exoskeletons in clinical neurorehabilitation: ready or bined BCI training and physiotherapy: a case report. Psychophysiol-
not? Neurorehabil Neural Repair 2022;36:747–56. ogy 2011;48:578–82.
22. Khan MA, Das R, Iversen HK, Puthusserypady S. Review on motor 44. Carino-Escobar RI, Carrillo-Mora P, Valdes-Cristerna R, et al. Longi-
imagery based BCI systems for upper limb post-stroke neurorehabili- tudinal analysis of stroke patients’ brain rhythms during an interven-
tation: from designing to application. Comput Biol Med 2020; tion with a brain-computer interface. Neural Plast 2019;2019:1–11.
123:103843. 45. Chae J, Hart R. Intramuscular hand neuroprosthesis for chronic
23. Yang S, Li R, Li H, et al. Exploring the use of brain-computer interfaces stroke survivors. Neurorehabil Neural Repair 2003;17:109–17.
in stroke neurorehabilitation. Biomed Res Int 2021;2021:9967348. 46. Chen S, Cao L, Shu X, et al. Longitudinal electroencephalography
24. Robinson N, Mane R, Chouhan T, Guan C. Emerging trends in BCI- analysis in subacute stroke patients during intervention of brain-com-
robotics for motor control and rehabilitation. Curr Opinion Biomed puter interface with exoskeleton feedback. Front Neurosci
Eng 2021;20:100354. 2020;14:809.
25. Carvalho R, Dias N, Cerqueira JJ. Brain-machine interface of upper 47. Chowdhury A, Meena YK, Raza H, et al. Active physical practice
limb recovery in stroke patients rehabilitation: a systematic review. followed by mental practice using BCI-driven hand exoskeleton: a
Physiother Res Int 2019;24:e1764. pilot trial for clinical effectiveness and usability. IEEE J Biomed
26. H€ohler C, Trigili E, Astarita D, Hermsd€orfer J, Jahn K, Krewer C. Health Inform 2018;22:1786–95.
The efficacy of hybrid neuroprostheses in the rehabilitation of upper 48. Frolov AA, Mokienko O, Lyukmanov R, et al. Post-stroke rehabilita-
limb impairment after stroke, a narrative and systematic review with tion training with a motor-imagery-based brain-computer interface
a meta-analysis. Artif Organs 2024;48:232–53. (BCI)-controlled hand exoskeleton: a randomized controlled multi-
27. Mendes LA, Lima IN, Souza T, do Nascimento GC, Resqueti VR, center trial. Front Neurosci 2017;11:400.
Fregonezi GA. Motor neuroprosthesis for promoting recovery of 49. Gharabaghi A, Kraus D, LeA£o e MT, et al. Coupling brain-machine
function after stroke. Cochrane Database Syst Rev 2020;2020: interfaces with cortical stimulation for brain-state dependent stimula-
CD012991. tion: enhancing motor cortex excitability for neurorehabilitation.
28. Fugl-Meyer AR, Jaasko L, Leyman I, Olsson S, Steglind S. The post- Front Hum Neurosci 2014;8:122.
stroke hemiplegic patient. 1. a method for evaluation of physical per- 50. Kawakami M, Fujiwara T, Ushiba J, et al. A new therapeutic applica-
formance. Scand J Rehabil Med 1975;7:13–31. tion of brain-machine interface (BMI) training followed by hybrid
29. Lyle RC. A performance test for assessment of upper limb function assistive neuromuscular dynamic stimulation (HANDS) therapy for
in physical rehabilitation treatment and research. Int J Rehabil Res patients with severe hemiparetic stroke: a proof of concept study.
1981;4:483–92. RNN 2016;34:789–97.

www.archives-pmr.org
2348 Y.T. Lo et al

51. Kawakami M, Okuyama K, Takahashi Y, et al. Change in reciprocal 69. Zhang X, Elnady AM, Randhawa BK, Boyd LA, Menon C. Combin-
inhibition of the forearm with motor imagery among patients with ing mental training and physical training with goal-oriented protocols
chronic stroke. Neural Plast 2018;2018:1–9. in stroke rehabilitation: a feasibility case study. Front Hum Neurosci
52. Kim T, Kim S, Lee B. Effects of action observational training plus 2018;12:125.
brain-computer interface−based functional electrical stimulation on 70. Curado MR, Cossio EG, Broetz D, et al. Residual upper arm motor
paretic arm motor recovery in patient with stroke: a randomized con- function primes innervation of paretic forearm muscles in chronic
trolled trial: effects of AOT plus BCI-FES on arm motor recovery. stroke after brain-machine interface (BMI) training. PLoS ONE
Occup Ther Int 2016;23:39–47. 2015;10:e0140161.
53. Kotov SV, Turbina LG, Kondur AA, Zaytseva EV, Biryukova EV. 71. Iba~nez J, Monge-Pereira E, Molina-Rueda F, et al. Low latency esti-
Recovery dynamics in patients with poststroke motor disorders after mation of motor intentions to assist reaching movements along multi-
multiple courses of neurorehabilitation using an exoskeleton con- ple sessions in chronic stroke patients: a feasibility study. Front
trolled by a brain−computer interface. Neurosci Behav Phys Neurosci 2017;11:126.
2018;48:1088–92. 72. Irastorza-Landa N, Garcıa-Cossio E, Sarasola-Sanz A, Br€otz D,
54. Li M, Liu Y, Wu Y, Liu S, Jia J, Zhang L. Neurophysiological sub- Birbaumer N, Ramos-Murguialday A. Functional synergy
strates of stroke patients with motor imagery-based brain-computer recruitment index as a reliable biomarker of motor function and
interface training. Int J Neurosci 2014;124:403–15. recovery in chronic stroke patients. J Neural Eng 2021;
55. Matarasso AK, Rieke JD, White K, Yusufali MM, Daly JJ. Com- 18:046061.
bined real-time fMRI and real time fNIRS brain computer interface 73. Jovanovic LI, Kapadia N, Lo L, Zivanovic V, Popovic MR, Mar-
(BCI): training of volitional wrist extension after stroke, a case series quez-Chin C. Restoration of upper limb function after chronic severe
pilot study. PLoS ONE 2021;16:e0250431. hemiplegia: a case report on the feasibility of a brain-computer inter-
56. Mukaino M, Ono T, Shindo K, et al. Efficacy of brain-computer face-triggered functional electrical stimulation therapy. Am J Phys
interface-driven neuromuscular electrical stimulation for chronic Med Rehabil 2020;99:e35–40.
paresis after stroke. J Rehabil Med 2014;46:378–82. 74. Knutson JS, Chae J, Hart RL, et al. Implanted neuroprosthesis for
57. Nam C, Rong W, Li W, Xie Y, Hu X, Zheng Y. The effects of upper- assisting arm and hand function after stroke: a case study. JRRD
limb training assisted with an electromyography-driven neuromuscu- 2012;49:1505.
lar electrical stimulation robotic hand on chronic stroke. Front Neu- 75. Zhao C-G, Ju F, Sun W, et al. Effects of training with a brain−com-
rol 2017;8:679. puter interface-controlled robot on rehabilitation outcome in patients
58. Nishimoto A, Kawakami M, Fujiwara T, et al. Feasibility of task- with subacute stroke: a randomized controlled trial. Neurol Ther
specific brain-machine interface training for upper-extremity paraly- 2022;11:679–95.
sis in patients with chronic hemiparetic stroke. J Rehabil Med 76. McCrimmon CM, King CE, Wang PT, Cramer SC, Nenadic Z, Do
2018;50:52–8. AH. Brain-controlled functional electrical stimulation therapy for
59. Rathee D, Chowdhury A, Meena YK, Dutta A, McDonough S, Pra- gait rehabilitation after stroke: a safety study. J NeuroEngineering
sad G. Brain−machine interface-driven post-stroke upper-limb func- Rehabil 2015;12:57.
tional recovery correlates with beta-band mediated cortical 77. Rabadi MH, Rabadi FM. Comparison of the action research arm
networks. IEEE Trans Neural Syst Rehabil Eng 2019;27:1020–31. test and the Fugl-Meyer assessment as measures of upper-
60. Remsik AB, Williams L, Gjini K, et al. Ipsilesional mu rhythm extremity motor weakness after stroke. Arch Phys Med Rehabil
desynchronization and changes in motor behavior following post 2006;87:962–6.
stroke BCI intervention for motor rehabilitation. Front Neurosci 78. Taub E, Uswatte G, Mark VW, Morris DMM. The learned nonuse
2019;13:53. phenomenon: implications for rehabilitation. Eura Medicophys
61. Sebastian-Romagosa M, Cho W, Ortner R, et al. Brain computer 2006;42:241–56.
interface treatment for motor rehabilitation of upper extremity of 79. Coscia M, Wessel MJ, Chaudary U, et al. Neurotechnology-aided
stroke patients—a feasibility study. Front Neurosci 2020;14:591435. interventions for upper limb motor rehabilitation in severe chronic
62. Tabernig CB, Lopez CA, Carrere LC, Spaich EG, Ballario CH. Neu- stroke. Brain 2019;142:2182–97.
rorehabilitation therapy of patients with severe stroke based on func- 80. Ward NS, Brander F, Kelly K. Intensive upper limb neurorehabilita-
tional electrical stimulation commanded by a brain computer tion in chronic stroke: outcomes from the Queen Square programme.
interface. J Rehabil Assist Technol Eng 2018;5:205566831878928. J Neurol Neurosurg Psychiatry 2019;90:498–506.
63. Tsuchimoto S, Shindo K, Hotta F, Hanakawa T, Liu M, Ushiba J. 81. Ma C, Chen N, Mao Y, Huang D, Song R, Li L. Alterations of mus-
Sensorimotor connectivity after motor exercise with neurofeedback cle activation pattern in stroke survivors during obstacle crossing.
in post-stroke patients with hemiplegia. Neuroscience Front Neurol 2017;8:70.
2019;416:109–25. 82. Byblow WD, Stinear CM, Barber PA, Petoe MA, Ackerley SJ. Pro-
64. Wu Q, Ge Y, Ma D, et al. Analysis of prognostic risk factors deter- portional recovery after stroke depends on corticomotor integrity.
mining poor functional recovery after comprehensive rehabilitation Ann Neurol 2015;78:848–59.
including motor-imagery brain-computer interface training in stroke 83. Agius Anastasi A, Falzon O, Camilleri K, Vella M, Muscat R. Brain
patients: a prospective study. Front Neurol 2021;12:661816. symmetry index in healthy and stroke patients for assessment and
65. Yuan K, Chen C, Wang X, Chu WC, Tong RK. BCI training effects prognosis. Stroke Res Treat 2017;2017:8276136.
on chronic stroke correlate with functional reorganization in motor- 84. Liepert J, Bauder H, Wolfgang HR, Miltner WH, Taub E, Weiller C.
related regions: a concurrent EEG and fMRI study. Brain Sci Treatment-induced cortical reorganization after stroke in humans.
2021;11:56. Stroke 2000;31:1210–6.
66. Bhagat NA, Yozbatiran N, Sullivan JL, et al. Neural activity modula- 85. Carey JR, Kimberley TJ, Lewis SM, et al. Analysis of fMRI and fin-
tions and motor recovery following brain-exoskeleton interface ger tracking training in subjects with chronic stroke. Brain
mediated stroke rehabilitation. NeuroImage Clin 2020;28:102502. 2002;125:773–88.
67. Chew E, Teo W-P, Tang N, et al. Using transcranial direct current 86. Johansen-Berg H, Dawes H, Guy C, Smith SM, Wade DT, Mat-
stimulation to augment the effect of motor imagery-assisted brain- thews PM. Correlation between motor improvements and altered
computer interface training in chronic stroke patients—cortical reor- fMRI activity after rehabilitative therapy. Brain 2002;125:2731–
ganization considerations. Front Neurol 2020;11:948. 42.
68. Marquez-Chin C, Marquis A, Popovic MR. BCI-triggered functional 87. Ang KK, Guan C, Chua KSG, et al. A large clinical study on the abil-
electrical stimulation therapy for upper limb. Eur J Transl Myol ity of stroke patients to use an EEG-based motor imagery brain-com-
2016;26:6222. puter interface. Clin EEG Neurosci 2011;42:253–8.

www.archives-pmr.org
Motor BCI for stroke rehabilitation 2349

88. Yilmaz O, Birbaumer N, Ramos-Murguialday A. Movement related convalescent stroke patients with moderate to severe hemiparesis. J
slow cortical potentials in severely paralyzed chronic stroke patients. Phys Ther Sci 2019;31:917–21.
Front Hum Neurosci 2014;8:1033. 95. Arya KN, Verma R, Garg RK. Estimating the minimal clinically
89. Pundik S, McCabe JP, Hrovat K, et al. Recovery of post stroke proxi- important difference of an upper extremity recovery measure in
mal arm function, driven by complex neuroplastic bilateral brain subacute stroke patients. Top Stroke Rehabil 2011;18(Suppl 1):
activation patterns and predicted by baseline motor dysfunction 599–610.
severity. Front Hum Neurosci 2015;9:394. 96. Zhang Z, Sun J. Interval censoring. Stat Methods Med Res
90. Buch ER, Modir Shanechi A, Fourkas AD, Weber C, Birbaumer N, 2010;19:53–70.
Cohen LG. Parietofrontal integrity determines neural modulation 97. Mitchell P, Lee SCM, Yoo PE, et al. Assessment of safety of a fully
associated with grasping imagery after stroke. Brain 2012;135: implanted endovascular brain-computer interface for severe paralysis
596–614. in 4 patients: the Stentrode With Thought-Controlled Digital Switch
91. Carter AR, Patel KR, Astafiev SV, et al. Upstream dysfunction of (SWITCH) study. JAMA Neurol 2023;80:270–8.
somatomotor functional connectivity after corticospinal damage in 98. Rastogi A, Willett FR, Abreu J, et al. The neural representation of
stroke. Neurorehabil Neural Repair 2012;26:7–19. force across grasp types in motor cortex of humans with tetraplegia.
92. Ma Z-Z, Wu J-J, Hua X-Y, et al. Evidence of neuroplasticity with eNeuro 2021;8. ENEURO.0231-20.2020.
brain-computer interface in a randomized trial for post-stroke reha- 99. Rubin DB, Ajiboye AB, Barefoot L, et al. Interim safety profile from
bilitation: a graph-theoretic study of subnetwork analysis. Front Neu- the feasibility study of the braingate neural interface system. Neurol-
rol 2023;14:1135466. ogy 2023;100:e1177–92.
93. Mansour S, Giles J, Ang KK, Nair KPS, Phua KS, Arvaneh M. 100. Normann RA, Fernandez E. Clinical applications of penetrating neu-
Exploring the ability of stroke survivors in using the contralesional ral interfaces and Utah Electrode Array technologies. J Neural Eng
hemisphere to control a brain-computer interface. Sci Rep 2016;13:061003.
2022;12:16223. 101. Sponheim C, Papadourakis V, Collinger JL, et al. Longevity and reli-
94. Hiragami S, Inoue Y, Harada K. Minimal clinically important differ- ability of chronic unit recordings using the Utah, intracortical multi-
ence for the Fugl-Meyer assessment of the upper extremity in electrode arrays. J Neural Eng 2021;18. 10.1088/1741-2552.

www.archives-pmr.org