학술

Motor imagery-based brain-computer interface training for upper-limb rehabilitation after subacute stroke: A randomized sham-controlled trial

PLOS ONE
CC BY
이 매체는 공공·자유 라이선스로 본문을 직접 표시합니다.

Figures
Abstract
Objective
To determine whether motor imagery (MI) based brain-computer interface (BCI) training improves upper-limb motor function in patients with subacute stroke compared with a high-fidelity sham-BCI control, and to explore accompanying structural and resting-state functional MRI changes.
Method
In this participant-blinded randomized controlled trial, 32 patients with subacute stroke were randomized to a MI-based BCI group or a high-fidelity BCI sham-control group (n = 16 each). Both groups received dose-matched training (30 min/session, 5 sessions/week for 4 weeks) in addition to standard rehabilitation. Upper-limb motor outcomes were assessed using the Fugl-Meyer Assessment-Upper Extremity (FMA-UE; primary outcome), Wolf Motor Function Test (WMFT), Motor Assessment Scale (MAS), and Modified Barthel Index (MBI). Structural MRI and resting-state functional MRI were acquired at baseline and post-intervention to assess gray matter volume (GMV) and functional connectivity (FC).
Results
Both groups showed significant improvement after the intervention. Compared with the sham-control group, the MI-based BCI group demonstrated greater gains in FMA-UE and secondary motor outcomes (WMFT and MAS), whereas the difference in MBI was not statistically significant. Voxel-based morphometry revealed a within-group increase in GMV in the contralesional supplementary motor area (SMA) in the MI-based BCI group; however, no significant between-group GMV difference was observed. The MI-based BCI group also exhibited increased resting-state FC between the SMA and the ipsilesional cerebellar lobule IX and fusiform gyrus. Changes in FC were associated with improvements in FMA-UE, WMFT, and MBI, but not MAS.
Conclusion
Compared with a high-fidelity sham-control group, MI-based BCI group training was associated with greater upper-limb motor recovery in patients with subacute stroke, accompanied by changes in structural and functional neuroimaging measures involving motor-related regions and connections. Larger randomized controlled trials with longer follow-up are required to confirm durability and clinical relevance.
Citation: Tong Q, Wu J, Fan P, Cai S, Wang M, Gong J, et al. (2026) Motor imagery-based brain-computer interface training for upper-limb rehabilitation after subacute stroke: A randomized sham-controlled trial. PLoS One 21(11): e0363060. https://doi.org/10.1371/journal.pone.0363060
Editor: Wang Zhan, University of Maryland at College Park, UNITED STATES OF AMERICA
Received: February 3, 2026; Accepted: November 2, 2026; Published: November 20, 2026
Copyright: © 2026 Tong et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All de-identified participant-level data underlying the clinical findings reported in this paper are provided in the Supporting information files.
Funding: This research was supported by the National Natural Science Foundation of China (Grant no. 62373326). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Abbreviations: ACC, anterior cingulate cortex; BCI, brain–computer interface; EEG, electroencephalogram; FC, functional connectivity; FFG, fusiform gyrus; FMA-UE, Fugl-Meyer Assessment-Upper Extremity; GMV, gray matter volume; MAS, Motor Assessment Scale; MBI, Modified Barthel Index; MCID, minimal clinically important difference; MI, motor imagery; rs-fMRI, resting-state functional magnetic resonance imaging; SMA, supplementary motor area; sMRI, structural magnetic resonance imaging; VBM, voxel-based morphometry; WMFT, Wolf Motor Function Test
1. Introduction
Stroke is a leading cause of long-term disability worldwide and frequently results in persistent upper-limb motor impairment, substantially limiting independence, daily participation, and quality of life [1–3]. Restoration of upper-limb function is therefore a key target of post-stroke rehabilitation. Rehabilitation initiated during the subacute phase is especially relevant, as this period is associated with an enhanced capacity for experience-dependent neural reorganization and may represent a critical window for motor relearning. Consequently, rehabilitation strategies that actively engage patients in goal-directed training during this stage are likely to play an important role in functional recovery.
Brain-computer interface (BCI)-based rehabilitation has emerged as an approach to enhance active patient engagement by linking movement intention to task-relevant sensory feedback. By translating volitional neural activity into control signals for external devices and providing real-time sensorimotor feedback, BCI systems create a closed-loop interaction designed to support motor relearning [4–6]. In noninvasive motor imagery (MI)-based BCI paradigms, neural correlates of imagined movement are decoded to drive assistive technologies such as robotic devices or functional electrical stimulation [7]. This intention-driven feedback mechanism may facilitate upper-limb recovery, particularly in patients with limited voluntary movement.
Accumulating evidence from randomized controlled trials and systematic reviews suggests that BCI-based interventions are associated with improvements in upper-limb motor outcomes after stroke, particularly when combined with conventional rehabilitation [8,9]. Neurophysiological and neuroimaging studies have also reported BCI-related changes in cortical excitability and functional network organization, findings that are consistent with experience-dependent neural reorganization [10–13]. However, treatment effects vary substantially across studies, reflecting heterogeneity in patient populations, training paradigms, outcome measures, and study quality.
Importantly, many prior investigations have focused on chronic stroke populations or have used designs that do not fully control for nonspecific influences, including attention, expectancy, device interaction, and passive movement [6,8]. Consequently, rigorously designed sham-controlled trials, particularly in subacute stroke using MI-based BCI paradigms, remain limited, and the specific contribution of neural contingency in active BCI training to observed clinical and neural changes has yet to be clearly established.
To address these limitations, we conducted a randomized, sham-controlled trial to examine the effects of MI-based BCI training on upper-limb motor recovery in patients with subacute stroke. The sham intervention was matched to the active BCI training in device setup, training duration, task instructions, and sensory experience, while withholding the neural contingency between MI-related brain activity and robotic movement to minimize nonspecific influences related to attention, expectancy, and passive movement. In addition to standardized clinical assessments, we acquired structural MRI (sMRI) and resting-state functional MRI (rs-fMRI) at baseline and after the intervention to characterize longitudinal changes in gray matter volume (GMV) and functional connectivity (FC). By integrating clinical and multimodal neuroimaging outcomes, this study sought to determine whether MI-based BCI training during the subacute phase is associated with greater upper-limb recovery and specific neural adaptations beyond nonspecific training effects.
2. Materials and methods
2.2 Study design, setting, and reporting
This study was designed as a single-center, randomized, sham-controlled, parallel-group clinical trial. The trial was conducted at Zhejiang Provincial People’s Hospital between December 2019 and December 2020. The study was reported in accordance with the Consolidated Standards of Reporting Trials (CONSORT) guidelines for randomized controlled trials. A CONSORT flow diagram illustrated in Fig 1.
2.3 Participants
2.3.1 Recruitment and eligibility.
Patients with subacute stroke who were admitted to the Department of Rehabilitation of Zhejiang Provincial People’s Hospital were screened for eligibility between December 2019 and December 2020. A total of 32 eligible participants were enrolled for this randomized trial and randomly assigned to either the MI-based BCI group or the sham-control group.
Exclusion criteria included: (1) unstable vital signs, such as uncontrolled hypertension, arrhythmia, or respiratory instability; (2) severe cognitive impairment (Mini-Mental State Examination score ≤21); (3) severe speech, visual, or hearing impairments, or aphasia interfering with comprehension or task execution; (4) bilateral stroke lesions; (5) history of epilepsy; (6) severe psychiatric disorders affecting cooperation; and (7) pre-existing upper-limb functional impairment prior to stroke.
2.4 Randomization, allocation concealment, and blinding
Participants were randomly assigned in a 1:1 ratio to either the MI-based BCI group or the sham-control group using a computer-generated random sequence. The randomization sequence was generated by an independent researcher who was not involved in participant recruitment, intervention delivery, or outcome assessment. Allocation concealment was ensured using sealed, opaque envelopes, which were opened only after participant enrollment.
To minimize assessment bias, outcome assessors were blinded to group allocation throughout the study. Participants were informed that they would receive one of two forms of MI-based BCI training but were not informed of the specific features distinguishing the active and sham interventions. Owing to the nature of the intervention, therapists delivering the training could not be blinded to group assignment; however, they were not involved in outcome assessment or data analysis.
Participant blinding was further supported by the use of identical training duration, task instructions, and user interfaces across groups. These procedures were implemented to reduce potential bias related to participant expectations and to enhance the validity of between-group comparisons.
2.5 Interventions
All participants received an additional intervention session for four consecutive weeks, in addition to standard rehabilitation therapy. The duration, frequency, and total number of intervention sessions were identical for the MI-based BCI group and the sham-control group. All intervention sessions were conducted in a dedicated rehabilitation setting within the hospital.
2.5.1 Standard rehabilitation.
In addition to the assigned BCI or sham intervention, all participants in both groups received standard rehabilitation in accordance with routine clinical practice at the study center. Standard rehabilitation primarily consisted of physical therapy and occupational therapy targeting upper-limb motor function, including range-of-motion exercises, task-oriented training, and functional activities related to activities of daily living.
The content, frequency, and duration of standard rehabilitation were comparable between the two groups throughout the intervention period. Therapists delivering standard rehabilitation were not involved in the BCI or sham training sessions, and the standard rehabilitation program was maintained consistently to minimize potential confounding effects from background therapy.
2.5.2 MI-based BCI intervention.
Participants assigned to the MI-based BCI group received BCI intervention targeting upper-limb rehabilitation. The intervention was delivered using BCI system (L-B300; Zhejiang Mailian Medical Technology Co., Ltd., Hangzhou, China), which integrates an electroencephalography (EEG) based MI paradigm with an upper-limb robotic orthosis and task-oriented training software. During each training session, participants were instructed to observe a task-oriented virtual swimming game and to perform MI of upper-limb swimming movements (Fig 2).
(A) Components of the BCI system (L-B300), including the EEG acquisition unit, signal-processing module, active-passive robotic arm, and evaluation software. (B) Electrode preparation and EEG cap placement following the International 10-20 System. (C) Signal-processing pipeline comprising EEG acquisition, preprocessing, feature extraction, pattern classification, and control output.
EEG signals were continuously acquired and analyzed in real time to quantify MI engagement, generating a MI-related control score that governed system output. When the MI score exceeded a predefined, individually adjusted threshold, the virtual game animation and the robotic orthosis were activated, with movement speed scaled proportionally to the MI score, thereby driving repetitive alternating upper-limb movements. If the MI score fell below the threshold, an auditory cue (“Please focus”) was presented to prompt attention. Sustained low MI engagement (lasting more than 5 seconds) resulted in gradual cessation of visual and robotic feedback until adequate MI activity was re-established.
Training difficulty was adjusted individually using predefined difficulty levels requiring progressively stronger MI engagement to maintain task challenge and participant involvement. To ensure participant safety, real-time resistance monitoring was implemented throughout all sessions to detect excessive human-machine interaction forces. When resistance exceeded predefined safety thresholds, the system automatically triggered an emergency stop, and a manual emergency stop button was available at all times. Each participant completed 20 training sessions in total, with each session lasting 30 minutes, five days per week for four consecutive weeks.
2.5.3 Sham-control intervention.
Participants assigned to the sham-control group received a sham BCI intervention designed to closely match the active MI-based BCI training in all aspects except for neural contingency. The sham intervention used the same BCI system, robotic orthosis, training interface, session duration, task instructions, and sensory feedback as the active intervention, and participants were instructed to perform MI tasks in the same manner.
In the sham condition, the Bluetooth connection between the EEG acquisition system and the host computer was disabled, thereby preventing real-time acquisition and decoding of neural signals. As a result, movements of the robotic orthosis and progression of the virtual task were not contingent on participants’ MI-related brain activity. Instead, the robotic device executed preprogrammed alternating upper-limb movements within a predefined velocity range comparable to that of the active BCI training.
Visual feedback, task presentation, and auditory cues were identical to those provided in the MI-based BCI group, including attention prompts. This sham design was implemented to preserve participant engagement and expectancy while withholding the neural contingency characteristic of MI-based BCI training, thereby minimizing nonspecific effects related to attention, device use, and passive movement.
2.6 Outcomes and assessment schedule
The primary clinical outcome of this study was upper-limb motor function, assessed using the Fugl-Meyer Assessment-Upper Extremity (FMA-UE). Prespecified secondary clinical outcomes included the Wolf Motor Function Test (WMFT), the Motor Assessment Scale (MAS), and the Modified Barthel Index (MBI), which together provided complementary measures of motor performance and functional independence.
All clinical outcome assessments were conducted at baseline prior to the intervention and immediately after completion of the 4-week training period. Assessments were performed by trained evaluators who were blinded to group allocation.
Neuroimaging outcomes were included as supportive measures to characterize intervention-related structural and functional brain changes. sMRI and rs-fMRI were acquired at the same two time points (baseline and post-intervention) to assess longitudinal changes in GMV and FC.
2.7 MRI acquisition, preprocessing, and analysis
2.7.1 MRI acquisition.
MRI data were acquired using a 3.0-T MRI scanner (Discovery MR750w, GE Healthcare, Milwaukee, WI, USA). Each participant underwent two MRI sessions: one at baseline, and one immediately after completion of the 4-week training period. sMRI data were obtained using a T1-weighted spoiled gradient recalled echo sequence (repetition time = 5.7 ms, echo time = 2.5 ms, flip angle = 9°, field of view = 240 × 240 mm², matrix size = 256 × 256, slice thickness = 1.0 mm, 160 slices). rs-fMRI data were acquired using a gradient-echo echo-planar imaging sequence (repetition time = 2000 ms, echo time = 30 ms, flip angle = 90°, field of view = 220 × 220 mm², matrix size = 64 × 64, slice thickness = 4 mm with a 0.8-mm interslice gap). During rs-fMRI acquisition, participants were instructed to remain still, keep their eyes closed, stay awake, and not engage in any specific cognitive task.
2.7.2 sMRI preprocessing and GMV analysis.
sMRI data were processed using the Computational Anatomy Toolbox implemented in SPM12 (Wellcome Centre for Human Neuroimaging, London, UK) running on MATLAB. Prior to preprocessing, all images underwent visual quality inspection to identify artifacts or excessive motion. For participants with right-hemisphere lesions, images were flipped along the midsagittal plane to align the lesioned hemisphere across subjects, facilitating group-level comparisons in accordance with common practice in stroke neuroimaging studies.
Stroke lesions were manually delineated on native T1-weighted images using MRIcron and ITK-SNAP, and binary lesion masks were generated for each participant. Lesion-masked T1 images were then created and used for subsequent preprocessing. Structural images were segmented into gray matter, white matter, and cerebrospinal fluid, spatially normalized to the Montreal Neurological Institute space using geodesic shooting registration, modulated to preserve volume information, and smoothed with a Gaussian kernel to reduce interindividual anatomical variability. Voxel-based morphometry (VBM) was performed to assess longitudinal changes in GMV following the intervention.
2.7.3 rs-fMRI preprocessing and FC analysis.
rs-fMRI data were preprocessed using the Data Processing & Analysis for Brain Imaging toolbox. Preprocessing steps included slice-timing correction, realignment for head-motion correction, co-registration to individual structural images, normalization to Montreal Neurological Institute space, spatial smoothing, temporal band-pass filtering, and regression of motion-related and physiological nuisance covariates.
To maintain consistency with the sMRI analysis, images from participants with right-hemisphere lesions were flipped along the midsagittal plane so that the lesioned hemisphere was consistently aligned across subjects. FC analyses were conducted to characterize intervention-related changes in resting-state brain networks (more details see S1 File).
3. Statistical analysis
Statistical analyses were conducted using SPSS software (IBM Corp., Armonk, NY, USA). Continuous variables are presented as mean ± SD or median (IQR), and categorical variables as counts (percentages). Baseline characteristics were compared between groups using independent-samples t-tests or Mann-Whitney U tests for continuous variables and chi-square tests for categorical variables. The primary clinical outcome was the FMA-UE score, and secondary clinical outcomes included the WMFT, MAS, and MBI. To estimate the intervention effect while accounting for baseline values, post-intervention scores were analyzed using baseline-adjusted linear regression models (analysis of covariance, ANCOVA). For each outcome, the post-intervention score was entered as the dependent variable, with treatment group and the corresponding baseline score included as independent variables. The treatment-group coefficient was interpreted as the adjusted between-group difference at post-intervention and is reported with its 95% confidence interval (CI) and p value. Standardized between-group effect sizes were additionally calculated to characterize the magnitude of the observed effects. All statistical tests were two-sided, with p < 0.05 considered statistically significant. Neuroimaging analyses employed VBM and FC to evaluate between-group differences in longitudinal changes from baseline to post-intervention. Multiple comparisons were corrected using voxel- or cluster-level methods. Neuroimaging results were interpreted in the context of clinical outcomes. All tests were two-tailed, with statistical significance set at p < 0.05.
3.1 Missing data and analysis population
All randomized participants were included in the intention-to-treat analysis. Missing post-intervention clinical outcomes for one sham-control participant were handled using multiple imputation with 50 imputed datasets. Baseline-adjusted ANCOVA was performed in each dataset, and estimates were pooled using Rubin’s rules. Complete-case ANCOVA was conducted as a sensitivity analysis. Neuroimaging analyses included participants with usable MRI data at both time points (n = 30; 15 per group).
3.2 Sample size calculation
The minimum sample size was determined a priori using PASS 2021 software (NCSS, LLC, Kaysville, UT, USA). Based on previous research in patients with stroke and our clinical experience [14], we assumed that the change in FMA score in the BCI group would reach the minimum clinically important difference of 4 points [15], whereas the expected change in the conventional training group was 1 point, with a common standard deviation of 2.3 points. With a two-sided α of 0.05 and 90% statistical power, the minimum required sample size was 28 participants. Considering an anticipated dropout rate of approximately 15%, we planned to enroll 32 participants, with 16 participants in each group. The prespecified assumptions corresponded to an anticipated standardized effect size of Cohen’s d = 1.30. In the observed data, the standardized between-group effect for change in FMA-UE was Cohen’s d = 0.87, corresponding to an exploratory retrospective power of approximately 65% based on the available complete outcome data.
4. Result
4.1 Participant flow and analysis sets
A total of 32 participants were randomized. One participant in the sham-control group discontinued the intervention because of worsening shoulder pain and did not complete the post-intervention clinical or MRI assessments. No serious adverse events occurred, and no other adverse events leading to treatment discontinuation were observed. The causal relationship between the worsening shoulder pain and the study intervention could not be established. All 32 participants were retained in the intention-to-treat clinical analysis, with missing clinical outcomes handled using multiple imputation. Complete-case clinical analyses included 31 participants. Neuroimaging analyses included 30 participants (15 per group).
4.2 Baseline characteristics
Baseline demographic, clinical, and lesion-related characteristics were comparable between the MI-based BCI group and the sham-control group. No statistically significant between-group differences were observed in age, sex, time since stroke onset, lesion side, lesion volume, stroke pathology, or baseline upper-limb motor function (Table 1 and S1 Fig).
4.3 Clinical outcomes
4.3.1 Primary clinical outcome.
Baseline FMA-UE scores did not differ significantly between the two groups. Changes in FMA-UE scores from baseline to post-intervention (ΔFMA-UE) were therefore used as the primary endpoint for between-group comparisons. Analysis of change scores indicated significantly greater improvement in upper-limb motor function in the MI-based BCI group compared with the sham-control group following the intervention (t = 2.479, p = 0.021; Table 2).
4.3.2 Secondary clinical outcomes.
Baseline scores on the WMFT, MAS, and MBI were comparable between the two groups. Between-group comparisons of change scores (Δ) revealed significantly greater improvements in the MI-based BCI group than in the sham-control group for WMFT (z = −3.419, p < 0.001) and MAS (z = −2.222, p = 0.028) following the intervention, but not for changes in MBI (z = −2.082, p = 0.052; Table 2).
4.4 sMRI outcomes
VBM analyses were conducted to examine longitudinal changes in GMV following the intervention. Between-group comparisons did not reveal significant differences in GMV change from baseline to post-intervention. Exploratory within-group analyses demonstrated a significant post-intervention increase in GMV in the MI-based BCI group, identified as a large cluster (6,439 voxels) with a peak located in the right supplementary motor area (SMA) and a subpeak in the right anterior cingulate cortex (ACC) (Figs 3 and S2 and S2 Table). No significant GMV changes were observed within the sham-control group. These within-group findings were considered exploratory and were further examined in relation to FC and clinical outcomes.
VBM revealed a significant GMV increase in the MI-based BCI group following the intervention (6,439-voxel cluster), with the peak located in the right SMA and a subpeak in the right ACC. Activation intensity is displayed on the cortical surface, with the color bar indicating t-values.
4.5 rs-fMRI outcomes
rs-fMRI based FC analyses were performed to assess longitudinal changes in functional brain networks from baseline to post-intervention. Between-group comparisons of FC changes from baseline to post-intervention revealed significantly increase in connectivity involving the right SMA in the MI-based BCI group. Specifically, compared with the sham-control group, the MI-based BCI group exhibited significantly increase in FC between the right SMA and two regions in the ipsilesional hemisphere, as defined after lesion-side normalization: the cerebellar lobule IX and the fusiform gyrus (FFG) (cluster-level p < 0.05, FDR-corrected; peak p < 0.001, uncorrected; Fig 4). No significant between-group differences in FC changes were observed for the remaining seed regions.
(a) Compared with the sham-control group, the MI-based BCI group showed a significantly greater increase in resting-state FC between the right SMA and the ipsilesional cerebellar lobule IX. (b) The MI-based BCI group also exhibited a significantly greater increase in FC between the right SMA and the ipsilesional FFG. Results are displayed at a cluster-level threshold of p < 0.05, false discovery rate (FDR)–corrected, with a peak-level threshold of p < 0.001 (uncorrected). Images are shown in neurological convention, with lesion-side normalization applied.
4.6 Associations between neuroimaging and clinical outcomes
Exploratory correlation analyses were conducted to examine associations between intervention-related changes in neuroimaging data and changes in clinical outcomes.
Changes in FC between the right SMA and the ipsilesional cerebellar lobule IX were associated with improvements in upper-limb motor function, as reflected by changes in the FMA-UE (r = 0.560, p = 0.030), the WMFT (r = 0.597, p = 0.020), and the MBI (r = 0.516, p = 0.049). No significant association was observed with changes in the MAS (S3 Fig).
Similarly, increases in FC between the right SMA and the ipsilesional FFG were associated with improvements in the FMA-UE (r = 0.658, p = 0.008), the WMFT (r = 0.609, p = 0.020), and the MBI (r = 0.807, p < 0.001) (S4 Fig).
In addition, exploratory analyses indicated that changes in FC between the right SMA and these two regions were positively associated with increases in GMV within the right SMA (cerebellar lobule IX: r = 0.524, p = 0.040; fusiform gyrus: r = 0.701, p = 0.004; Fig 5).
(a) Changes in FC between the right SMA and the ipsilesional cerebellar lobule IX were positively correlated with changes in GMV within the right SMA. (b) Changes in FC between the right SMA and the ipsilesional FFG were also positively correlated with changes in right SMA GMV. Correlation coefficients (r) and p values were derived from Spearman rank correlation analyses. Solid lines represent linear regression fits, and dotted lines indicate 95% confidence intervals. These analyses were exploratory in nature.
5. Discussion
In this randomized, sham-controlled trial, we examined whether MI-based BCI training improves upper-limb motor recovery in patients with subacute stroke. Compared with a closely matched sham intervention, the MI-based BCI group showed greater improvement in the prespecified primary endpoint (change in FMA-UE score) and a consistent direction of benefit across secondary motor outcomes. Neuroimaging measures were included to characterize accompanying structural and functional brain changes under a rigorously controlled intervention paradigm; these outcomes were not intended to serve as primary efficacy endpoints.
The clinical relevance of these findings is supported by evaluation against thresholds of meaningful change for commonly used upper-limb measures. For the FMA-UE, clinically important improvement has been reported to fall in the range of approximately 4–7 points in chronic stroke cohorts, with larger values reported in more impaired or convalescent populations, underscoring the context dependence of minimal clinically important difference (MCID) estimates [15]. Consistent with this, both groups in the present study demonstrated clinically relevant improvement over the intervention period, reflecting the expected recovery trajectory in the subacute phase under active background rehabilitation. Importantly, however, participants receiving MI-based BCI training achieved significantly greater gains in upper-limb motor outcomes than those receiving the matched sham intervention, including superior change scores on FMA-UE, WMFT, and MAS. Because the sham condition closely matched device exposure, task instructions, and sensory experience while withholding neural contingency, the observed between-group differences are consistent with a potential additional benefit of contingency-based training beyond nonspecific training effects [16].
The present findings are broadly consistent with randomized trials and meta-analyses showing that BCI-assisted rehabilitation can enhance post-stroke upper-limb motor recovery alongside conventional therapy, while also reflecting considerable heterogeneity reported across studies [17,18]. Variability in efficacy has been attributed to differences in stroke stage, baseline impairment, MI paradigms, and the degree to which neural activity is contingently linked to task-relevant movement. Notably, many prior studies have focused on chronic stroke populations, whereas interventions delivered during the subacute phase may yield greater gains [19]. In this context, the present study targeted the subacute recovery window and employed a robotic system supporting coordinated, multi-joint upper-limb movements within an intention-contingent closed loop, an approach that may be particularly suited to leveraging experience-dependent plasticity early after stroke [20]. Although improvements in global functional independence were observed, the comparatively weaker between-group separation on disability-level measures is consistent with the limited sensitivity of broad ADL instruments to short-term, impairment-focused interventions and underscores the need for larger trials with extended follow-up.
VBM analyses identified post-intervention increases in GMV within the MI-based BCI group; however, between-group comparisons did not reveal significant structural differences. The absence of significant between-group structural differences indicates that these findings should be interpreted as exploratory within-group observations rather than definitive evidence of a group-specific effect. This pattern is consistent with recent work showing that post-stroke structural and functional brain changes can variably contribute to upper-limb motor outcomes without necessarily producing robust voxel-wise group differences in traditional VBM analysis [21,22]. The anatomical distribution of observed increases aligns with regions implicated in motor planning and control, such as the SMA, but such structural adaptations are heterogeneous across individuals and may reflect compensatory processes or secondary reorganization rather than direct effects of the intervention alone [23,24]. The selective presence of structural changes in motor-related areas are consistent with the possibility of experience-dependent plasticity, but does not establish causality attributable to the MI-based BCI protocol.
Resting-state FC analyses indicated that MI-based BCI training was associated with selective modulation of motor-related networks rather than global changes in connectivity. Compared with the sham-control group, participants receiving MI-based BCI training showed greater post-intervention increases in FC between the SMA and two ipsilesional regions, the cerebellar lobule IX and the FFG. The restriction of these effects to predefined motor-related seeds is consistent with a network-specific pattern of FC changes associated with the intervention. At the network level, the SMA plays a central role in motor intention and preparation. Enhanced coupling between the SMA and cerebellar lobule IX is conceptually consistent with the closed-loop MI-based BCI paradigm, in which MI–related neural activity is contingently linked to robotic assistance and multisensory feedback. Beyond motor coordination, cerebellar lobule IX has been implicated in attentional and integrative processes that may be engaged during sustained, feedback-driven training [25,26]. Increased connectivity between the SMA and FFG may similarly reflect enhanced visuomotor integration and imagery-related processing relevant to visually guided MI tasks [27,28]. Importantly, these FC findings should be interpreted as descriptive and associative. While compatible with the theoretical framework of MI-based BCI rehabilitation, they do not establish a causal neural mechanism underlying motor recovery.
When considered together, the clinical improvements and accompanying neuroimaging findings suggest a convergent pattern of associations between clinical recovery and changes in brain structure and FC under a rigorously controlled intervention paradigm [29]. Exploratory analyses indicated that greater improvements in upper-limb motor performance were associated with larger changes in SMA-centered FC, as well as with increased GMV within the contralesional SMA [30,31]. These associations spanned clinical, functional, and structural measures, providing convergent-but not causal-evidence of coordinated neural adaptations.
Such convergence is consistent with contemporary models of post-stroke recovery, which emphasize distributed reorganization across interconnected motor-related networks rather than isolated changes within single regions. In the context of MI-based BCI training, repeated coupling of motor intention with contingent multisensory feedback may preferentially engage networks involved in motor planning, sensory integration, and cognitive control. However, given the exploratory nature of the correlation analyses, the modest sample size, and the absence of direct causal manipulation, these findings should be interpreted as descriptive associations rather than definitive evidence linking neural reorganization to functional gains.
Several limitations of the present study should be acknowledged. First, although a randomized, sham-controlled design was employed, the sample size was modest, particularly for the neuroimaging analyses, which may have limited statistical power to detect subtle between-group effects. Although the study was designed to detect a relatively large standardized effect, the observed effect for the primary outcome was smaller (Cohen’s d = 0.87), resulting in limited retrospective power (~65%). Therefore, smaller clinical effects and particularly exploratory neuroimaging effects may have been insufficiently powered. In addition, potential therapist-level clustering could not be formally accounted for because of the limited sample size. Therefore, some non-independence among participants treated by the same therapist cannot be excluded. Second, the MI-based BCI system used an asynchronous paradigm, which may permit occasional false-positive detections of MI-related activity and introduce non-optimal feedback during training. Future studies incorporating synchronous paradigms or advanced decoding algorithms may help improve signal specificity and training precision. Third, EEG acquisition was limited to a relatively small number of channels, constraining spatial resolution and the ability to characterize fine-grained cortical dynamics associated with MI and recovery. In addition, neuroimaging assessments were restricted to sMRI and rs-fMRI connectivity, precluding direct evaluation of corticospinal tract integrity or task-evoked neural activity. Multimodal approaches combining diffusion imaging, transcranial magnetic stimulation, or task-based functional MRI may provide more comprehensive insight into the neural substrates underlying BCI-assisted rehabilitation. Finally, outcomes were assessed immediately after the 4-week intervention period, and the durability of the observed motor gains and associated neural changes remains unknown. Larger-scale trials with extended follow-up will be required to determine the long-term clinical impact of MI-based BCI training, to optimize intervention parameters, and to identify patient subgroups most likely to benefit.
6. Conclusions
In patients with subacute stroke, MI-based robot-assisted BCI training was associated with greater improvements in upper-limb motor function than the closely matched sham intervention. Although exploratory, the findings support the potential value of MI-based BCI as an adjunct to conventional rehabilitation during the subacute phase. Larger randomized controlled trials with longer follow-up are needed to confirm durability and clinical applicability.
Supporting information
S1 File. Details of MRI scanning and analysis.
https://doi.org/10.1371/journal.pone.0363060.s001
(DOCX)
S1 Fig. Patient demographics and behavioral scale scores.
Baseline characteristics including FMA-UE scores, disease duration, age, pre-intervention lesion volume, pathology, location, sex, and affected hemisphere sides showed no significant differences between groups pre-intervention (all p > 0.05).
https://doi.org/10.1371/journal.pone.0363060.s005
(TIF)
S2 Fig. Horizontal brain slices demonstrating GMV increases following BCI intervention.
The 6,439-voxel cluster peak localized to the right supplementary motor area (SMA_R), with a subpeak in the right anterior cingulate cortex (ACC_R). Contrast: BG post- BG pre; Voxel p < 0.05, Cluster p < 0.05, FDR corrected; The color bar represents the t-value.
https://doi.org/10.1371/journal.pone.0363060.s006
(TIF)
S3 Fig. Spearman correlation test for the enhanced values of FC between the SMA_R and the Cerebellum_9_L, and the improvement values of each behavioral score in the BG.
(a) The enhanced FC values show a significant positive correlation with the improvement values of the FMA-UE. (b) The enhanced FC values also exhibit a significant positive correlation with the improvement values of the WMFT. (c) No significant correlation exists between the enhanced FC values and the MAS. (d) Additionally, the enhanced FC values are significantly positively correlated with the improvement values of the MBI. FC: Functional Connectivity; SMA_R: Right Supplementary Motor Area; Cerebellum_9_L: Left Cerebellar Lobule IX; BG: BCI Group; FMA-UE: Fugl-Meyer Assessment-Upper Extremity; WMFT: Wolf Motor Function Test; MAS: Motor Assessment Scale; MBI: Modified Barthel Index.
https://doi.org/10.1371/journal.pone.0363060.s007
(TIF)
S4 Fig. Spearman correlation test results for the enhanced FC values between the SMA_R and FFG_L, and the improvement values of each behavioral score in the BG.
(a) The enhancement values of FC showed a significant positive correlation with the improvement values of the FMA-UE; (b) The FC enhancement values also exhibited a significant positive correlation with the improvement values of the WMFT; (c) No significant correlation was found between the FC enhancement values and the MAS; (d) Furthermore, the FC enhancement values demonstrated a significant positive correlation with the improvement values of the MBI. FC: Functional Connectivity; SMA_R: Right Supplementary Motor Area; FFG_L: Left Fusiform Gyrus; BG: BCI Group; FMA-UE: Fugl-Meyer Assessment-Upper Extremity; WMFT: Wolf Motor Function Test; MAS: Motor Assessment Scale; MBI: Modified Barthel Index.
https://doi.org/10.1371/journal.pone.0363060.s008
(TIF)
S1 Table. Demographic data of stroke patients.
https://doi.org/10.1371/journal.pone.0363060.s009
(DOCX)
S2 Table. Brain regions showing significant differences in FC and GMV values.
https://doi.org/10.1371/journal.pone.0363060.s010
(DOCX)
References
- 1. GBD 2021 Stroke Risk Factor Collaborators. Global, regional, and national burden of stroke and its risk factors, 1990-2021: A systematic analysis for the Global Burden of Disease Study 2021. Lancet Neurol. 2024;23(10):973–1003.
- 2. Feigin VL, Brainin M, Norrving B, et al. World stroke organization: Global stroke fact sheet 2025. Int J Stroke. 2025;20(2):132–44.
- 3. Chan NH, Ng SSM. Contribution of perceived upper limb function to the participation and activity levels among community-dwelling people with chronic stroke. Ann Rehabil Med. 2025;49(3):175–86. pmid:40524518
- 4. Cervera MA, Soekadar SR, Ushiba J, Millán JDR, Liu M, Birbaumer N, et al. Brain-computer interfaces for post-stroke motor rehabilitation: A meta-analysis. Ann Clin Transl Neurol. 2018;5(5):651–63. pmid:29761128
- 5. Mokienko OA, Lyukmanov RK, Bobrov PD, Suponeva NA, Piradov MA. Brain-computer interfaces for upper limb motor recovery after stroke: Current status and development prospects (review). Sovrem Tekhnol Med. 2023;15(6):63–73. pmid:39944367
- 6. Tonin A, Semprini M, Kiper P, Mantini D. Brain-computer interfaces for stroke motor rehabilitation. Bioengineering (Basel). 2025;12(8):820. pmid:40868333
- 7. Oliveira I, Russo M, Almeida AI, Vourvopoulos A, Mendes Pereira C. Recommendations for combining brain-computer interface, motor imagery, and virtual reality in upper limb stroke rehabilitation: Qualitative participatory design study. JMIR Rehabil Assist Technol. 2025;12:e71789. pmid:41092418
- 8. Liu J, Li Y, Zhao D, Zhong L, Wang Y, Hao M, et al. Efficacy and safety of brain-computer interface for stroke rehabilitation: An overview of systematic review. Front Hum Neurosci. 2025;19:1525293. pmid:40115885
- 9. Jiang Y, Yin J, Zhao B. Motor imagery brain-computer interface in rehabilitation of upper limb motor dysfunction after stroke. J Vis Exp. 2023;199:10.3791/65405.
- 10. Cantillo-Negrete J, Carino-Escobar RI, Carrillo-Mora P, Rodriguez-Barragan MA, Hernandez-Arenas C, Quinzaños-Fresnedo J, et al. Brain-computer interface coupled to a robotic hand orthosis for stroke patients’ neurorehabilitation: A crossover feasibility study. Front Hum Neurosci. 2021;15:656975. pmid:34163342
- 11. Wu Q, Yue Z, Ge Y, Ma D, Yin H, Zhao H, et al. Brain functional networks study of subacute stroke patients with upper limb dysfunction after comprehensive rehabilitation including BCI training. Front Neurol. 2020;10:1419. pmid:32082238
- 12. Ma Z-Z, Wu J-J, Cao Z, Hua X-Y, Zheng M-X, Xing X-X, et al. Motor imagery-based brain-computer interface rehabilitation programs enhance upper extremity performance and cortical activation in stroke patients. J Neuroeng Rehabil. 2024;21(1):91. pmid:38812014
- 13. Mane R, Chew E, Phua KS, Ang KK, Robinson N, Vinod AP, et al. Prognostic and monitory EEG-biomarkers for BCI upper-limb stroke rehabilitation. IEEE Trans Neural Syst Rehabil Eng. 2019;27(8):1654–64. pmid:31247558
- 14. Pichiorri F, Morone G, Petti M, Toppi J, Pisotta I, Molinari M, et al. Brain-computer interface boosts motor imagery practice during stroke recovery. Ann Neurol. 2015;77(5):851–65. pmid:25712802
- 15. Page SJ, Fulk GD, Boyne P. Clinically important differences for the upper-extremity Fugl-Meyer Scale in people with minimal to moderate impairment due to chronic stroke. Phys Ther. 2012;92(6):791–8. pmid:22282773
- 16. Lin K, Hsieh Y, Wu C, Chen C, Jang Y, Liu J. Minimal detectable change and clinically important difference of the Wolf Motor Function Test in stroke patients. Neurorehabil Neural Repair. 2009;23(5):429–34. pmid:19289487
- 17. Ren C, Li X, Gao Q, Pan M, Wang J, Yang F, et al. The effect of brain-computer interface controlled functional electrical stimulation training on rehabilitation of upper limb after stroke: A systematic review and meta-analysis. Front Hum Neurosci. 2024;18:1438095. pmid:39391265
- 18. Li D, Li R, Song Y, Qin W, Sun G, Liu Y, et al. Effects of brain-computer interface based training on post-stroke upper-limb rehabilitation: a meta-analysis. J Neuroeng Rehabil. 2025;22(1):44. pmid:40033447
- 19. Wang A, Tian X, Jiang D, Yang C, Xu Q, Zhang Y, et al. Rehabilitation with brain-computer interface and upper limb motor function in ischemic stroke: A randomized controlled trial. Med. 2024;5(6):559-569.e4. pmid:38642555
- 20. Zhang Y, Gao Y, Zhou J, Zhang Z, Feng M, Liu Y. Advances in brain-computer interface controlled functional electrical stimulation for upper limb recovery after stroke. Brain Res Bull. 2025;226:111354. pmid:40280369
- 21. Zich C, Ward NS, Forss N, Bestmann S, Quinn AJ, Karhunen E, et al. Post-stroke changes in brain structure and function can both influence acute upper limb function and subsequent recovery. Neuroimage Clin. 2025;45:103754. pmid:39978147
- 22. Yu P, Dong R, Wang X, Tang Y, Liu Y, Wang C, et al. Neuroimaging of motor recovery after ischemic stroke - functional reorganization of motor network. Neuroimage Clin. 2024;43:103636. pmid:38950504
- 23. Hanakawa T, Hotta F, Nakamura T, Shindo K, Ushiba N, Hirosawa M, et al. Macrostructural cerebellar neuroplasticity correlates with motor recovery after stroke. Neurorehabil Neural Repair. 2023;37(11–12):775–85. pmid:37882368
- 24. Ward NS, Brown MM, Thompson AJ, Frackowiak RSJ. Neural correlates of motor recovery after stroke: A longitudinal fMRI study. Brain. 2003;126(Pt 11):2476–96. pmid:12937084
- 25. Grigoryan KA, Mueller K, Wagner M, Masri D, Pine KJ, Villringer A, et al. Short-term BCI intervention enhances functional brain connectivity associated with motor performance in chronic stroke. Neuroimage Clin. 2025;46:103772. pmid:40228398
- 26. Tanamachi K, Kuwahara W, Okawada M, Sasaki S, Kaneko F. Relationship between resting-state functional connectivity and change in motor function after motor imagery intervention in patients with stroke: A scoping review. J Neuroeng Rehabil. 2023;20(1):159. pmid:37980496
- 27. Zhang P, Duan L, Ou Y, Ling Q, Cao L, Qian H, et al. The cerebellum and cognitive neural networks. Front Hum Neurosci. 2023;17:1197459. pmid:37576472
- 28. Wang Y, Sun Y, Yu J, Cheng X, Niu Z. Brain network of athletes in motor imagery and action anticipation: An ALE meta-analysis and MACM analysis. Front Sports Act Living. 2025;7:1652165. pmid:40927390
- 29. Rehme AK, Grefkes C. Cerebral network disorders after stroke: Evidence from imaging-based connectivity analyses of active and resting brain states in humans. J Physiol. 2013;591(1):17–31. pmid:23090951
- 30. Cui F, Zhao L, Lu M, Liu R, Lv Q, Lin D, et al. Functional and structural brain reorganization in patients with ischemic stroke: a multimodality MRI fusion study. Cereb Cortex. 2023;33(19):10453–62. pmid:37566914
- 31. Thiel A, Vahdat S. Structural and resting-state brain connectivity of motor networks after stroke. Stroke. 2015;46(1):296–301. pmid:25477218

관련 뉴스