The application of brain-computer interface combined with neuromodulation for stroke upper-limb rehabilitation: a systematic review and network meta-analysis

Abstract

Background

In conventional rehabilitation, upper-limb recovery in patients with stroke is slow and limited. Motor Imagery-Based Brain-Computer Interface (MI-BCI) may promote cortical plasticity by translating brain signals into commands to control external devices; however, the effects of different intervention approaches remain unclear.

Methods

This study followed the PRISMA 2020 statement and searched PubMed, EBSCO, Web of Science, Scopus, the Cochrane Library, and the China National Knowledge Infrastructure (CNKI) for randomised controlled trials (RCTs) reporting outcomes related to upper-limb function or activities of daily living. Risk of Bias 2 (RoB 2) and Confidence in Network Meta-Analysis (CINeMA) were used to assess risk of bias and confidence in the evidence. Network meta-analysis was conducted using the network and mvmeta modules in Stata 18.0 to assess heterogeneity, network consistency, robustness of the results, and publication bias.

Results

A total of 49 studies involving 2,209 participants and 13 intervention strategies were included. For the Fugl-Meyer Assessment of the Upper Extremity (FMA-UE), interventions showing relatively favourable effects included MI-BCI combined with transcranial ultrasound stimulation and functional electrical stimulation (FES) (MD = 17.65, 95% CI: 6.88 to 28.43), MI-BCI combined with repetitive transcranial magnetic stimulation (rTMS) and rehabilitation robotics (ROB) (MD = 14.22, 95% CI: 2.98 to 25.47), and MI-BCI combined with transcranial direct current stimulation (tDCS) and ROB (MD = 13.59, 95% CI: 6.90 to 20.28). For the Action Research Arm Test (ARAT), favourable interventions included MI-BCI combined with ROB and tDCS (MD = 17.26, 95% CI: 5.96 to 28.55) and MI-BCI combined with FES (MD = 14.94, 95% CI: 9.10 to 20.78). For the Wolf Motor Function Test (WMFT), MI-BCI combined with ROB showed the largest effect (MD = 10.72, 95% CI: 9.44 to 17.18). For the Modified Barthel Index (MBI), favorable interventions included MI-BCI combined with tDCS and ROB (MD = 20.40, 95% CI: 11.21 to 29.60), MI-BCI combined with rTMS and ROB (MD = 19.46, 95% CI: 9.33 to 29.59), and MI-BCI combined with intermittent theta-burst stimulation and ROB (MD = 18.24, 95% CI: 5.93 to 30.55). Overall, the network evidence showed no obvious inconsistency; between-study heterogeneity was within an acceptable range, and no clear publication bias was detected.

Conclusion

Some non-invasive MI-BCI combined interventions like neuromodulation, ROB, or FES might improve upper-limb function and daily activities in stroke patients. However, rankings don't imply clinical importance. Due to small samples, limited evidence, and differences, these findings need confirmation in larger trials.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12883-026-05343-0.

Introduction

Population ageing increases the global health burden from age-related diseases and disabilities [4, 52]. Stroke is a major health issue due to its high incidence, long-term disability risk, and ongoing care needs, threatening individuals’ independence and quality of life [22]. The Global Burden of Disease Study 2021 reported 93.8 million stroke cases, 11.9 million new cases, 7.3 million deaths, and 160.5 million disability-adjusted life-years worldwide [1]. Stroke survivors often face lasting motor impairments and other long-term effects [24, 37]. Upper-limb dysfunction limits self-care, fine motor skills, daily activities, and social participation. Recovering upper-limb motor function is essential for regaining independence after stroke [17, 33, 51]. Improving upper-limb function is a key goal in stroke rehab and a crucial marker for assessing the clinical value and potential of new therapies [11].

Task-oriented training, constraint-induced movement therapy, mirror therapy, functional electrical stimulation (FES), and robot-assisted training (ROB) may all improve upper-limb function post-stroke, but benefits vary among interventions and patients [15, 35, 53]. Moreover, improvements in motor-function scores may not fully translate into daily activities [2]. When rehab resources are limited, or patients have restricted voluntary movement, conventional rehab may not offer enough training dosage, active participation, or task-specific sensory feedback [14]. Technology-assisted rehabilitation has gained attention for increasing training repetitions, ensuring consistent treatment, and linking patients' motor intentions with sensorimotor feedback [13, 25].

Motor imagery-based brain-computer interface (MI-BCI) detects sensorimotor rhythm changes during motor imagery or attempted movement and translates motor intention into feedback (visual, auditory, robot-assisted, or electrical), forming an intention-dependent closed loop [6, 42]. Unlike training with preset feedback schedules, MI-BCI provides feedback after detecting neural activity, enabling goal-directed training without overt movement. The tight link between motor intention and sensory feedback may promote sensorimotor network reorganization through repetitive practice and feedback-based learning [26, 45, 47]. Changes in sensorimotor and frontoparietal networks' activity or connectivity, noted in earlier studies, support the potential neuroplastic effects of MI-BCI [3, 32].

Combining MI-BCI with other technologies may target different components of stroke rehab. Depending on stimulation frequency, target, and stroke stage, repetitive transcranial magnetic stimulation (rTMS) can modulate cortical excitability and motor-network activity [29, 58]. Transcranial direct current stimulation (tDCS) modulates neuronal excitability and neuroplasticity thresholds via subthreshold membrane polarisation [39, 40, 48]. Applying neuromodulatory techniques before or during MI-BCI training preconditions neural networks, enhancing responsiveness for task-specific training. Conversely, MI-BCI-triggered FES induces muscle contractions post-motor intention detection, repeatedly pairing motor intention with movement and sensory feedback [3]. MI-BCI-controlled robotic devices offer repetitive, quantifiable movement support with real-time feedback. Thus, multimodal interventions may combine motor-intention detection, neural modulation, and sensorimotor feedback.

Functional recovery after stroke may be affected by attention, working memory, executive control, and motor-learning ability. Using rTMS or theta-burst stimulation on cognitive-control networks, combined with cognitive training, can help improve post-stroke cognition [54]. Cognitive function was not an outcome of this review; it is included only as background and does not suggest that the interventions improve cognition.

Previous meta-analyses suggest BCI or MI-BCI may improve stroke upper-limb function over conventional rehab, sham feedback, or non-contingent training [6, 27, 30]. Previous studies often pooled various feedback devices, neuromodulation techniques, and training structures, making it hard to compare the effectiveness of specific interventions. A recent network meta-analysis compared BCI-FES, tDCS, FES, and conventional treatment, but did not fully include dual- or triple-component interventions combining MI-BCI with ROB, rTMS, ITBS, or transcranial ultrasound. [60]. Existing studies vary in intervention components, parameter settings, training doses, comparators, and assessment times, with limited direct comparisons among combined interventions.

Network meta-analysis combines direct and indirect evidence if transitivity and consistency assumptions hold [8, 20, 49]. In this study, 13 treatment nodes were defined based on intervention components and their functional roles with technologies supporting theoretical interpretation. The analysis included MI-BCI alone, combined with FES or ROB, and further with tDCS, rTMS, ITBS, or ultrasound, comparing upper-limb impairment, activity performance, and daily activities within the same evidence network. The primary effect window was immediately post-intervention. RoB 2 and CINeMA assessed bias and evidence certainty. FMA-UE evaluated upper-limb impairment, ARAT and WMFT assessed activity and task performance, and MBI measured daily living activities [16, 31, 50, 55, 59]. This review aimed to compare the effectiveness and evidence certainty of MI-BCI-centered multimodal rehabilitation strategies and support future direct comparative studies.

Materials and methods

This systematic review and network meta-analysis followed PRISMA 2020 guidelines and was registered with PROSPERO (CRD420251239250) [44].

Literature search

Six electronic databases—PubMed, EBSCO, Web of Science, Scopus, Cochrane Library, and CNKI—were searched from inception to December 26, 2025, using MeSH and free-text terms. Search strategies are in Supplementary Table S1.

Inclusion and exclusion criteria

Eligibility criteria followed the PICOS framework. Studies were included if they met all listed criteria. (1) participants were patients with stroke aged ≥ 18 years who had received a clinical diagnosis confirmed by computed tomography (CT) or magnetic resonance imaging (MRI); (2) the experimental group received MI-BCI training plus at least one additional rehab intervention; (3) The study was a randomized controlled trial comparing MI-BCI-based rehab with other treatments or conventional rehab.; (4) at least one eligible outcome was reported, including upper-limb motor impairment assessed with the Fugl-Meyer Assessment of the Upper Extremity (FMA-UE), upper-limb activity using the Action Research Arm Test (ARAT) or Wolf Motor Function Test (WMFT), or activities of daily living via the Modified Barthel Index (MBI); (5) sufficient data were available to calculate or extract an effect estimate for at least one relevant outcome, with post-intervention data reported; and (6) the full text was in English or Chinese.

Studies were excluded if they met any of these criteria: (1) the experimental and control groups received identical interventions with no treatment difference; (2) essential statistical data were unavailable, preventing outcome data extraction; (3) the publication was not an original research study, including reviews, case reports, and abstracts; (4) the full text was unavailable or data incomplete after contacting authors; or (5) the study was a duplicate, retaining the most complete report.

Study selection and data extraction

Two researchers (CYY and LYY) independently extracted data using a standardised form. Disagreements were resolved through discussion, with a third researcher involved in adjudication when necessary. Extracted information covered study design, participants, interventions, comparators, outcomes, and assessment times. Study-level data included publication year, location, funding, treatment groups, intervention duration, weekly frequency, and session length. Participant data included age, sex, stroke subtype, time from stroke onset, affected side, baseline limb motor function, comorbidities, and cognitive screening results. If key information for study inclusion, treatment-node classification, or effect-size calculation was missing, the team contacted the relevant authors.

To assess transitivity, we extracted prespecified clinical and methodological effect modifiers that might influence treatment response, including participant age, sex, stroke subtype, time from stroke onset to study enrollment, stroke stage, severity of baseline upper-limb motor impairment, intervention duration, weekly intervention frequency, session length, total sessions, estimated total treatment exposure, concomitant rehab, and outcome assessment timing. Stroke stage was classified as acute, subacute, or chronic. Studies with participants across multiple stages were classified as mixed-stage unless separate data were available. Baseline impairment severity was primarily indicated by mean baseline FMA-UE score; if unavailable, other validated measures or severity classifications were used. When possible, severity was recorded as a continuous variable or according to original study categories. No reclassification was done if data were insufficient.

Intervention-level details include: (1) BCI paradigm; (2) control signals and decoding methods like neural activity for motor imagery or attempted movement; (3) peripheral devices such as FES, upper-limb robots, exoskeletons, and soft robotic gloves; (4) neuromodulatory interventions like tDCS, rTMS, ITBS, and transcranial ultrasound stimulation (TUS); (5) system mode, including intention-dependent closed-loop and fixed settings; and (6) conventional rehab details such as physiotherapy, occupational therapy, and range-of-motion exercises. Dosage is described by session length, frequency, total duration, number, and estimated exposure, calculated when data are available. Intervention parameters like stimulation frequency, pulses, intensity, current, pulse width, duty cycle, target muscles, robotic assistance, resistance, and movement repetitions are also reported.

All prespecified outcomes meeting eligibility criteria were extracted. The FMA-UE assessed upper-limb impairment, while ARAT and WMFT evaluated activity-level upper-limb function. When studies reported multiple versions of the same scale or measures, a hierarchy was applied. For activities of daily living, if both MBI and Functional Independence Measure (FIM) were reported, MBI data were preferred. Data at the end of intervention and follow-up were extracted separately, recording exact timing. If means and SDs were not reported, estimates were made from medians, IQRs, or figures using Get Data Graph Digitizer v2.25. Studies lacking necessary data after contacting authors were excluded from quantitative analysis.

Risk of bias and certainty of evidence assessment

The revised Cochrane risk-of-bias tool for randomised trials (RoB 2) assessed the risk of bias of included studies. RoB 2 evaluates five domains: bias from randomisation, deviations from interventions, missing outcome data, outcome measurement, and selection of reported results [36]. Two reviewers independently judged each domain as “low risk,” “some concerns,” or “high risk.” Disagreements were resolved through discussion or, if needed, a third reviewer. Results were visualised using a macro-enabled Excel tool.

Following Development and Evaluation (GRADE) framework principles, the Confidence in Network Meta-Analysis (CINeMA) framework assessed the certainty of network meta-analysis evidence [41]. CINeMA evaluates six domains—bias, reporting, indirectness, imprecision, heterogeneity, and incoherence—and rates evidence as high, moderate, low, or very low [46]. When concerns affected certainty, comparisons were downgraded per GRADE/CINeMA criteria [38].

Treatment node classification

As shown in Table 1, the final evidence network had 13 treatment nodes (see Supplementary Table S2). Nodes were classified based on operational definitions, intervention components, and criteria used in neurorehabilitation network meta-analyses. Interventions with similar core components and goals were grouped, while those differing significantly, like neuromodulation techniques or peripheral devices, were separate nodes [57]. The mechanisms of action mainly helped interpret differences rather than serve as the sole basis for classification.

Table 1.

Final treatment nodes included in the network meta-analysis

Treatment Node	Description
MI-BCI + ITBS + ROB	A closed-loop, tri-component intervention combining intermittent theta-burst stimulation (ITBS) with MI-BCI–triggered robotic upper-limb training. ITBS is applied as an adjunctive neuromodulatory element (usually before or alongside training), while MI-BCI decoding provides intention-dependent control of robotic movement
MI-BCI + tDCS + ROB	A closed-loop, three-component approach where transcranial direct current stimulation (tDCS) is used alongside motor imagery-based brain-computer interface (MI-BCI)–driven robotic training. tDCS is applied either before or during the training session, and MI-BCI decoding activates or influences robotic upper-limb movement based on the user's intention
MI-BCI + rTMS + ROB	A closed-loop, three-component intervention combining repetitive transcranial magnetic stimulation (rTMS) with MI-BCI–triggered robotic upper-limb training. rTMS acts as an adjunctive neuromodulatory element, while MI-BCI decoding supports intention-dependent robotic execution
MI-BCI + TUS + FES	A closed-loop, three-part intervention combines transcranial ultrasound stimulation (TUS) with MI-BCI–triggered functional electrical stimulation (FES). TUS acts as a supplementary neuromodulatory component, while MI-BCI decoding triggers FES on the affected limb in real time
MI-BCI + FES	A closed-loop intervention where MI-BCI decoding triggers FES of the paretic limb in real time, thereby connecting motor intention with peripheral muscle activation (intention-contingent BCI–FES training)
MI-BCI + ROB	A closed-loop intervention in which MI-BCI decoding enables intention-contingent control of an upper-limb robotic device, such as an exoskeleton or an end-effector robot, to perform movements
MI-BCI	Standalone MI-BCI training where EEG-based decoding of motor imagery (or attempted movement, where relevant) provides intention-contingent feedback, such as visual or auditory cues, without using peripheral actuation devices like FES or robotic systems
ITBS	Intermittent theta-burst stimulation delivered as a standalone neuromodulation treatment without using MI-BCI or peripheral actuation
rTMS	Repetitive transcranial magnetic stimulation administered as an independent neuromodulation intervention without utilising MI-BCI or peripheral actuation
tDCS	Transcranial direct current stimulation administered as an independent neuromodulation intervention, excluding MI-BCI or peripheral methods of actuation
FES	Functional electrical stimulation is delivered without MI-BCI triggering, meaning it is pre-programmed or fixed-parameter stimulation
ROB	Standalone robot-assisted upper-limb training provided without MI-BCI control or supplementary neuromodulation
CR	Conventional rehabilitation, excluding MI-BCI, neuromodulation, robotic devices, or FES, usually involves physiotherapy, occupational therapy, range-of-motion exercises, and basic activities-of-daily-living training

All intervention arms additionally received standard conventional rehabilitation

MI-BCI-centred interventions were classified into seven nodes based on system components, whether transcranial neuromodulation was included, and the peripheral effector device type. These included MI-BCI alone, MI-BCI + FES, MI-BCI + ROB, and four triple-component interventions: MI-BCI + tDCS + ROB, MI-BCI + rTMS + ROB, MI-BCI + ITBS + ROB, and MI-BCI + TUS + FES. MI-BCI was the core component, combined with peripheral technologies and neuromodulation to enhance motor intention coupling with sensorimotor feedback and promote neuroplasticity [32, 42, 45]. Node separation was based on distinct additional treatments, including neuromodulation (tDCS, rTMS, ITBS, TUS) and effector devices (ROB, FES) [5, 7, 12]. Interventions with only intrinsic MI-BCI modules, like neural decoding, intention detection, and feedback, were classified as MI-BCI-alone.

Interventions without a BCI component were classified into six nodes based on main technological features: ROB, FES, tDCS, rTMS, ITBS, and conventional rehabilitation (CR) [39, 40]. These have been used in post-stroke upper-limb neurorehabilitation and served as important comparators for evaluating MI-BCI-centred interventions. Because CR was the most common comparator across trials and represented routine practice, it was chosen as the main reference for the network meta-analysis [56, 60]. This allowed consistent comparison of treatment effects relative to CR and improved interpretability, without changing the relative effect estimates between nodes [34].

Statistical method

All statistical analyses were performed using Stata 18.0, and the network meta-analysis was conducted using the network and mvmeta commands. Evidence network plots illustrated connections among treatment nodes, sample sizes, and direct comparison evidence. For each outcome, a random-effects model estimated mean differences (MDs) and 95% confidence intervals (CIs). Results were summarized with forest plots and league tables. P-scores ranked treatments, with higher values indicating better rankings. Node-splitting models compared direct and indirect evidence to assess inconsistency. Funnel plots and Egger’s regression tested for small-study effects and publication bias. All tests were two-sided, with p < 0.05 indicating significance. The effect direction was standardised so that MD > 0 signified greater improvement with the intervention.

Results

Literature search results

As shown in Fig. 1, a total of 1,272 records were obtained from database searches. After removing duplicates, 816 records remained. After screening titles and abstracts, 702 records were excluded. The remaining 104 records were screened in full text, and 8 articles that could not be accessed in full were removed. Finally, 106 studies were selected based on inclusion criteria, and 60 studies that did not meet the requirements were excluded, leaving 46 studies. Additionally, 3 more studies were identified through manual searches (such as Google Scholar and other open-access resources). In total, 49 studies were included.

Fig. 1.

PRISMA flow diagram of included and excluded studies

Characteristics of included studies and assessment of transitivity

This review included 49 studies with 2,238 participants (see Supplementary Table S3). Of those reporting sex, 1,302 were male, 665 female; sex data was incomplete for others. Participants ranged from 18 to 90 years old, averaging about 58.1 years. The studies involved ischemic, hemorrhagic, or mixed strokes, with varying distributions. Overall, participants shared similarities in age, sex, and stroke type across treatment groups.

The time from stroke onset to study enrolment ranged from 3 days to 2.4 years. One study mainly enrolled participants in the acute stage, 28 in the subacute, 16 in the chronic, and four included mixed-stage populations spanning subacute and chronic. Most evidence came from subacute and chronic patients, while evidence from acute stage was limited. Some overlap in time since stroke existed across treatment nodes, but several peripheral nodes were informed by only one or a few studies from specific stages.

The severity of baseline upper-limb motor impairment varied across studies, using measures like Brunnstrom stage, Modified Ashworth Scale (MAS), muscle-strength grade, and FMA-UE. These captured different clinical aspects but provided overall baseline motor and neuromuscular status. Most participants had moderate-to-severe impairment, with some studies restricting eligibility to certain Brunnstrom stages, MAS scores ≤ 2–3, muscle strength ≤ grade 2, inability to extend the wrist, or hand dysfunction. A few included a broader range from severe to mild-to-moderate impairment. Despite different definitions and methods, some clinical overlap in baseline severity was noted across treatment comparisons.

Intervention dose varied across studies. Most used ~ 30-min training sessions, five times weekly for about 4 weeks. Session lengths ranged from 10 to 180 min, weekly sessions from 1 to 7, and duration from 2 to 8 weeks. Median total exposure was around 10 h, with some studies exceeding this. Most delivered 10–24 specialised sessions, but some used longer sessions or higher doses. Concomitant rehabilitation, including physiotherapy, occupational therapy, and other therapies, varied across trials. These treatments were generally provided to both groups, but overall treatment exposure still differed.

Outcome assessments were mainly done at the end of the intervention. The FMA-UE was the most common measure of upper-limb motor impairment, while the ARAT and WMFT, used less often, mainly assessed activity-level function and task performance. The MBI primarily measured activities of daily living. Most studies reported only end-of-treatment effects, with few including follow-up assessments at 2 weeks, 4 weeks, 1 month, 3 months, or longer, with variable timing. To improve comparability, end-of-study assessments were the primary effect window, while longer-term follow-ups were mainly supplementary.

The studies evaluated seven MI-BCI interventions: MI-BCI alone, combined with FES, ROB, tDCS, rTMS, ITBS, and TUS. Trials differed in decoding methods, feedback, devices, neuromodulation, and timing. Six non-BCI treatments (rTMS, tDCS, FES, ROB, CR, ITBS) were also assessed, totalling 13 treatment nodes.

From a transitivity perspective, participant age, sex, stroke type, and baseline limb motor function showed some overlap across treatment comparisons. Most studies focused on subacute and chronic patients with short-term, high-frequency rehab, with outcomes assessed at intervention end. This overlap supports, but doesn't confirm, transitivity, as some factors like stroke stage, time since stroke, and network nodes were unevenly balanced or supported by few studies. Variability in impairment measures, treatment duration, and rehab content further complicates comparisons. Few studies had long-term follow-up, with inconsistent assessment timing. These issues affect treatment response and increase uncertainty, so transitivity is considered reasonable but uncertain.

Risk of bias and certainty of evidence

Forty-nine studies were included; RoB 2 results are in Supplementary Table S4. Most studies had low risk of bias or some concerns, especially in randomization, missing data, outcome measurement, and reported results. More concerns appeared in deviations from intended interventions. Few studies had high risk, mainly due to issues in randomization and assessors knowing intervention allocation, potentially biasing outcomes.

The network evidence certainty assessed by CINeMA is in Supplementary Tables S5-1 to S5-4. Evidence quality varied across outcomes and comparisons. For FMA-UE, MI-BCI + TUS + FES, MI-BCI + ITBS + ROB, and MI-BCI + tDCS + ROB versus CR had moderate certainty, mainly limited by within-study bias. MI-BCI + FES and MI-BCI + ROB versus CR had low certainty, due to bias and heterogeneity. MI-BCI alone versus CR had moderate certainty; other standalone interventions versus CR were rated low or very low due to imprecision.

For the ARAT, MI-BCI + tDCS + ROB versus CR was rated as moderate certainty, mainly due to within-study bias and imprecision. MI-BCI + FES and MI-BCI + ROB versus CR were rated low because of bias and imprecision. For the WMFT, MI-BCI + ROB and MI-BCI + FES versus CR were rated moderate, but MI-BCI alone versus CR was low due to imprecision.

For the MBI, evidence certainty for principal MI-BCI interventions versus CR was generally low. Comparisons of MI-BCI + tDCS + ROB and MI-BCI + ITBS + ROB versus CR were limited mainly by bias, imprecision, and heterogeneity. MI-BCI + rTMS + ROB versus CR was limited mainly by bias and imprecision, while MI-BCI + FES and MI-BCI + ROB versus CR were mainly affected by bias and heterogeneity. Several standalone neuromodulation interventions versus CR had very low certainty due to serious imprecision.

The network effect estimates, the number of studies with direct evidence, CINeMA certainty ratings, and main reasons for downgrading for key treatment comparisons are summarised in Table 2. Several comparisons for FMA-UE and WMFT had moderate-certainty evidence, while ARAT comparisons ranged from moderate to low certainty. MBI-related evidence was mostly low certainty due to bias, imprecision, and heterogeneity.

Table 2.

Summary of network estimates and CINeMA confidence ratings for principal comparisons

Outcome	Comparison	MD (95% CI)	Direct evidence, n	CINeMA rating	Downgrading reasons
FMA-UE	MI-BCI + TUS + FES	17.65 (6.88 to 28.43)	0	Moderate	Within-study bias
FMA-UE	MI-BCI + ITBS + ROB	14.22 (2.98 to 25.47)	0	Moderate	Within-study bias
FMA-UE	MI-BCI + tDCS + ROB	13.59 (6.90 to 20.28)	0	Moderate	Within-study bias
FMA-UE	MI-BCI + FES	10.98 (6.77 to 15.20)	9	Low	Within-study bias; heterogeneity
FMA-UE	MI-BCI + ROB	9.09 (5.22 to 12.97)	8	Low	Within-study bias; heterogeneity
ARAT	MI-BCI + tDCS + ROB	17.26 (5.96 to 28.55)	0	Moderate	Within-study bias; imprecision
ARAT	MI-BCI + FES	14.94 (9.10 to 20.78)	1	Low	Within-study bias; imprecision
ARAT	MI-BCI + ROB	13.57 (2.85 to 24.29)	1	Low	Within-study bias; imprecision
WMFT	MI-BCI + ROB	10.72 (4.03 to 17.41)	3	Moderate	Within-study bias
WMFT	MI-BCI + FES	8.09 (4.90 to 11.29)	0	Moderate	Within-study bias
WMFT	MI-BCI	8.00 (2.42 to 13.58)	0	Low	Within-study bias; imprecision
MBI	MI-BCI + tDCS + ROB	20.40 (11.21 to 29.60)	0	Low	Within-study bias; imprecision; heterogeneity
MBI	MI-BCI + rTMS + ROB	19.46 (9.33 to 29.59)	1	Low	Within-study bias; imprecision
MBI	MI-BCI + ITBS + ROB	18.24 (5.93 to 30.55)	0	Low	Within-study bias; imprecision; heterogeneity
MBI	MI-BCI + ROB	11.24 (5.87 to 16.61)	0	Low	Within-study bias; heterogeneity
MBI	MI-BCI + FES	10.24 (5.23 to 15.25)	0	Low	Within-study bias; heterogeneity

Findings from the network meta-analysis

In the network plot, each node shows a treatment modality, size proportional to the sample size. Connecting lines indicate direct comparisons, with thickness reflecting the number of trials.

FMA-UE

The FMA-UE evidence network included 44 randomized controlled trials, with 11 treatment nodes and 14 direct comparisons, forming a connected network (Fig. 2, left). CR, MI-BCI + ROB, and MI-BCI + FES were supported by more direct evidence and participants, while nodes like ITBS, tDCS, and MI-BCI + TUS + FES had limited direct head-to-head evidence. No significant differences were found between direct and indirect evidence in node-splitting analysis (all p > 0.05; Supplementary Figure S6.1 and Table S6.1), with p-values from 0.164 to 0.945.

Fig. 2.

Network plot (left) and P-score ranking (right) for FMA-UE

The random-effects model found that MI-BCI + TUS + FES had the largest network effect (MD = 17.65, 95% CI: 6.88–28.43), followed by MI-BCI + ITBS + ROB (MD = 14.22, 95% CI: 2.98–25.47), and MI-BCI + tDCS + ROB (MD = 13.59, 95% CI: 6.90–20.28). MI-BCI + FES (MD = 10.98, 95% CI: 6.77–15.20) and MI-BCI + ROB (MD = 9.09, 95% CI: 5.22–12.97) also had favorable effects, with all five CIs not crossing zero. MI-BCI and FES alone showed smaller effects, while ITBS, tDCS, and ROB alone had smaller estimates with CIs crossing zero.

The P-score ranking generally aligned with the network effect estimates (Fig. 2, right). MI-BCI + TUS + FES ranked highest, then MI-BCI + tDCS + ROB and MI-BCI + ITBS + ROB. These three strategies were supported by only one, four, and one study involving 34, 86, and 22 participants, respectively. Thus, high P-scores, especially for single-study nodes, should not be seen as definitive but considered alongside network effects, 95% CIs, and CINeMA certainty.

The league-table comparisons (Supplementary Figure S6.3) showed that MI-BCI + FES, MI-BCI + ROB, MI-BCI + tDCS + ROB, MI-BCI + rTMS + ROB, and MI-BCI + TUS + FES were significantly better than CR, with the 95% CIs not crossing zero. No asymmetry was in the funnel plot, and Egger’s test did not detect small-study effects (p = 0.951; Supplementary Figure S6.4).

ARAT

The ARAT evidence network included 12 RCTs, with 8 treatment nodes and 8 direct comparisons (Fig. 3, left). Nodes included CR, FES, MI-BCI, MI-BCI + FES, MI-BCI + ROB, MI-BCI + tDCS + ROB, ROB, and tDCS. CR was well connected, and MI-BCI + ROB and MI-BCI + tDCS + ROB had more studies. For three comparisons assessed by node-splitting, no significant differences were found between direct and indirect evidence (all p > 0.05; Supplementary Figure S7.1 and Table S7.1).

Fig. 3.

Network plot (left) and P-score ranking (right) for ARAT

The random-effects model found that MI-BCI + tDCS + ROB had the largest network effect (MD = 17.26, 95% CI: 5.96 to 28.55), followed by MI-BCI + FES (MD = 14.94, 95% CI: 9.10 to 20.78) and MI-BCI + ROB (MD = 13.57, 95% CI: 2.85 to 24.29) (Supplementary Figure S7.2). All three interventions favored the intervention group, with CIs not crossing zero. Standalone ROB (MD = 11.44, 95% CI: − 0.16 to 23.04) and tDCS (MD = 9.65, 95% CI: − 1.94 to 21.25) also favored the intervention group, but their CIs crossed zero. FES and MI-BCI alone showed smaller effect estimates with more uncertainty.

The P-score ranking was MI-BCI + tDCS + ROB, followed by MI-BCI + FES and MI-BCI + ROB (Fig. 3, right). However, the highest-ranked intervention, MI-BCI + tDCS + ROB, was supported by only three studies involving 76 participants. Therefore, interpretation of the treatment rankings should consider the number of studies, number of participants, network relative-effect estimates and their 95% CIs, and CINeMA certainty of evidence. A high P-score should not be interpreted in isolation as definitive evidence of clinical superiority.

The league-table comparisons (Supplementary Figure S7.3) showed that MI-BCI + tDCS + ROB, MI-BCI + FES, and MI-BCI + ROB were statistically significantly superior to CR, with the corresponding 95% CIs not crossing zero. These findings further support potentially favourable effects of these MI-BCI-based combined interventions relative to CR; however, they do not establish that the combined interventions are necessarily superior to their corresponding individual components or that the different components act synergistically. No obvious asymmetry was observed in the comparison-adjusted funnel plot, and Egger’s regression test did not detect a statistically significant small-study effect (p = 0.805; Supplementary Figure S7.4).

WMFT

The WMFT evidence network included 12 RCTs with 6 treatment nodes and 6 direct comparisons (Fig. 4, left). The nodes were CR, FES, MI-BCI, MI-BCI + FES, MI-BCI + ROB, and ROB. CR was well connected, while MI-BCI + ROB and MI-BCI + FES were common combined interventions. Node-splitting showed no significant differences between direct and indirect evidence for MI-BCI + ROB vs CR, ROB vs CR, and MI-BCI + ROB vs ROB; p values were 0.400, 0.554, and 0.766 (Supplementary Figure S8.1 and Table S8.1).

Fig. 4.

Network plot (left) and P-score ranking (right) for WMFT

The random-effects model showed that, compared with CR, MI-BCI + ROB (MD = 10.72, 95% CI: 4.03 to 17.41), MI-BCI + FES (MD = 8.09, 95% CI: 4.90 to 11.29), and MI-BCI alone (MD = 8.00, 95% CI: 2.42 to 13.58) all showed favourable network relative effects, with none of the corresponding 95% CIs crossing zero (Supplementary Figure S8.2). In contrast, standalone ROB (MD = 3.04, 95% CI: − 3.82 to 9.91) and standalone FES (MD = 2.68, 95% CI: − 1.11 to 6.46) showed smaller effect estimates, and their 95% CIs crossed zero. No marked statistical heterogeneity was detected within the network (I² = 0.0%, τ = 0.1122, p for heterogeneity = 0.4867).

P-score rankings placed MI-BCI + ROB first, followed by MI-BCI + FES and MI-BCI alone (Fig. 4, right). MI-BCI + ROB was supported by four studies with 61 participants. Although larger than some peripheral nodes' evidence, it remains limited. Thus, a high P-score should not be seen as definitive but alongside network effect estimates, 95% CIs, evidence size, and CINeMA certainty.

The league-table comparisons (Supplementary Figure S8.3) showed significant favourable effects relative to CR for MI-BCI + ROB, MI-BCI + FES, and MI-BCI alone, with 95% CIs not crossing zero. These suggest potential benefits of MI-BCI interventions on WMFT outcomes but do not confirm they are superior to individual components or act synergistically. No asymmetry was seen in the funnel plot, and Egger’s test found no small-study effect (p = 0.625; Supplementary Figure S8.4).

MBI

The MBI evidence network included 26 RCTs with 12 treatment nodes and 16 comparisons. The structure was complex but connected (Fig. 5, left). CR, MI-BCI + ROB, and MI-BCI + FES had more evidence and participants, while tDCS, rTMS, ITBS, and related MI-BCI interventions were mainly at the network's periphery. For five comparisons assessed via node-splitting analysis, no significant differences were found between direct and indirect evidence, with p values from 0.301 to 0.950 (Supplementary Figure S9.1, Supplementary Table S9.1). Specifically, p values for rTMS versus CR, MI-BCI + rTMS + ROB versus CR, MI-BCI + ROB versus rTMS, MI-BCI + ROB versus MI-BCI + rTMS + ROB, and tDCS versus MI-BCI + tDCS + ROB were 0.301, 0.516, 0.463, 0.350, and 0.950.

Fig. 5.

Network plot (left) and P-score ranking (right) for MBI

The random-effects model showed that, compared with CR, MI-BCI + tDCS + ROB had the largest network relative-effect estimate (MD = 20.40, 95% CI: 11.21 to 29.60), followed by MI-BCI + rTMS + ROB (MD = 19.46, 95% CI: 9.33 to 29.59) and MI-BCI + ITBS + ROB (MD = 18.24, 95% CI: 5.93 to 30.55) (Supplementary Figure S9.2). Among standalone neuromodulation interventions, tDCS (MD = 14.98, 95% CI: 5.64 to 24.33) and rTMS (MD = 12.62, 95% CI: 2.19 to 23.04) showed effects favouring the intervention, with 95% CIs not crossing zero. ITBS also showed a positive point estimate, but with lower precision (MD = 10.77, 95% CI: − 1.89 to 23.43). MI-BCI + ROB (MD = 11.24, 95% CI: 5.87 to 16.61) and MI-BCI + FES (MD = 10.24, 95% CI: 5.23 to 15.25) showed favourable effects, with 95% CIs not crossing zero. In contrast, the 95% CIs for MI-BCI alone (MD = 6.77, 95% CI: − 1.08 to 14.62), ROB alone (MD = 1.90, 95% CI: − 5.52 to 9.31), and FES alone (MD = 0.68, 95% CI: − 6.47 to 7.83) crossed zero.

P-score rankings placed MI-BCI + tDCS + ROB first, followed by MI-BCI + rTMS + ROB and MI-BCI + ITBS + ROB (Fig. 5, right). The top-ranked MI-BCI + tDCS + ROB was supported by only two studies with 29 participants, and comparison with tDCS was also based on two studies. Thus, a high P-score isn't definitive evidence of clinical superiority but should be considered with the network effects, 95% CIs, evidence base size, and CINeMA certainty, which was low for all three interventions versus CR due to bias, imprecision, and heterogeneity.

The league-table comparisons (Supplementary Figure S9.3) further supported favourable effects of MI-BCI + tDCS + ROB, MI-BCI + rTMS + ROB, and MI-BCI + ITBS + ROB relative to CR. However, these comparisons alone cannot establish that triple-component strategies are generally superior to other standalone or dual-component interventions, nor can they demonstrate definitive synergistic effects among the different treatment components. No obvious asymmetry was observed in the comparison-adjusted funnel plot, and Egger’s regression test did not detect a statistically significant small-study effect (p = 0.463; Supplementary Figure S9.4).

Discussion

This study used network meta-analysis to compare MI-BCI strategies for post-stroke upper-limb rehab. Compared with CR, MI-BCI combined interventions showed beneficial effects on motor impairment, activity, and daily living. Multimodal strategies with FES, ROB, or neuromodulation ranked higher, while some standalone interventions had smaller effects or confidence intervals crossing zero. This aligns with evidence that motor-intention BCI training aids post-stroke recovery by connecting neural activity with real-time feedback [6, 27, 30]. Some highly ranked interventions were supported by few studies with limited participants. P-scores only show relative ranking within the evidence network and do not indicate high certainty or clinical superiority. Rankings should be interpreted alongside effects, confidence intervals, evidence size, and CINeMA certainty.

The clinical relevance of the effects relates to minimal clinically important differences (MCIDs). Studies report FMA-UE MCIDs of about 4.25–7.25 points in chronic stroke patients with mild-to-moderate upper-limb impairments [43], an MCID of 12.4 points has been reported in patients with moderate-to-severe hemiparesis during convalescence [18]. This study found network effect estimates for MI-BCI interventions versus CR ranged from 9.09 to 17.65 points, exceeding the 4.25–7.25-point reference range, with some triple-component strategies over 12.4 points. These effects may be clinically important rather than just statistically significant. However, MCID thresholds vary by population and are not direct equivalents to network meta-analysis effects. They should be used as supportive references, not definitive evidence of clinical importance.

The ARAT findings showed a similar pattern. Previous studies report about 5.7 points as a meaningful reference for chronic stroke patients, while MCIDs of 12 and 17 points are reported in early-stage stroke, varying by factors like limb dominance [28, 55]. The network effect estimates for MI-BCI + tDCS + ROB, MI-BCI + FES, and MI-BCI + ROB versus CR ranged from 13.57 to 17.26 points, all exceeding 5.7 points and overlapping with the 12–17-point range for early stroke patients. This suggests their average effects at the activity level are likely clinically important, not just statistically significant. Some network estimates had wide confidence intervals, and evidence size varied across treatment nodes. Thus, a point estimate above the MCID threshold doesn't mean all patients will see meaningful improvement.

For activities of daily living, [19] Estimated an MCID of 1.85 points for the 20-point Barthel Index, roughly 9.25 points on a 100-point scale. In this study, MI-BCI combined interventions vs. CR showed effects from 10.24 to 20.40 points, reaching or exceeding this reference. This suggests some interventions may be clinically important for daily activities. Different Barthel Index versions vary in scoring and measurement, so linear conversion from 20- to 100-point scales is only approximate. Also, MCIDs and average effects from network meta-analysis differ conceptually; their comparison helps interpret clinical relevance but doesn't guarantee individual patient achievement of MCID.

The large network effects in multimodal MI-BCI strategies likely stem from the functional complementarity among motor-intention detection, neural-network modulation, and sensorimotor feedback. MI-BCI detects neural activity during motor imagery or movement and provides feedback when the target signal is identified. Unlike feedback given on a schedule or randomly, this motor-intention-contingent closed-loop coupling creates a closer temporal link between cortical activity and sensorimotor feedback. Pichiorri et al. [45] found that BCI-supported motor imagery training improved motor function, increased sensorimotor rhythmic synchronisation in the affected hemisphere, and changed intrahemispheric functional connectivity Krueger et al. [26]. Further showed that BCI-FES synchronised with movement-related activity led to greater improvements than random FES, with changes in cortical and corticomuscular connectivity, supporting Hebbian-like plasticity. However, a direct causal link between these neurophysiological changes and clinical recovery is unproven, so they cannot be confirmed as mediators.

This mechanism may explain how FES and ROB complement MI-BCI. MI-BCI-triggered FES repeatedly links motor intention with muscle contraction, sensation, and proprioception, possibly enhancing the connection between motor commands and sensory feedback. Biasiucci et al. [3] reported that improvements in motor function after brain-activity-triggered FES lasted 6–12 months and increased functional connectivity within the motor network. ROB offers repetitive, quantifiable movement assistance and proprioceptive feedback, connecting motor intention with sensorimotor outcomes. FES and ROB can strengthen closed-loop coupling by increasing feedback salience and consistency. Research shows MI-contingent BCI feedback yields greater motor improvements and functional connectivity changes than non-motor imagery feedback, highlighting the importance of matching feedback to motor intention in BCI rehab [23].

tDCS, rTMS, and ITBS do not replace motor training but may precondition for task-specific training by modulating neuronal excitability, cortical balance, and network connectivity. Multimodal strategies may work through the temporal and functional complementarity of components. First, residual neural network plasticity is modulated, then MI-BCI detects motor intentions, followed by peripheral feedback via FES or robotics. This process involves neural-state modulation, motor-intention training, and peripheral reinforcement, but currently remains mostly neurophysiological. The study only compared clinical outcomes and did not test interactions among treatments; thus, it cannot show biological or statistical synergy. Variations in neuromodulation parameters, targets, timing, and dose may affect outcomes.

rTMS has a strong neurophysiological basis, as stroke recovery involves not just local changes but also reorganisation across bilateral motor networks. Depending on factors like stimulation frequency, target, stroke stage, and patient traits, rTMS may boost excitability in the ipsilesional cortex or modulate increased activity in the contralesional cortex, affecting the motor network's excitatory–inhibitory balance. Chen et al. [9] found that, in 63 acute stroke patients, low-frequency stimulation of the contralesional M1 combined with high-frequency stimulation of the ipsilesional M1 altered intra- and interhemispheric connectivity, some of which were linked to motor recovery. This mode of action may complement MI-BCI by first modulating the motor network's excitability and connectivity with rTMS, then recruiting task-related networks through MI-BCI, which reinforces neural activity via sensorimotor feedback, creating a process linking network modulation with intention-dependent training.

Clinical studies support this combined strategy. Johnson et al. [21] combined rTMS with virtual-reality BCI in three chronic stroke patients and showed changes in interhemispheric inhibition and increased activation of the affected cortex under real-rTMS-plus-BCI. These preliminary findings support the feasibility of integrating both technologies. In a randomized trial with 44 stroke patients, Choi et al. [10] found that low-frequency rTMS combined with motor imagery training improved FMA-UE, WMFT, and ARAT scores more than rTMS alone, with increased M1 activation, supporting the idea of functional complementarity between rTMS and motor-intention training. However, the study did not directly evaluate rTMS with MI-BCI, providing only indirect clinical evidence. Similarly, Tringali et al. [54] reviewed RCTs on post-stroke cognitive rehab, supporting that rTMS can modulate cortical excitability and plasticity. Yet, since their focus was on cognitive function, the findings can't be directly applied to upper-limb motor rehab, so their support for rTMS plus MI-BCI remains indirect.

From a clinical translation perspective, some triple-component strategies showed large network effects but also increased complexity, equipment needs, staff training, treatment time, and costs. In contrast, dual-component strategies like MI-BCI + FES and MI-BCI + ROB demonstrated favourable effects across outcomes and may be easier to integrate into routine rehab. For example, MI-BCI + ROB was supported by moderate evidence for WMFT, while some highly ranked triple strategies relied on few studies and participants, indicating that high network ranking doesn't always mean greater evidence certainty or broader clinical applicability.

Treatment response may be influenced by patient characteristics such as stroke stage, initial severity of upper-limb impairment, corticospinal tract integrity, motor imagery ability, and cognitive function. These factors affect a patient's ability to generate stable motor-related brain signals, engage in sensorimotor feedback learning, and develop neuroplasticity. Additionally, the trials varied in stroke duration, impairment severity, session length, treatment exposure, and concurrent rehabilitation. Hence, an optimal treatment dose remains undetermined.

Limitations and future directions

This study has limitations. First, evidence supporting different treatment nodes was imbalanced. CR, MI-BCI + ROB, and MI-BCI + FES had more studies and participants, while some triple-component interventions at the network's periphery were supported by only one or a few studies, and some high-ranked strategies had small samples. As a result, network effect estimates and P-scores for these nodes may be influenced by individual studies and errors. Higher rankings shouldn't be seen as definitive clinical superiority but should be considered with confidence intervals, evidence size, and CINeMA certainty. The low or moderate evidence certainty for key comparisons also indicates uncertainty in treatment rankings.

Second, there was significant variation in MI-BCI system implementations, including BCI paradigms, signal processing, decoding algorithms, feedback modes, and closed-loop triggers. Peripheral devices like FES, exoskeletons, robotic arms, and soft gloves also differed. While treatment strategies were grouped by core tech components, specific technical differences within these groups could affect motor-intention detection, feedback latency, training success, sensorimotor feedback quality, and ultimately, the physiological and clinical outcomes of BCI systems.

Third, Neuromodulation parameters were highly variable across studies of tDCS, rTMS, ITBS, and TUS, differing in target, side, frequency, intensity, pulse count, duration, and timing relative to MI-BCI training. Although data on these factors were extracted, the average effects for each treatment node reflect interventions with similar core components but varying parameters. Thus, the current network estimates cannot identify which stimulation parameter yields the greatest therapeutic effect.

Fourth, the intervention-specific treatment dose and the content and dose of concomitant CR varied across studies. Most studies used about 30-min sessions five times weekly for around 4 weeks; however, session durations ranged from 10 to 180 min, weekly sessions from 1 to 7, intervention durations from 2 to 8 weeks, and total exposure from 1.5 to 53.3 h. These differences in technological components, session length, training frequency, total treatment time, and background rehab dose may affect relative effects and increase uncertainty in transitivity and indirect comparison interpretations.

Fifth, Attribution of effects was challenging for multicomponent interventions. Key treatment nodes mainly involved MI-BCI combined with FES or ROB, with some strategies adding tDCS, rTMS, ITBS, or TUS. Network meta-analysis estimates the effects of entire treatment combinations, not individual components. Larger effects from MI-BCI + tDCS + ROB or MI-BCI + rTMS + ROB cannot be solely attributed to any single part. The apparent superiority of some triple strategies over dual or standalone interventions does not confirm biological synergy. Variations in dose and timing further complicate attribution of effects.

Sixth, evidence on long-term effectiveness remains limited. Most outcomes were at the intervention's end, with few studies conducting follow-up at 2 weeks, 4 weeks, 1 month, 3 months, or longer. Follow-up timing was inconsistent. Thus, findings mainly show end-of-treatment or short-term effects and can't confirm if improvements last for months or more.

Seventh, studies varied in stroke stage, baseline impairment, and rehabilitation, with some clinical details incomplete. While effect modifiers overlapped across treatments, the transitivity assumption wasn't fully verified. Technology-assisted rehab also hindered blinding of participants and therapists, and some studies lacked adequate blinding of outcome assessors, increasing bias risk due to deviations and measurement bias.

Future research should include adequately powered, multicentre, head-to-head RCTs; standardise reporting of MI-BCI systems and neuromodulation; standardise intervention and rehabilitation doses; use study designs to distinguish treatment components; and extend follow-up. Mechanistic measures like EEG, MEP, fMRI/fNIRS, electromyography, and kinematics could be linked to outcomes to identify beneficial components and optimise MI-BCI combinations balancing effectiveness, evidence, and feasibility.

Conclusions

This network meta-analysis found that combining MI-BCI with FES, ROB, or neuromodulation improved post-stroke upper-limb impairment, activity, and daily functions compared to CR. Several triple-component strategies showed strong effects on FMA-UE, ARAT, and MBI. MI-BCI + ROB and FES had promising results for clinical use. However, some top-ranked treatments were supported by few small studies, with variations in BCI paradigms, neuromodulation parameters, intervention dose, rehabilitation, and follow-up timing. The meta-analysis estimates relative effects of treatment combinations but cannot isolate contributions of MI-BCI, neuromodulation, or peripheral technologies nor confirm synergy. The study identifies MI-BCI strategies for further research, but current evidence doesn't support clinical recommendations or treatment priorities.

Authors’ contributions

H.Z., C.Y., and L.Y. jointly conceived and designed the study. C.K., N.S.M., and D.K. supervised the research. C.Y. developed the literature search strategy. H.Z., L.Y., and C.Y. carried out study selection, data extraction, and statistical analyses. N.S.M. and D.K. contributed to preparing the results and visualisations and supported the compilation of tables. X.Z.H., C.Y., and L.Y. drafted the manuscript. L.Z. critically reviewed and revised it, providing significant intellectual input. C.Y. and L.Y. completed the final proofreading. All authors read and approved the final manuscript.

Funding

This research was funded by the Dazhou City Key Research Base Project in Philosophy and Social Sciences (Eastern Sichuan Sports Culture Development Research Center), titled “Research on the Inheritance and Development of Ethnic Traditional Sports Culture in Eastern Sichuan within the National Fitness Programme” (CDTY202407).

Data availability

All data analysed in this study were drawn from published reports and their supplementary materials. Extracted datasets and analytical files are available from the corresponding author upon request. Some materials may be subject to third-party restrictions or institutional policies and therefore cannot be made publicly available without prior authorisation; interested researchers are advised to contact the corresponding author to request access.

Declarations

Ethics approval and consent to participate

Ethical approval and informed consent for the original trials were obtained from the respective institutional review boards, as reported in the included publications. Since this study is a systematic review and meta-analysis of published data, no new human participants were recruited, and no additional ethical approval or consent was required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.