Abstract

Background:

Brain-computer interface-based rehabilitation represents an emerging neurorehabilitation approach for post-stroke motor recovery, yet its comprehensive effects on patients in the early phase after stroke, typically defined as within 3 months of onset, remain to be fully established. This systematic review and meta-analysis evaluated effects of this intervention on upper limb function, activities of daily living, and adverse events in individuals with early stroke.

Methods:

This study was conducted following PRISMA guidelines. Eligibility criteria were established for randomized controlled trials that encompassed: (1) participants were adults (≥18 years) within 3 months of stroke onset with upper limb motor impairment; (2) interventions included brain-computer interface-based rehabilitation, and (3) outcomes that measured upper limb function, activities of daily living, and adverse events. A systematic search was performed across PubMed, Embase, Cumulative Index to Nursing and Allied Health Literature, Cochrane Library, and China National Knowledge Infrastructure databases from their inception to August 23, 2025. Two independent reviewers assessed eligibility, compiled data, and appraised methodological rigor, potential bias, and reliability of the evidence. Meta-analysis was performed using RevMan 5.4 (Cochrane Collaboration, UK) and Stata 18 (StataCorp., USA), applying random-effects models to calculate mean differences (MD) or risk ratios (RR) with 95% confidence intervals (CI). Subgroup analyses, meta-regression, sensitivity analyses, and publication bias assessments were conducted where appropriate.

Results:

Nine studies involving 642 participants (212 females and 430 males) with a mean age of 59.77 years were included. For primary outcomes, brain-computer interface-based rehabilitation significantly improved upper limb function in patients with early stroke (MD = 5.02, 95% CI: 3.20, 6.84). Subgroup analyses revealed that no statistically significant differences were observed in the improvement of upper limb functionality among various patient demographics and intervention characteristics (all p > 0.05). For secondary outcomes, the pooled analysis suggested a potential improvement in activities of daily living with BCI-based rehabilitation (MD = 7.68, 95% CI: 0.32, 15.03), although this finding was accompanied by very high heterogeneity (I2 = 88%) and was not robust in sensitivity analyses, indicating low certainty of evidence. Subgroup analyses indicated that greater benefits might be observed in patients within 30 days after stroke onset and with intervention durations not exceeding 3 weeks. Regarding safety, preliminary data from a single study suggested no significant difference in adverse events between groups (p = 0.87), but the evidence base is currently insufficient to draw firm conclusions.

Conclusions:

Brain-computer interface-based rehabilitation is effective in improving upper limb motor function in patients with early stroke. Current evidence suggests a potential benefit for activities of daily living, but the evidence is of low certainty due to substantial heterogeneity and limited robustness. Subgroup analyses identified time from onset and intervention duration as potential effect modifiers for activities of daily living. Preliminary safety data from a single study are encouraging but insufficient to establish a safety profile. Further well-designed randomized controlled trials are needed to establish optimal brain-computer interface-based rehabilitation protocols, to confirm the potential benefit on activities of daily living with more robust evidence, and to evaluate long-term efficacy and safety.

Systematic Review Registration:

PROSPERO [Register number: CRD420251144151].

1 Introduction

Stroke poses a major challenge to global health due to its high incidence, significant disability rate, and substantial socioeconomic costs, collectively contributing to a significant disease burden (Katan and Luft, 2018; Phipps and Cronin, 2020). With the improvement of medical management of early stroke, defined as within 3 months of onset, the survival rate of early stroke patients has increased significantly, but most early stroke survivors still experience persistent functional impairments (Phipps and Cronin, 2020; Stinear et al., 2020; Hilkens et al., 2024). Upper limb dysfunction stands out as one of the most common and devastating sequelae of early stroke, with previous studies indicating that approximately 60–80% of patients experience this dysfunction during the early phase (Tang et al., 2024; Persson et al., 2012; Schwarz et al., 2019). Research suggests that the critical window for neurological functional recovery typically occurs within the first 3 months post-stroke, covering the acute phase (often ≤ 30 days) and the subacute phase (generally 1–3 months after onset), during which focused interventions can significantly enhance functional rehabilitation, while 15%-30% of individuals may still experience lasting permanent dysfunction beyond this critical timeframe (Zhang et al., 2024; Patel et al., 2019; Hebert et al., 2016). It is worth noting that the dynamic process of neural plasticity and functional recovery is not homogeneous within these 3 months (Chen et al., 2026). The acute phase is often characterized by more rapid spontaneous recovery and heightened cortical excitability, whereas the subsequent subacute phase involves more experience-dependent plastic remodeling (Chen et al., 2026; Qiao et al., 2023). This indicates that functional recovery of upper limb dysfunction is particularly critical within this early post-stroke period. Therefore, this study focuses on the effects of rehabilitation interventions in patients within 3 months of stroke onset, a timeframe that captures a period of heightened neuroplasticity and potential for meaningful recovery.

Early stroke patients primarily experience upper limb dysfunction due to disruption of neural pathways caused by brain injury (Dolganov and Karpova, 2019; Geng et al., 2022; Langhorne et al., 2011). Stroke lesions can trigger a series of pathophysiological reactions like neuroinflammation, oxidative stress, metabolic abnormalities, excitotoxicity and cell apoptosis, which not only directly damage key structures such as the motor cortex and corticospinal tract, but also interrupt the nerve conduction pathways related to upper limb movement, further affecting the integration and regulation of motor sensory input, and ultimately leading to decreased motor control ability and limited upper limb movement (Geng et al., 2022; Langhorne et al., 2011; Alsbrook et al., 2023; Iadecola and Anrather, 2011). Additionally, patients in the early stroke phase often develop secondary issues such as muscle atrophy, joint contractures, and reduced activity, which can lead to abnormal movement patterns, altered muscle tone, and diminished ability to execute functional tasks, thereby exacerbating upper limb dysfunction (Langhorne et al., 2011; Pollock et al., 2014). This dysfunction presents unique challenges to both patients and clinicians. For patients, it directly impairs voluntary movement, fine motor skills, and bimanual coordination, which severely restricts independence in daily activities, limits social participation, and reduces overall quality of life, leading to increased reliance on care (Patel et al., 2019; Geng et al., 2022; Langhorne et al., 2011). For healthcare professionals, managing this condition demands close functional monitoring and personalized interventions, adding to clinical workloads and requiring sustained resources to address sequelae like spasticity and pain while preventing functional decline (Patel et al., 2019; Langhorne et al., 2011; Pollock et al., 2014). Overall, upper limb dysfunction not only delays the overall recovery process but may also lead to long-term functional impairment, thereby significantly increasing the patient's risk of disability (Langhorne et al., 2011; Pollock et al., 2014). Therefore, seeking effective rehabilitation interventions to enhance upper limb function and promote overall recovery in early stroke patients is of considerable clinical importance.

Conventional rehabilitation therapies like physical and occupational therapy primarily rely on task-oriented training, muscle strengthening, and compensatory strategies to promote functional recovery, yet they exhibit significant limitations in patients with severe motor impairments, particularly during the early stages of stroke (Hebert et al., 2016; Langhorne et al., 2011; Roesner et al., 2024). Since these methods depend heavily on residual motor function, they often fail to deliver the necessary repetitive practice when voluntary movement is minimal, and patients frequently struggle to maintain adequate training intensity due to fatigue or diminished motivation (Ambrosini et al., 2021; Lo et al., 2024). More critically, traditional rehabilitation therapies are generally unable to precisely activate impaired motor neural circuits or effectively modulate cortical excitability, thereby hindering the induction of beneficial neuroplastic changes (Sebastián-Romagosa et al., 2020; Li et al., 2025). Moreover, such programs tend to adopt standardized protocols that lack individualization, making it difficult to adapt to dynamic neural states during recovery, limiting the potential for neural remodeling, and resulting in particularly poor outcomes for fine motor skills and coordination training (Sebastián-Romagosa et al., 2020; Xie et al., 2022; Ren et al., 2024). Therefore, there is an urgent need to explore innovative rehabilitation strategies that can make up for these shortcomings.

Brain-computer interface (BCI) is an emerging neuromodulation technology, which decodes the patient's motor imagery (MI) or other neural activity intentions in real time and drives external devices, such as functional electrical stimulation (FES), soft robotic gloves or virtual reality tasks to provide immediate, closed-loop sensorimotor feedback, thereby promoting motor function reconstruction (Lo et al., 2024; Li et al., 2025; Ren et al., 2024). The core physiological mechanism is based on the Hebbian plasticity principle, utilizing repetitive task-specific brain activation training to enhance functional reorganization between damaged brain regions and motor networks, promoting coordinated activation across bilateral hemispheres, thereby inducing neuroplastic changes conducive to functional recovery (Lo et al., 2024; Xie et al., 2022; Ren et al., 2024). Building upon this technological and physiological foundation, BCI-based rehabilitation is conceptualized as a standardized neurorehabilitation treatment program that integrates an active, closed-loop BCI system (Ma et al., 2025). Its core feature is the establishment of an intention–decoding–feedback reinforcement learning loop via the BCI system, which translates the patient's actively generated motor intention (e.g., MI) in real time into multimodal sensorimotor feedback, such as completing movements via FES or robotic assistance (Ma et al., 2025; Wang et al., 2025). This approach emphasizes that the patient's active participation and effort are pivotal for driving therapeutic neuroplasticity, with the fundamental aim of strengthening or reconstructing impaired motor control pathways through high-intensity, task-specific, and intention-synchronized repetitive training (Wang et al., 2025). It thus differs from conventional rehabilitation techniques that provide primarily passive assistance or open-loop stimulation, representing a more interactive and potentially individualized strategy. Current clinical studies indicate that BCI-based rehabilitation demonstrates a positive effect on improving upper limb function in stroke patients (Brunner et al., 2024; Zanona et al., 2023; Wu et al., 2020), yet evidence regarding its impact on early stroke patients remains unclear. Therefore, it is necessary to conduct a systematic review of existing clinical studies to evaluate the efficacy and safety of BCI-based rehabilitation in these specific patients.

In recent years, the number of systematic reviews and meta-analyses on the efficacy of BCI in stroke rehabilitation has increased. Existing evidence shows that BCI-based rehabilitation can effectively improve upper limb function (Lo et al., 2024; Li et al., 2025; Xie et al., 2022; Ren et al., 2024; Baniqued et al., 2021; Carvalho et al., 2019; Peng et al., 2022; Yang et al., 2021; Shou et al., 2023; Nojima et al., 2022) and daily living activities (Li et al., 2025; Xie et al., 2022; Peng et al., 2022; Shou et al., 2023) in stroke patients. Notably, only one meta-analysis has specifically assessed the safety of BCI-based rehabilitation in stroke, reporting no significant adverse effects (Xie et al., 2022). However, these systematic reviews did not separately examine the effects of BCI-based rehabilitation on early stroke patients, instead including studies with mixed cohorts of both early and chronic stroke participants. This approach may obscure phase-specific treatment effects, as the recovery trajectory and responsiveness to intervention can differ substantially between early and chronic phases. Combining these populations may lead to heterogeneity in results and unclear conclusions, making it difficult to reveal the potential therapeutic advantages of BCI-based rehabilitation during the critical early post-stroke period. Therefore, conducting a systematic review and meta-analysis to comprehensively evaluate the effects of BCI-based rehabilitation in early-stage stroke holds significant clinical importance and research value. Such a review would focus specifically on this phase and explore potential effect modifiers, such as time since stroke onset.

Thus, the objectives of this systematic review and meta-analysis were to evaluate the effects of BCI-based rehabilitation on upper limb function, activities of daily living, and adverse events in patients with early stroke.

2 Methods

This meta-analysis was performed in accordance with the guidelines established by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (Page et al., 2021).

2.1 Eligibility criteria

The criteria for inclusion were established in accordance with the Population-Interventions-Comparison-Outcomes of interest-Study design (PICOS) framework. Specifically, it consisted of the following components: (1) participants were adults (≥18 years) within 3 months of stroke onset with upper limb motor impairment; (2) the intervention group (IG) must receive BCI-based rehabilitation, which includes specific protocols for training that utilize BCI technology to facilitate upper limb movement; (3) the control group (CG) receive either conventional rehabilitation programs without BCI-based upper limb functional rehabilitation or a placebo intervention that mimics BCI without providing actual therapeutic benefits; (4) the primary outcome should be alterations in upper limb function evaluated by the Fugl-Meyer Assessment (FMA); secondary outcomes may include activities of daily living evaluated through the Modified Barthel Index (MBI) or the Functional Independence Measure (FIM), alongside the occurrence of adverse events associated with the intervention, including but not limited to discomfort, injury, or any serious adverse events; (5) randomized controlled trials (RCTs). The exclusion criteria were delineated as follows: (1) abstracts, observational studies, qualitative studies, reviews, meta-analyses, case reports, letters, protocols, or studies not published in peer-reviewed journals; (2) duplicate or inaccessible full-text studies; (3) studies involving participants with severe cognitive impairments or other neurological conditions that may confound the results related to upper limb motor impairment; (4) studies lacking adequate specifics regarding the intervention, such as duration, frequency, and intensity; (5) studies failing to report the primary outcome; (6) studies with insufficient data for effect size (ES) and 95% confidence interval (CI); (7) studies exhibiting inadequate methodological rigor, characterized by a Physiotherapy Evidence Database (PEDro) score falling below 6 (Zhang et al., 2024; Fabero-Garrido et al., 2022; Wood et al., 2008).

2.2 Information sources

A thorough literature search was performed across five electronic databases, namely PubMed, Embase, Cumulative Index to Nursing and Allied Health Literature (CINAHL), Cochrane Library, and China National Knowledge Infrastructure (CNKI) databases, spanning from their inception until August 23, 2025. To guarantee exhaustive coverage, supplementary resources were also investigated, including clinical trial registries (ClinicalTrials.gov), websites (Google), and reference lists from all retained articles.

2.3 Search strategy

The search strategy was developed based on the predefined PICOS framework. To effectively capture the core concepts, keywords and associated terms were identified for each key component: (1) “acute stroke,” “sub-acute stroke,” “early stroke,” “cerebrovascular accident,” “stroke,” “brain vascular accident,” “brain attack,” “cerebral stroke,” “apoplexy,” “hemiplegia,” “CVA”; (2) “brain-computer interface,” “BCI,” “brain-machine interface,” “BMI,” “neural interface,” “neurotechnology”; (3) “upper limb function,” “arm function,” “upper extremity function,” “upper limb mobility,” “Fugl-Meyer Assessment,” “FMA,” “activities of daily living,” “ADL,” “Modified Barthel Index,” “MBI,” “Functional Independence Measure,” “FIM,” “adverse events,” “adverse effects,” “side effects,” “untoward effects,” “harmful effects,” “complications”; (4) “random* control* trials.” Boolean operators were systematically utilized to integrate these concept blocks, employing the OR operator to connect related concepts and the AND operator to link different concepts together. No language restrictions applied during the search process. After the initial search of the primary database, a supplementary manual search was performed across the additional resources mentioned above to guarantee the retrieval of comprehensive literature. As an example, the complete and reproducible search strategy for PubMed is presented below. Strategies for the other databases are available in Supplementary material.

(stroke [MeSH Terms] OR stroke [Text Word] OR acute stroke [Text Word] OR sub-acute stroke [Text Word] OR early stroke [Text Word] OR cerebrovascular accident [Text Word] OR brain vascular accident [Text Word] OR brain attack [Text Word] OR cerebral stroke [Text Word] OR apoplexy [Text Word] OR hemiplegia [Text Word] OR CVA [Text Word]) AND (brain-computer interface [MeSH Terms] OR brain-computer interface [Text Word] OR BCI [Text Word] OR brain-machine interface [Text Word] OR BMI [Text Word] OR neural interface [Text Word] OR neurotechnology [Text Word]) AND (upper limb function [Text Word] OR arm function [Text Word] OR upper extremity function [Text Word] OR upper limb mobility [Text Word] OR Fugl-Meyer Assessment [Text Word] OR FMA [Text Word] OR activities of daily living [Text Word] OR ADL [Text Word] OR Modified Barthel Index [Text Word] OR MBI [Text Word] OR Functional Independence Measure [Text Word] OR FIM [Text Word] OR adverse events [Text Word] OR adverse effects [Text Word] OR side effects [Text Word] OR untoward effects [Text Word] OR harmful effects [Text Word] OR complications [Text Word]) AND (randomized controlled trial [Publication Type]).

2.4 Selection process

All identified records from databases and supplementary sources were imported into EndNote 20 (Clarivate Analytics, UK) for duplicate removal. To enhance the efficiency of initial screening, a semi-automated step was implemented prior to manual review. After duplicate removal, all bibliographic records were exported to a spreadsheet (Microsoft Excel). A standardized list of exclusion keywords (e.g., review, protocol, and terms clearly unrelated to stroke or BCI-based rehabilitation) was applied to titles and abstracts. This permitted the rapid exclusion of records that were manifestly irrelevant. Two independent reviewers (JX and YM) then conducted a formal, manual screening of titles and abstracts against the predefined eligibility criteria, excluding those that failed to qualify. Full texts of potentially relevant articles were retrieved and independently assessed for inclusion by the same reviewers. The reviewers ultimately engaged in direct dialogue and meticulous proofreading of the final included studies. Any discrepancies during the selection process were resolved through discussion or by consulting a third reviewer (SS) when consensus could not be reached. The entire selection process was documented using a PRISMA flow diagram, which visually represented the number of records identified, excluded, and included at each stage, along with specific reasons for exclusion. This approach ensured transparency and clarity in the selection methodology.

2.5 Data collection process

A standardized, pre-piloted data extraction form was developed in Microsoft Excel to ensure systematic and consistent data collection. Two reviewers (QF and LH) independently extracted data from each included study. The extracted data encompassed the following key domains: study characteristics (authors and year), participant demographics (sample size, age, sex, time since stroke onset, body mass index, and baseline function score), interventional specifics (type, intensity, frequency, duration, and supervision), control composition, outcome parameters, and other relevant data as recommended by the Cochrane Handbook for Systematic Reviews of Interventions (Cumpston et al., 2019). If necessary, the authors of the primary studies would be approached by the investigator (YH) via email to request missing or unclear data. Any inconsistencies in the gathered data were corrected through mutual agreement or by seeking the expertise of a third-party reviewer (ZT).

2.6 Quality assessment

The methodological evaluation integrated three complementary tools to comprehensively assess study validity. Methodological quality of included RCTs was assessed using the PEDro scale, which is a well-validated and widely adopted tool for assessing methodological quality in physiotherapy research and is frequently employed in systematic reviews and meta-analyses (Cashin and McAuley, 2020; Moseley et al., 2019). This 11-item instrument evaluates key methodological dimensions: one item addresses external validity (eligibility criteria), eight items assess risk of bias (including random allocation, concealed allocation, baseline comparability, blinding of participants/therapists/assessors, adequate follow-up, and intention-to-treat analysis), and two items evaluate statistical reporting completeness (between-group comparisons and measures of variability) (Moseley et al., 2019). Scored on a 0–10 scale (excluding the first eligibility item), higher scores indicate superior methodological rigor (Maher et al., 2003). Based on established thresholds, studies are quality-graded as excellent (9–10), good (6–8), fair (4–5), or poor (< 4) (Gonzalez et al., 2018). This research established a methodological quality criterion for inclusion by requiring a PEDro score exceeding 6. The pre-specified threshold was established to include studies with at least good methodological quality, thereby enhancing the internal validity and reliability of the pooled effect estimates and allowing for a more accurate assessment of the intervention effect. This approach aligns with widely adopted practices in high-quality systematic reviews within the field of neurorehabilitation, striking a reasonable balance between methodological rigor and evidence representativeness (Zhang et al., 2024; Dhariwal et al., 2025). Furthermore, given that BCI-based early stroke rehabilitation constitutes an emerging and rapidly evolving field, the methodological quality of initial studies may be heterogeneous. Studies with scores below this threshold typically have deficiencies in key areas of bias control, such as allocation concealment and assessor blinding, which may affect the authenticity of efficacy assessments. Therefore, from a methodological standpoint, prioritizing the synthesis of a more reliable evidence base is particularly crucial for interpreting therapeutic effects in this domain. Concurrently, the items of the Cochrane Risk of Bias tool were employed to evaluate the risk of bias in the included studies (Cumpston et al., 2019). All assessments were documented and visualized using Review Manager version 5.4 (Cochrane Collaboration, UK) (Cumpston et al., 2019). This tool assesses six key domains of bias: selection bias, performance bias, detection bias, attrition bias, reporting bias, and other biases. Each domain was judged as “low,” “high,” or “unclear” risk of bias based on specific criteria within the tool (Cumpston et al., 2019). Additionally, the certainty of evidence for all outcomes was graded as high, moderate, low, or very low using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework through systematic evaluation of five key domains: serious risk of bias (based on PEDro scores < 6), inconsistency (I2 > 50%), indirectness of populations/interventions/outcomes, imprecision (95% CIs crossing clinical decision thresholds or optimal information size not met), and publication bias (Brignardello-Petersen and Guyatt, 2025; Gonzalez-Padilla and Dahm, 2021). Notably, each domain could be downgraded by up to two levels, and the baseline evidence quality was set as high for RCTs and as low for observational studies (Gonzalez-Padilla and Dahm, 2021; Zhang et al., 2019). Two independent reviewers (JH and JW) conducted all assessments using standardized protocols, with discrepancies resolved through consensus or third-reviewer arbitration (YS), and results were visualized through summary tables, graphs, and evidence profiles, ultimately forming a three-dimensional evaluation system of methodological quality, risk of bias, and strength of evidence.

2.7 Data synthesis and analysis

All data synthesis and analyses were performed using RevMan 5.4 (Cochrane Collaboration, UK) (Cumpston et al., 2019) and Stata 18 (StataCorp., USA) (StataCorp, 2023). Meta-analysis was only conducted when three or more RCTs reported data for a given outcome; otherwise, a descriptive synthesis was provided (Fabero-Garrido et al., 2022; Huang et al., 2025). A random-effects model was applied for all meta-analyses to account for potential clinical and methodological heterogeneity (Halme et al., 2023). Study heterogeneity was assessed using Cochran's Q statistic and the I2 index (Huedo-Medina et al., 2006). Interpretation of I2 followed Cochrane guidelines: 0%−40% (unimportant), 30%−60% (moderate), 50%-90% (substantial), 75%−100% (considerable) (Huedo-Medina et al., 2006). Statistically significant heterogeneity was defined as Q statistic p < 0.1 or I2 > 50% (Huedo-Medina et al., 2006). For continuous outcomes related to upper limb function (e.g., FMA scores), the mean difference (MD) from baseline to post-intervention and its standard deviation (SD) (i.e., change scores) were used for analysis. The pooled MD with 95% CI or standardized mean difference (SMD) with 95% CI was then calculated, depending on whether the same assessment scale was used across studies. For dichotomous outcomes (e.g., adverse events), risk ratios (RR) with 95% CI were computed. When studies reported median and interquartile range (IQR) without mean and SD, the CI calculation was derived by converting median to mean and IQR to SD using the formula SD=IQR/1.35, consistent with Cochrane methods for skewed data transformation (Cumpston et al., 2019). The specific studies and outcomes that required this data conversion were explicitly reported. Statistical significance was defined as p < 0.05 for all analyses.

To investigate possible sources of clinical heterogeneity, subgroup analyses and meta-regression assessments were conducted. Where sufficient data were available (Cumpston et al., 2019), subgroup analyses were conducted based on factors such as intervention duration, BCI type, and baseline severity. Meta-regression analyses exploring associations between the pooled ES and baseline covariates (sex distribution expressed as % female, mean patient age) were conducted when at least 10 studies reported primary outcomes, with subgroup analyses performed when study numbers were insufficient for meta-regression (Cumpston et al., 2019, 2022). Sensitivity analyses were performed to evaluate the robustness of the results through conducting leave-one-out analyses, which involve iteratively removing each study to assess its individual influence on the pooled ES. Publication bias was assessed visually using funnel plots and statistically using Egger's test if more than 10 studies were included in a meta-analysis (Cumpston et al., 2019).

3 Results3.1 Study selection

The comprehensive search process identified a total of 1071 studies, including 155 studies from PubMed, 276 studies from Embase, 140 studies from CINAHL, 445 studies from Cochrane, and 55 studies from CNKI. Additionally, 75 studies were identified from clinical trial registers. After removing 326 duplicate records, a further 153 records were excluded by the automated keyword screening process, leaving 592 records for manual screening. Following a detailed evaluation of titles and abstracts, 524 records were excluded for failing to satisfy the inclusion criteria, resulting in 68 studies that were eligible for full-text retrieval. Unfortunately, the full texts of two reports could not be retrieved. In addition to database searches, 14 studies were identified through website and citation searches, though three reports were not retrievable. Thus, 77 studies were retrieved and assessed for eligibility. Ultimately, nine studies (He et al., 2025; Zhang et al., 2025; Ji et al., 2025; Hou et al., 2024; Liu et al., 2023; Liao et al., 2023; Dun et al., 2023; Wang et al., 2024, 2022) fulfilled the eligibility criteria and were incorporated into the systematic review and meta-analysis, while the remaining 68 studies were excluded due to factors such as unsuitable patient demographics, inappropriate interventions and outcomes, as well as a PEDro score of less than 6 points. Notably, five studies (Liu et al., 2023; Liao et al., 2023; Dun et al., 2023; Wang et al., 2024, 2022) that had been included in previous systematic reviews were also identified and retrieved through the independent search. These studies underwent the same rigorous screening and eligibility assessment as all newly identified studies. Figure 1 presents the PRISMA search flow diagram, which illustrates the rigorous selection process employed to ensure high-quality evidence for the analysis.

PRISMA search flow diagram.

3.2 Study characteristics

The included nine studies (He et al., 2025; Zhang et al., 2025; Ji et al., 2025; Hou et al., 2024; Liu et al., 2023; Liao et al., 2023; Dun et al., 2023; Wang et al., 2024, 2022) were published between 2022 and 2025. Four additional studies (He et al., 2025; Zhang et al., 2025; Ji et al., 2025; Hou et al., 2024) published in 2024 and 2025 were identified, providing novel insights and emerging evidence in this field. To enable cross-study comparisons, Table 1 summarizes the demographic and baseline characteristics of participants from all included studies, whereas Table 2 provides details on intervention protocols, outcome measures, and key results.

StudySample sizeAge (years)Female (%)Time from onset (days)BMI (kg/m2)Initial FMA-UL scoreSample sizeAge (years)Female (%)Time from onset (days)BMI (kg/m2)Initial FMA-UL scoreIGCG(He et al. 2025)2558.60 ± 12.805 (20.0)33.30 ± 18.50NR23.00 ± 25.932358.30 ± 12.89 (39.1)41.80 ± 19.10NR18.00 ± 13.33(Zhang et al. 2025)2052.55 ± 11.087 (35.0)27.45 ± 11.60NRNR2051.25 ± 9.7910 (50.0)22.70 ± 7.43NRNR(Ji et al. 2025)2061.75 ± 10.358 (40.0)49.50 ± 23.3624.32 ± 2.6622.00 ± 6.301960.05 ± 14.354 (21.1)42.42 ± 24.5024.61 ± 2.5121.00 ± 4.44(Hou et al. 2024)2968.93 ± 5.2016 (55.2)42.30 ± 23.7032.67 ± 2.50NR2867.64 ± 5.2013 (46.4)46.80 ± 22.8023.41 ± 2.10NR(Wang et al. 2024)15060.00 ± 11.1140 (26.7)15.00 ± 9.63024.44 ± 3.3928.00 ± 20.0014658.00 ± 10.3731 (21.2)13.00 ± 7.4124.69 ± 3.0631.00 ± 26.67(Liu et al. 2023)3052.50 ± 10.598 (26.7)18.50 ± 9.11NRNR3053.00 ± 15.5611(36.7)18.00 ± 9.78NRNR(Liao et al. 2023)2061.50 ± 3.8012 (60.0)20.10 ± 2.20NR18.50 ± 6.462061.00 ± 3.7011 (55.0)20.60 ± 2.40NR18.70 ± 6.23(Dun et al. 2023)1253.47 ± 4.234 (33.3)51.00 ± 12.00NRNR1061.00 ± 3.004 (36.4)57.00 ± 15.00NRNR(Wang et al. 2022)2069.05 ± 5.7911 (55.0)42.00 ± 25.8023.84 ± 2.76NR2067.25 ± 4.788 (40.0)47.10 ± 23.1023.61 ± 2.30NRAverage3659.82 ± 8.3312 (39.1)33.24 ± 16.3026.32 ± 2.8322.88 ± 14.673559.72 ± 8.8411 (38.4)34.38 ± 14.6124.08 ± 2.4922.18 ± 12.67

Participant characteristics across the included studies.

IG, Intervention group; CG, Control group; BMI, Body mass index; FMA, Fugl-Meyer Assessment; UL, Upper limb; NR, Not reported in the source trial. Data are mean ± standard deviation and percentages.

StudyIntervention