Brain-computer interface for upper limb functional recovery post-stroke: a meta-analysis of improvement in upper limb motor function and changes in neurophysiological markers Authors: Huanhuan Zhang, Runzhi Xian, Yuchi Zhang, Xiaoyu Liu, Guoqiang Yang, Guiquan Chen Published: 2026 ABSTRACT BackgroundBrain-Computer Interface (BCI) has emerged as a promising intervention, facilitating recovery of upper limb motor function, enhancing associated neurophysiological markers, and improving activities of daily living (ADL) performance, which is measured using the Modified Barthel Index (MBI). Nevertheless, most existing research has centered on individual outcome domains, and thus critical questions remain unanswered. These questions include heterogeneity in treatment efficacy across different post-stroke phases, in addition to the consistency of BCI’s effects across specific functional assessment scales, such as the Fugl-Meyer Upper Extremity Scale (FMA-UE) and the Action Research Arm Test (ARAT)—and on neurophysiological markers.ObjectiveTo perform a systematic review and meta-analysis of the effects of BCI-mediated rehabilitation versus conventional rehabilitation on upper limb motor function, functional activities, neurophysiological markers, and MBI performance in patients with stroke.MethodsSystematic searches were performed in 7 databases for this study: PubMed, Embase, Cochrane Library, Web of Science, China National Knowledge Infrastructure (CNKI), Wan Fang Data, and China Biology Medicine Database (CBM). Inclusion criteria were randomized controlled trials (RCTs) comparing BCI-mediated rehabilitation with conventional rehabilitation in patients with upper limb dysfunction following stroke. The primary outcome measure was FMA-UE scores. Secondary outcome measures included functional activity levels as assessed by ARAT, neurophysiological markers (e.g., motor evoked potentials, MEPs, electroencephalography, EEG indices), and MBI scores. The methodological quality of the included studies was assessed using the validated evidence-based Cochrane Collaboration’s RoB 2.0. In the meta-analysis, if the results of the heterogeneity test were less than 25%, the fixed-effect model was employed; otherwise, the random-effects model was used. For each outcome measure, the mean difference (MD) or standardized mean difference (SMD), together with their corresponding 95% confidence intervals (95% CI), were calculated.ResultsAfter screening against the inclusion and exclusion criteria, 16 eligible randomized controlled trials (RCTs) were ultimately included, encompassing a total of 1761 patients. Compared with conventional rehabilitation interventions, BCI rehabilitation interventions exert a certain effect on improving patients ‘upper limb motor function (SMD = 0.49, 95%CI;0.26–0.73, p = 0.037).Additionally, significant improvements were observed in ARAT scores (SMD = 0.53, 95% CI: 0.14–0.91, p = 0.035),neurophysiological markers(SMD = 0.53, 95% CI;0.08–0.98, p < 0.05)and MBI scores(SMD = 0.85, 95% CI;0.55–1.14, p < 0.001).Subgroup analyses of BCI-mediated rehabilitation interventions for upper limb motor function recovery in stroke patients demonstrated certain subgroup-specific disparities in the magnitude of therapeutic effects across different subgroups. With respect to stroke staging, BCI-mediated rehabilitation interventions conferred superior efficacy for motor function recovery in patients with subacute stroke, a finding plausibly attributable to the temporal course of post-stroke neural remodeling. Subgroup analyses stratified by intervention dosage demonstrated that both 10–15 and 20–30 sessions of BCI-mediated rehabilitation interventions resulted in clinically meaningful improvements in patients’ upper limb function. No statistically significant difference was detected between these two treatment regimens, yet the 20-30-session protocol was associated with a relatively larger effect size. Furthermore, subgroup analyses stratified by intervention modality demonstrated that no statistically significant differences emerged across the distinct intervention approaches, and this observation plausibly suggests the rehabilitative benefits of divergent BCI paradigms in patients are uniformly mediated through neural circuit remodeling.ConclusionBCI emerges as a promising therapeutic modality for stroke rehabilitation, delivering substantial, statistically significant improvements across multiple key outcomes, including FMA-UE, ARAT, neurophysiological markers, and MBI scores with consistent statistical significance observed across all aforementioned domains.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12984-026-02080-w. Introduction On a global scale, stroke is characterized by exceptionally high incidence and disability rates and has emerged as a major public health concern [1]. After a stroke, dysregulation of neuromuscular function leads to the development of motor dysfunction. This condition is typically characterized by muscle weakness, motor coordination deficits, and abnormal muscle tone [2–4]. These motor-related impairments pose considerable challenges to post-stroke rehabilitation. Furthermore, post-stroke neurological dysfunction impairs both sensorimotor and neurocognitive functions, which subsequently has a significant adverse impact on patients ‘ability to perform daily activities and participate in social interactions [5, 6]. Therefore, post-stroke recovery of upper limb motor function ranks among the core objectives of rehabilitation, and neurophysiological measures serve as critical biological markers for predicting post-stroke neural functional reorganization [7–9]. Consequently, improvements in upper limb motor function and neurophysiological parameters carry considerable clinical value for the refinement of clinical rehabilitation strategies. In clinical practice, various programs for stroke rehabilitation are mainly carried out around two core paths, namely functional reconstruction and compensation replacement [10]. Reconstruction refers to the re-establishment of the basic exercise mode through conventional physical rehabilitation; however, when the reconstruction method is not feasible in clinical practice, the achievement of such exercise modes can only rely on compensation or alternative strategies [11]. Therefore, conventional physical rehabilitation currently has inherent limitations in the therapeutic effect of stroke rehabilitation, and it is difficult to directly induce neurophysiological remodeling after stroke. In contrast, as a cutting-edge neurological rehabilitation technology, the brain-computer interface (BCI) can bypass the foci and establish a connection between the brain and the external environment [12]. This technology has a dual function: it can not only promote the rehabilitation process, but also realize the functions of compensation and substitution, while providing real-time feedback to encourage patients to actively participate in rehabilitation training [13, 14]. Currently, BCIs are classified into invasive (iBCI) and non-invasive (nBCI) types [15], From the perspective of the mechanism of action, both types of brain-computer interfaces play a role by establishing direct brain-device communication and providing real-time sensory motor feedback, thereby enhancing neuroplasticity and promoting the improvement of upper limb motor function and neurophysiological indicators [16]. However, there are significant differences between these two types of technologies in terms of technical characteristics, safety and clinical practicality. The invasive brain-computer interface requires electrodes to be implanted in the cerebral cortex, and there are certain surgical risks and complications; however, this technology has better signal quality and space-time resolution, which is convenient for capturing higher-frequency EEG signals [17]. Such strengths support the accurate decoding of motor intentions and targeted regulation of impaired neural circuits [18, 19]. nBCIs—including those based on EEG, FNIRS, or SEMG—employ scalp electrodes or non-contact sensors, obviating the requirement for surgery and providing excellent safety, ready accessibility, and favorable tolerability [20–22]. Although nBCIs demonstrate lower signal quality and spatiotemporal resolution relative to invasive BCIs, they have been more widely adopted in clinical practice owing to the absence of surgical intervention, showing to improving upper limb motor function across different post-stroke phase via subgroup analysis. Therefore, the objective of this meta-analysis is to systematically evaluate the effects of BCI interventions compared to conventional rehabilitation on upper limb motor function (assessed by Fugl-Meyer Assessment and Action Research Arm Test), activities of daily living (Modified Barthel Index), and neurophysiological markers in post-stroke patients, and to explore the influence of stroke phase, intervention frequency, and BCI modality on treatment efficacy. Data and methods Data sources We conducted a systematic review and meta-analysis to investigate whether brain-computer interface (BCI) interventions can improve upper limb motor function, MBI, and neurophysiological parameters in stroke survivors. To ensure methodological rigor, the protocol for this systematic review was registered on the International Prospective Register of Systematic Reviews (PROSPERO) on 6 November 2025 (registration number: CRD420251185457). This study was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement and the methodological guidelines of the Cochrane Handbook for Systematic Reviews of Interventions (Version 6). This analysis was carried out in strict accordance with the pre-registration program. Only one outcome indicator has been revised before the analysis was implemented, and the revision has been recorded on the PROSPERO platform. Search strategy In order to comprehensively screen all relevant studies, this study formulated and implemented a comprehensive strategy for multi-database joint retrieval, and followed the optimal method to improve the sensitivity and specificity of retrieval. The retrieval work covers the four major English databases (PubMed, Embase, Cochrane Library, Web of Science) and the three major Chinese databases (China National Knowledge Network, Wan fang Data, China Biomedical Literature Database) to ensure the comprehensive coverage of relevant literature at home and abroad. The retrieval time limit is set from November 6, 2025 for the construction of each database, and there is no starting time limit. The search terms adopt the strategy of combining thematic words and free words, and combine the medical subject words list (Mesh) and the exclusive search rules of each database to optimize and combine the search terms. Search terms were derived from the following conceptual domains: brain-computer interfaces (BCI), stroke, and rehabilitation. The final search strategies were optimized through the application of Boolean operators, truncation, and field-specific restrictions, thus ensuring a comprehensive search with minimal omission risk. Eligibility criteria and study selection Inclusion and exclusion criteria were pre-specified according to the PICOS framework. The following criteria were applied: (1) participants were stroke patients; (2) the intervention was BCI-based therapy; (3) the comparator was conventional rehabilitation or a sham control, with matched intervention intensity; (4) outcomes included the FMA-UE and at least one secondary measure (ARAT, neurophysiological markers, or MBI); and (5) the study design was an RCT. Studies were excluded if they: (1) were not randomized controlled trials; (2) involved patients with comorbid severe neurological disorders; (3) did not report Fugl-Meyer Assessment for Upper Extremity (FMA-UE) scores; (4) had incomplete data preventing effect size calculation; (5) were duplicate publications; (6) were animal studies, in vitro studies, or involved healthy participants; or (7) had unavailable full texts or critical methodological flaws (e.g., unreported randomization procedure, irreconcilable data inconsistencies). Study eligibility was independently assessed by two reviewers (HHZ and RZX), with any discrepancies resolved through discussion or by consultation with a third reviewer. Data extraction Data extraction was performed using a standardized protocol in order to ensure accuracy and completeness and to mitigate potential biases. After a pilot test to calibrate criteria and data item definitions, two researchers independently extracted data using a standardized form. This pilot test was implemented to align the understanding of inclusion/exclusion criteria and data item definitions prior to formal extraction. A standardized data extraction form was developed to systematically collate relevant data. The form encompassed study characteristics, participant demographics, details of interventions and controls, outcome data, and methodological quality items. Methodological quality assessment of included studies This study used the RoB 2.0 bias risk assessment tool to conduct a methodological quality evaluation of all included randomized controlled trials (RCTs). This tool is a verified standardized evaluation tool that provided a special bias evaluation framework for randomized controlled trials, covering all core methodological areas related to clinical research in the field of post-stroke rehabilitation. Three professionally trained researchers used this tool to independently complete the bias risk assessment. After completing a pre-calibration tests of the three studies to ensure the consistency of the evaluation criteria, the researchers then conducted a formal evaluation of all the included tests. In the formal evaluation stage, two researchers independently evaluated the grade and cross-checked the evaluation results. Any differences were decided by the third researcher, and the final bias risk level of each study was determined by consensus. The results of bias risk assessment were visualized in the form of a summary chart and a risk of bias graph, which provides a methodological basis for interpreting the subsequent meta-analysis results and exploring the source of heterogeneity. Statistical analysis methods This meta-analysis aimed to evaluate the efficacy of BCI interventions. The main outcome indicators were upper limb motor function, which was evaluated by the upper limb Fugl-Meyer Assessment; secondary outcome indicators included ARAT score, neurophysiological index and MBI score. All statistical analyses in this study were performed using Stata 18.0 and Review Manager (RevMan) 5.4 software, and strictly followed the methodological guidelines of the Cochrane Interventional System Review Manual (6th Edition). In the data preprocessing stage, the continuous data reported in the form of quartile spacing (Q1, Q3) in some studies were converted to mean ± standard deviation (mean ± SD) format. For outcome indicators with different naming methods but equivalent functional attributes, they were combined and classified after confirming that their measurement connotations were consistent. The therapeutic effect was pooled using the standard mean difference (SMD) as the effect index, and the 95% confidence interval (CI) was calculated. Heterogeneity was quantitatively analyzed using I² statistics and the Cochran Q test; if I² ≥ 25% or the Q test P value was < 0.10, the random effects model was used for data consolidation; on the contrary, the fixed effect model was used. This study conducted a sensitivity analysis to evaluate the robustness of the combined results by eliminating single studies one by one. Prespecified subgroup analyses were performed based on total intervention frequency and stroke phase (acute, subacute, chronic). Publication bias was evaluated using funnel plots, Egger’s test, and the trim-and-fill method. A two-tailed P < 0.05 was considered statistically significant. Results Literature retrieval and screening were conducted in strict accordance with the PRISMA Statement guidelines. The databases searched included PubMed (n = 1043), Web of Science (n = 1309), Embase (n = 54), Wan Fang Data (n = 20), China Biomedical Literature Database (SinoMed, n = 15), and China National Knowledge Infrastructure (CNKI, n = 1), with an initial search yielding a total of 2442 relevant articles. We removed a total of 602 records (including duplicates, books, conference proceedings, reviews, study protocols, and meta-analyses) using EndNote software, with 1840 articles remaining for title and abstract screening. We excluded 1725 articles for inconsistency with the study topic, and ultimately, 115 articles were subjected to full-text eligibility assessment. During full-text screening, we excluded 56 reviews or meta-analyses, 25 clinical trial registrations, 5 non-randomized controlled trials (non-RCTs), 5 studies focusing on lower limb rehabilitation, 4 case reports, 2 articles with unavailable full texts, and 2 articles without publication dates. Ultimately, 16 RCTs meeting the inclusion criteria were included in this meta-analysis. The flow diagram illustrating the search strategy and study selection process is shown in Figs. 1 and 2. study selection process. Fig. 1diagram illustrating the search strategy diagram illustrating the search strategy Fig. 2study selection process study selection process Methodological quality and risk of bias The risk of bias for the 16 included studies was assessed using the Cochrane-recommended Risk of Bias tool (RoB 2.0). The eligible included trials exhibited robust methodological quality across key domains, including the integrity of outcome measure data and selective reporting, with the overall methodological rigor of the meta-analysis well substantiated (Fig. 3 Risk of bias graph and Fig. 4 Risk of bias summary). For random sequence generation, all 16 eligible trials (100%) were rated as low risk of bias. For allocation concealment, 14 eligible trials (87.5%) were rated as low risk of bias. Only two trials (Guo [23]; Liang [24]) were judged to be at unclear risk of bias, owing to the absence of specific descriptions regarding their allocation concealment methodologies. No trials were rated as high risk of bias, with the overall risk profile being well controlled. For the domains of integrity of outcome measure data, selective reporting, outcome assessor blinding, and other sources of bias, all 16 eligible trials (100%) were rated as low risk of bias. However, for the domain of participant and personnel blinding, a high proportion of trials (14 out of 16, 87.5%) were rated as high risk of bias. Such elevated risk stemmed from the technical attributes of BCI-based interventions, which necessitate specialized equipment (e.g., EEG electrode caps) and active engagement of patients in therapeutic rehabilitation training. As a result, blinding of participants and interventionists through placebo devices was not achievable, thereby rendering intergroup interventions readily distinguishable. Notably, this represents an intrinsic risk inherent to the BCI rehabilitation field, rather than a methodological flaw in study design. Fig. 3Risk of bias graph Risk of bias graph Fig. 4Risk of bias summary Risk of bias summary Effects of BCI-based training on FMA-UE in stroke survivors Compared to conventional rehabilitation, BCI-based interventions yielded a statistically significant and moderate improvement in FMA-UE scores (pooled random-effects SMD = 0.49, 95% CI: 0.26 to 0.73, p = 0.037), favouring BCI. The pooled estimate was robust to heterogeneity. The forest plot for this analysis is presented in Fig. 5 forest plot for the FMA-UE. The presence of moderate statistical heterogeneity (I² = 43.4%, p = 0.037) justified the use of a random-effects model. In order to verify the robustness of the combined results, this study carried out a sensitivity analysis to eliminate single studies one by one. The results showed that after excluding the study of Wang [25], the range of the combined effect is narrowed to 0.22–0.73 (Fig. 6 Sensitivity Analysis for the FMA-UE), suggesting that the study had a great impact on the overall heterogeneity of this analysis. In addition, the analysis results of the funnel diagram showed that the effect points of most studies are distributed within the within the 95% confidence interval confidence interval (Fig. 7 Funnel Plot for the FMA-UE), showing only mild asymmetry. This result confirmed that there was no significant publication bias in this study, and further verified the robustness of the combined results. Fig. 5Forest plot for the FMA-UE Forest plot for the FMA-UE Fig. 6Sensitivity analysis for the FMA-UE Sensitivity analysis for the FMA-UE Fig. 7Funnel Plot for the FMA-UE Funnel Plot for the FMA-UE Subgroup analyses by stroke phases, frequencies and modalities To clarify the effects of distinct stroke phases, varying intervention frequencies and diverse BCI intervention modalities on upper limb motor function recovery, a meta-analysis was performed on the 15 enrolled randomized controlled trials (RCTs) (Fig. 8 Subgroup Analysis of the FMA-UE). Fig. 8Subgroup analysis of the FMA-UE Subgroup analysis of the FMA-UE The analysis findings demonstrated (Fig. 9 Subgroup Analysis of FMA-UE by Stroke Stage) that BCI interventions exerted ameliorative effects for patients across all stroke phases, with the most significant improvement in FMA-UE scores observed in subacute stroke patients (SMD = 0.62, 95% CI; 0.20 to 1.04, I² = 59.3%, p = 0.031). Among chronic stroke patients, the meta-analysis yielded a non-significant effect size in favour of BCI (SMD = 0.40, 95% CI; 0.08 to 0.72, p = 0.152). Although the upper bound of the CI approached significance and heterogeneity was low to moderate (I² = 34.7%), the overall effect did not reach statistical significance. In the acute phase, the analysis showed a non-significant point estimate with high uncertainty (SMD = 0.45, 95% CI; -0.07 to 0.96). Although the p-value of the correlation analysis in this study was < 0.001 and there was no statistical heterogeneity (I²=0.0%), the confidence interval range was relatively wide and the ineffective values were spanning. This limitation was caused by the small sample size, so it was impossible to draw an exact conclusion about the effectiveness of the intervention [9]. Fig. 9Subgroup analysis of FMA-UE by stroke stage Subgroup analysis of FMA-UE by stroke stage This study carried out a stratified subgroup analysis according to the number of intervention courses, aiming to evaluate its impact on the recovery of upper limb motor function (Fig. 10 Subgroup Analysis by Number of BCI Sessions). Among them, the subgroup analysis results of 20 to 30 courses of intervention showed that the patient’s upper limb FMA-UE score showed a statistically significant improvement (SMD = 0.58, 95% CI; 0.21 to 0.94); this analysis was characterized by moderate heterogeneity (I² = 60.9%, p = 0.009). For the subgroup of 10–15 sessions, a significant yet smaller effect was found (SMD = 0.37, 95% CI; 0.09 to 0.64). No heterogeneity was detected in this subgroup (I² = 0.0%, p = 0.673). These findings indicate that, within the investigated range, increasing the number of BCI sessions may enhance upper limb functional recovery. Fig. 10Subgroup analysis by number of BCI sessions Subgroup analysis by number of BCI sessions Subgroup analysis was performed to compare different BCI modalities (Fig. 11 Subgroup Analysis by Type of BCI Intervention), Among the BCI modalities analyzed, MI-BCI with FES showed the highest effect size (SMD = 0.60, 95% CI; 0.08 to 1.13, I² = 33.5%, P = 0.025), followed by Non-MI BCI (SMD = 0.50, 95% CI; 0.04 to 0.96, I² =58.4%, P = 0.026) and MI-BCI combined with conventional rehabilitation (SMD = 0.46, 95% CI; 0.13 to 0.79, I² = 7.1%, P = 0.007). These findings suggest that MI-BCI protocols incorporating FES should be a priority for future investigation. Fig. 11Subgroup analysis by type of BCI intervention Subgroup analysis by type of BCI intervention Effects of BCI on ARAT scores in stroke survivors To evaluate upper limb function using the common clinical tool ARAT, a meta-analysis was performed based on 6 randomized controlled trials (Fig. 12 Forest plot for the ARAT). Given the presence of moderate heterogeneity (I² = 58.2%, p = 0.035), a random-effects model was used to pool the effect sizes. Meta-analysis indicated a significant benefit of BCI over conventional rehabilitation on ARAT scores, with a moderate effect size (SMD = 0.53, 95% CI; 0.14 to 0.91, p = 0.035). These findings demonstrate the efficacy of BCI intervention for improving upper limb motor function post-stroke. Fig. 12Forest plot for the ARAT Forest plot for the ARAT Effect of BCI on MBI scores in stroke patients Activities of daily living (ADL) were assessed using the MBI. A meta-analysis of the six included randomized controlled trials was performed, the results of which were presented in Fig. 13 Forest plot for the MBI. The MBI score of patients in the BCI intervention group was significantly higher than that of patients in the conventional rehabilitation control group (fixed-effect model; SMD = 0.86, 95% CI; 0.59 to 1.12). The analysis results showed that the heterogeneity between studies is low (I² = 21.5%, p = 0.272), suggesting that the results of each study were consistent and supporting the suggesting that BCI intervention can may improve ADL in stroke patients. Fig. 13Forest plot for the MBI Forest plot for the MBI BCI intervention on post-stroke neurophysiological markers One aspect that needs to be focused on is the role of brain-computer interface (BCI) on neurophysiological indicators. Such indicators can reflect neuroplasticity and reveal the mechanism of motor function recovery in stroke patients. This study carried out a meta-analysis of the neurophysiological outcome indicators in the seven randomized controlled trials (RCTs) included (Fig. 14 Forest plot of neurophysiological outcomes). Meta-analysis using a random-effects model yielded a significant positive effect size (SMD = 0.53, 95% CI; 0.08 to 0.98, p = 0.021), The analysis results showed that BCI intervention produced benign changes in neurophysiological outcome indicators. Regardless of the stage of the stroke (subacute or chronic phase), neurophysiological indicators showed a positive response to BCI intervention, and no significant differences were found between different subgroups (p = 0.824), as detailed in Fig. 15 Subgroup analysis of neurophysiological outcomes by stroke stage. It should be noted, however, that substantial heterogeneity was present within the chronic-phase subgroup (I² = 72.8%). In view of the moderate degree of heterogeneity, this study further carried out sensitivity analysis. The results showed that the combined effect has always been stable in the range of 0.3–0.6, and the 95% confidence interval of all results did not contain invalid values, suggesting that the main conclusions of this study had good robustness (Fig. 16 Sensitivity Analysis of Neurophysiological Outcomes). Fig. 14Forest plot of neurophysiological outcomes Forest plot of neurophysiological outcomes Fig. 15Subgroup analysis of neurophysiological outcomes by stroke stage Subgroup analysis of neurophysiological outcomes by stroke stage Fig. 16Sensitivity analysis of neurophysiological outcomes Sensitivity analysis of neurophysiological outcomes Discussion In order to comprehensively evaluate the rehabilitation effect of BCI-mediated stroke, the meta-analysis carried out in this study covered 16 randomized controlled trials (RCTs), which systematically explored their impact on rehabilitation outcome indicators from upper limb motor function, daily life activity ability to neurophysiological changes.According to the research findings, stroke patients who received BCI training showed certain degrees of change in motor function, activities of daily living, and neurophysiological indicators, suggesting a potential association between BCI training and the aforementioned improvements. The results of subgroup analysis show that the therapeutic effect of BCI varies with the stage of stroke. Subgroup analysis indicated that BCI intervention was associated with larger effect sizes in subacute stroke compared to chronic stroke.This law of change is consistent with the clear principle of neuroplasticity. There is a critical period for plasticity enhancement in the early stage after stroke, and this critical period narrows in the chronic period [26, 27]. The period of maximal neural reorganization after stroke is generally considered to occur within 2 to 8 weeks [28–31]. The mechanism of action of BCI is considered to promote targeted and adaptive functional reorganization of the brain through synergy with endogenous neural activity waves after stroke. After entering the chronic stage, neuroplasticity gradually weakened, and the pathological compensation mode has tended to be solidified [31, 32]. Therefore, BCI training may not be enough to reactivate adaptive neuroplastic changes, and it may not bring additional functional improvements compared with standard rehabilitation. The results of the subgroup analysis based on the therapeutic dose showed that the motor function of the upper limbs was improved after 10 to 15 and 20 to 30 BCI sessions in stroke patients. No statistically significant difference in effect was detected between the two regimens, suggesting comparable effects. Therefore, it is urgent to carry out research in the future to clarify the optimal number of BCI training sessions to maximize the recovery of upper limb motor function in stroke patients. Further subgroup analysis results show that the differences in the efficacy of different types of BCI intervention schemes (motor imagination brain-computer interface combined with functional electrical stimulation, motor imagination brain-computer interface combined with conventional treatment and non-motor imagination brain-computer interface) are not statistically significant. The result that there is no significant difference in efficacy may indicate that the various BCI intervention schemes tested in this study are all targeted at similar functional recovery-related neural pathways. Therefore, future research should focus on the specific parameters of the combined BCI scheme with standardization and optimization to clarify the optimal treatment mode. In addition to functional improvements, BCI training also confers measurable benefits at the neurophysiological level following stroke. The seven studies included in this analysis consistently demonstrated that BCI training enhances post-stroke neurophysiological function. Specifically, the intervention was consistently associated with the enhancement of “voluntary control of sensorimotor rhythm (SMR)” and “event-related desynchronization/synchronization (ERD/ERS)“ [33]. This improvement of regulation ability may then trigger a series of neuroplasticity changes: enhancing the activation of the interhemispheric balance restoring the balance of the bilateral hemisphere, improving the excitability of the corticospinal tract, and strengthening the functional connection of the movement-related brain space, thus jointly promoting the recovery of motor function [34–36]. The preliminary work of Ramos and Fu [37, 38] has demonstrated a strong correlation between such neurophysiological changes and clinically measurable functional improvements. In addition to standardized clinical evaluation, certain degrees of change in activities of daily living were also observed in stroke patients who received BCI intervention, suggesting that BCI may be beneficial in this regard. Notably, several included studies found a positive correlation between improvements in MBI scores and concurrent changes in neurophysiological indicators (e.g., enhanced modulation of sensorimotor rhythm, increased event-related desynchronization) [9, 39, 40]. Although correlation does not imply causation, this association may, to some extent, suggest that BCI-driven neuroplasticity could contribute to improvements in patients’ activities of daily living [41].Considering the temporal sequence, dose-response trends, and the biological plausibility of closed-loop neural feedback, these findings are consistent with a potential causal relationship, but do not prove causation. However, future studies are needed to more rigorously examine whether neurophysiological changes mediate the effect of BCI on ADL outcomes, using methods such as mediation analysis or structural equation modeling. Limitations There are some limitations in this study. First, the total number of RCTs included is only 16, and there are fewer studies that can be combined with some outcome indicators. Moreover, the majority of included studies had small sample sizes (typically fewer than 30 participants per group), which limits the precision and generalizability of the pooled estimates.Some subgroups (such as the acute phase) have a small sample size, which may weaken the statistical efficiency and robustness of subgroup analysis, all findings related to subgroups with small sample sizes should be interpreted with caution. In addition, this study did not conduct a formal quantitative evaluation of bias (such as the funnel plot regression test) for all outcomes, which is also an acknowledged constraint. Second, although subgroup analysis has been carried out according to the dose and method of intervention, the original study reported that key parameters (task difficulty, feedback mode, etc.) are very heterogeneous, so it is impossible to stratify these potential influencing factors in more detail. Finally, this article only compares BCI with conventional rehabilitation, and has not yet touched on the emerging and promising direction of “BCI combined auxiliary nutriredgulation (such as TMS/tDCS)”, and its synergy needs to be explored in the future. Conclusion The findings of this meta-analysis indicate a positive association between BCI-based rehabilitation training and improvements in upper limb motor function, activities of daily living, and certain neurophysiological indicators in post-stroke patients. Subgroup analyses suggest that patients in the subacute phase may experience more pronounced functional improvements compared to those in the chronic phase. Based on the available evidence, BCI intervention shows potential for clinical application in subacute stroke rehabilitation, and it is recommended that training protocols be individualized according to specific patient conditions. Future efforts should focus on further optimizing BCI training parameters and exploring its combination with other intervention approaches, so as to provide a more substantial basis for clinical practice. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.