We systematically searched the literature published up to September 11, 2025, across three English databases (PubMed, Web of Science, Embase) and four Chinese databases (China National Knowledge Infrastructure, SinoMed, VIP Database, and Wanfang Data). The search strategy utilized a combination of Medical Subject Headings (MeSH) and free-text terms for the core concepts of “stroke” and “brain-computer interfaces”. The search was restricted to randomized controlled trials (RCTs) published in Chinese or English. The complete search strategy for PubMed is provided in the Supplementary Table 1.

The inclusion and exclusion criteria were formulated a priori to identify studies that directly evaluated the effect of BCI-based training on lower limb recovery after stroke. Inclusion criteria were as follows: (1) Randomized controlled trials (RCTs). (2) Adult patients (age ≥ 18 years) diagnosed with stroke, confirmed by computed tomography or magnetic resonance imaging, with lower limb motor dysfunction. (3) The experimental group received BCI-based training, which could be combined with conventional rehabilitation. (4) The control group received conventional rehabilitation. (5) Studies had to report at least one of the following outcome measures related to lower limb motor function or activities of daily living: the Fugl-Meyer Assessment for Lower Extremity (FMA-LE), the Fugl-Meyer Assessment for Balance (FMA-B), Time UP and GO Test (TUGT), the Berg Balance Scale (BBS), 6-min Walk Distance (6MWD), or the Modified Barthel Index (MBI). The FMA-LE was prespecified as the primary outcome for meta-analysis given its widespread use and validation for assessing lower-limb motor recovery after stroke. Exclusion criteria included: (1) Duplicate publications. (2) Non-clinical studies (e.g., reviews, meta-analyses, animal studies, case reports). (3) Full-text unavailable after exhaustive search. Two reviewers (CS and JX) independently assessed the studies for inclusion, with any discrepancies resolved by consensus.

Data extraction was performed by two independent reviewers (CS and JX). The following data were extracted from each included study: first author's name, year of publication, participant characteristics, detailed descriptions of both experimental and control interventions, specific BCI paradigms, outcome measures, and follow-up assessments when available. When studies reported mean change scores and standard deviations, these values were extracted directly. When change scores were not explicitly provided, they were calculated using the following formulas based on the Cochrane Handbook for Systematic Reviews of Interventions (18). In cases where studies reported median and interquartile range, these values were converted to mean and SD estimates using established transformation methods: mean ≈ median and standard deviation ≈ IQR × 1.35. Any divergences were settled through deliberation, and unsettled matters were assessed and resolved by a third reviewer (YH) in order to achieve final consensus.

Two reviewers (CS and JX) independently assessed the methodological quality and risk of bias of the included studies using the Physiotherapy Evidence Database (PEDro) scale. This validated tool, widely used in systematic reviews, evaluates clinical trials through 11 criteria encompassing randomization, allocation concealment, blinding, dropout rates, intention-to-treat analysis, and statistical reporting (19). The first item is related to external validity and is not included in the total score. The remaining 10 items are each scored 1 point if clearly satisfied, yielding a total score ranging from 0 to 10. Based on this score, studies were classified as follows: 9–10 (excellent), 6–8 (good), 4–5 (fair), and below 4 (poor).

The search strategy yielded 2,278 articles from various sources: 512 from PubMed, 988 from Web of Science, 101 from Embase, 224 from China National Knowledge Infrastructure, 135 from SinoMed, 123 from VIP Database, 195 from Wanfang Data. After removing duplicates and, the titles and abstracts of 1,508 articles were screened for possible inclusion. Subsequently, 63 articles were assessed for eligibility through full-text screening. Finally, 13 articles were included in the meta-analysis (24–36). The flowchart of the search strategy and selection steps is presented in Figure 1.

Table 1 summarizes the baseline characteristics of the included studies, including details of the participants, interventions, and outcome measures. This systematic review and meta-analysis included 13 randomized controlled trials (RCTs), all conducted in China, with 12 published in Chinese and 1 in English. A total of 582 stroke patients were involved, with the mean age ranging approximately from 44 to 66 years across the studies. The experimental interventions consisted of BCI combined with conventional rehabilitation or other physical therapies (e.g., robotics, exoskeletons, rTMS), while the control groups received the corresponding therapy without the BCI component. Few studies clearly reported the specific disease stage of the included stroke patients. Moreover, all outcome measures were assessed using validated instruments and standardized methods. The included studies demonstrated a moderate to good level of methodological quality (Table 2).

Five studies involving 246 participants reported that BCI-based training significantly improved MBI scores compared with control group (MD = 6.72, 95%CI: 1.74–11.69, P = 0.008) (Figure 4). Heterogeneity assessment indicated considerable heterogeneity across the studies (I² = 96%, P < 0.00001). The leave-one-out sensitivity analysis confirmed that the pooled results were robust, however, substantial heterogeneity persisted.

Visual inspection of the funnel plot for the primary outcome revealed a symmetric pattern with no marked asymmetry (Figure 5). These results suggest a low probability of significant publication bias, though we recognize the limitation of a small number of studies for a conclusive assessment.

There were no adverse events documented in any of the included studies.

FMA-LE is an internationally recognized gold standard for assessing the recovery of limb motor function after stroke. This meta-analysis included 13 studies that evaluated the effect of BCI on FMA-LE scores. The pooled results showed that BCI intervention significantly increased patients' FMA-LE scores compared to the control group, with no heterogeneity observed among the studies. This indicates that BCI training is clearly effective in improving lower limb motor impairment in stroke patients. The improvement in lower limb motor function following BCI therapy depends on its task-specific closed-loop modulation that precisely targets the remodeling of cortico-spinal pathways governing gait and postural control (37–40). This mechanistic perspective is corroborated by a recent comprehensive review covering advances in BCI technology from 2023 to 2024, which highlights that BCI systems facilitate cortical–spinal pathway remodeling through intention-driven, closed-loop paradigms, thereby enhancing motor recovery after neurological insults (41). By decoding movement intention from the sensorimotor cortex to trigger exoskeletons or functional electrical stimulation for performing task-oriented actions like weight-bearing and stepping (42), the system establishes direct efferent-afferent coupling (43, 44). This process not only enhances coordinated activation between the primary motor cortex and motor planning regions (45). More critically, delivers precisely timed proprioceptive feedback that is spatially congruent with the intended movement (42, 46, 47). This exact spatiotemporal correspondence is considered pivotal for optimizing the function of central pattern generators at the spinal level and rebalancing interhemispheric inhibition (40, 45, 48), thereby effectively promoting the functional reorganization of neural networks essential for weight-supported standing and coordinated locomotion (49–52).

Balance dysfunction is a common issue after stroke, severely affecting patients‘ transfer ability and walking safety (53). This analysis used the BBS for assessment, and the results showed that BCI training significantly improved patients' balance function. The positive effect on functional mobility was further supported by the single study reporting the Timed Up and Go Test, which demonstrated significantly greater improvement in the BCI group compared to the control. The underlying mechanism likely involves the BCI system promoting the functional reorganization of cortical areas critical for postural control, by reinforcing the closed-loop integration of balance intention and successful execution (54–56). However, this result exhibited substantial heterogeneity. Through leave-one-out sensitivity analysis, we identified the study by Wan et al. (35) as the primary source of heterogeneity. Excluding this study significantly reduced heterogeneity (I² = 0%). This might be due to differences in the “pedaling training” mode used in that study compared to the posture control and weight-bearing focused balance training modes in other studies (57), leading to inconsistent effects on balance improvement. After excluding this study, the beneficial effect of BCI on balance remained significant, indicating that the conclusion that BCI improves postural stability and dynamic balance by enhancing neural control over trunk and proximal lower limb muscles is reliable (58).

For the FMA-B and 6MWD, each outcome was reported by only a single study. No significant between-group differences were found for either measure. The current evidence remains insufficient to draw definitive conclusions regarding the effects of BCI on these specific aspects of balance and walking endurance.

The MBI is an important indicator for assessing patients' ability to perform ADL independently (59). This meta-analysis indicates that BCI training can improve patients' MBI scores, suggesting that BCI intervention helps enhance patients' independent living ability (48). However, this result exhibited extremely high heterogeneity. Our subgroup analyses provided a crucial insight into this issue, revealing that the cumulative treatment time was a significant moderator (P = 0.02). Interventions exceeding 500 min demonstrated a large, significant, and consistent effect on MBI, whereas shorter interventions yielded an uncertain and highly variable effect. This finding suggests that the observed overall heterogeneity largely stems from the dose-response relationship of BCI therapy. Other factors, such as differences in BCI systems, patient characteristics (e.g., time post-stroke, baseline disability), and varied rehabilitation foci, likely contribute further to the heterogeneity (45). Therefore, it can be concluded that while BCI shows promise in improving ADL, a sufficient cumulative dose is critical to achieving a reliable benefit, and the optimal parameters for different patient subgroups require further refinement.

We performed a set of subgroup analyses to identify the sources of heterogeneity and how stable the interventions' effects were. For the FMA-LE and BBS outcomes, parameters including treatment dosage patient age and crucially, the specific BCI paradigm, did not exhibit significant moderating effects. These non-significant findings firmly back up the primary results for these outcomes, suggesting that BCI is stably effective in improving motor impairment and balance across a range of clinical protocols age groups and technological implementations. This consistency supports the view that the core therapeutic element is the “intention-decoding–closed-loop feedback” process itself, which modulates corticospinal circuits irrespective of the peripheral actuator used. In contrast, for the MBI outcome, both cumulative treatment time and BCI paradigm significantly influenced results. The former indicates that the translation of BCI therapy into gains in daily living activities is highly dependent on receiving an adequate cumulative dose. These subgroup results sharpen the clinical application of BCI: its benefits for motor and balance function prove versatile and widely applicable, even across different stages of recovery, while its impact on ADL calls for a more intensive, structured approach to dosage. The latter indicates that the translation of neuromodulatory gains into real-world function depends critically on the specific nature of the peripheral interface. In this analysis, BCI+FES—which provides physiologically matched, somatotopic feedback-demonstrated a clear advantage over BCI+Active Cycling for improving activities of daily living. This refines the clinical application: while benefits for basic motor function are generalizable, optimizing ADL outcomes requires not only sufficient dosage but preferably pairing with a peripheral interface like FES that enables direct functional integration.

It is important to acknowledge the limitations of this study:

This study confirms BCI as an effective and reliable tool for improving lower limb motor and balance function in stroke patients. Most critically, the uncovered dose-response relationship for MBI (activities of daily living) offers a critical reference point for clinical practice, indicating that cumulative intervention plans of no fewer than 500 min prove critical to maximizing functional recovery. Future work should focus on precisely defining the minimum and optimal effective doses of BCI, and exploring patient subgroups most likely to benefit, as suggested by the current trends. Additionally, confirming these findings in large-scale, multicenter international trials is essential to establish their generalizability across diverse healthcare systems, rehabilitation protocols, and patient populations. Advancing the standardization and long-term efficacy evaluation of BCI technology are equally critical next steps. Furthermore, it is noteworthy that among the 13 included RCTs, only 4 were prospectively registered in the Chinese Clinical Trial Registry, while the remaining 9 did not provide registration information. Moving forward, prospective registration of clinical trials should be emphasized as a fundamental practice to enhance transparency, reduce reporting bias, and improve the overall methodological quality of future BCI research.