The design, conduct, and reporting of this umbrella review followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines (Multimedia Appendix 1) and its PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) checklist (Multimedia Appendix 2) [24,25]. Additionally, we adhered to the PRIOR (Preferred Reporting Items for Overviews of Reviews) checklist (Multimedia Appendix 3) to ensure comprehensive reporting [26].

Given the complexity of upper limb motor function and its critical role in daily life, upper limb rehabilitation robots have reached a relatively mature stage of research, development, and application. These systems enable high-precision, high-intensity, and personalized training, which is designed to promote functional recovery, quality of life, and self-care ability. Accordingly, this review specifically investigated the effects of robotic interventions on upper limb recovery post stroke. We excluded reviews that focused only on general stroke outcomes or those evaluating robotic interventions for lower limb rehabilitation. Furthermore, included studies were required to demonstrate clinical relevance through rigorous methodological quality and provide complete datasets to support their conclusions. To account for the rapid technological evolution in this field, we limited inclusion to meta-analyses published within the past 7 years.

Eligible studies involved adult patients with stroke undergoing upper limb rehabilitation and compared interventions incorporating RAT (as a stand-alone or adjunct treatment) against conventional therapy. The outcomes focused on clinical efficacy, encompassing upper limb motor function [27], muscle strength [28], spasticity [29], and ADL [30]. Meta-analyses lacking sufficient data and non-English publications were excluded.

We performed a systematic literature search on December 26, 2025, using the Sichuan University Library Discovery System. The following databases were searched: PubMed, Web of Science, Embase, the Cochrane Library, IEEE Xplore, and Scopus. The search was restricted to peer-reviewed journal articles published in English between January 1, 2019, and the search date. The complete search syntax is detailed in Table S1 in Multimedia Appendix 4. The search strategy was developed de novo through team discussion to ensure conceptual adequacy. It was not based on any previously published strategy and was not updated after its initial execution. No supplementary search techniques were used, such as citation tracking, manual searches of grey literature, or contacting authors for unpublished data. The study protocol was prospectively registered on PROSPERO (CRD42024497183), and no additional trial registries were consulted. All retrieved records were imported into EndNote 20 for management. Duplicates were removed using the software’s automated function, supplemented by manual verification, to compile a unique set of citations for subsequent screening.

Two reviewers (XZ and LZ) independently appraised the methodological quality of each included meta-analysis using the AMSTAR 2 tool [31]. Any disagreements were resolved through discussion with a third reviewer (SL) to reach consensus. The overall quality of each study was rated as high, moderate, low, or critically low.

Two reviewers (XZ and LZ) independently performed data extraction using Microsoft Excel 2019. For the qualitative evidence synthesis, key information was extracted from each meta-analysis. This included all outcome comparisons between RAT and conventional therapy, as well as the moderators examined in prespecified subgroup analyses. The extracted moderators comprised intervention intensity, stroke stage (subacute or chronic), robot type (unilateral vs bilateral), training design (exoskeleton vs end-effector), and baseline upper limb motor impairment level.

To assess the degree of primary study overlap across the included systematic reviews, we constructed a citation matrix in tabular form and used the corrected covered area (CCA) index to quantify the overall extent of overlap, calculated using the following formula: CCA = (N – r) × (r × c) – r

where represents the total number of included publications (including duplicates), r denotes the number of index publications (the count of unique primary studies or the number of rows in the matrix), and is the number of reviews (or the number of columns in the matrix).

The CCA value was interpreted using commonly applied thresholds derived from Pieper et al [32] work: <5% (slight), 5%-10% (moderate), 11%-15% (high), >15% (very high). For meta-analyses with incomplete information, the following handling approach was adopted: (1) Maximized matching: using the available author and year information, cross-referencing and matching were performed against the references in the other 20 meta-analyses. (2) Conservative estimation: for studies that could not be uniquely matched, they were treated as independent studies in the calculation of the overall CCA to avoid underestimating overlap.

Given the methodological limitations of statistically pooling effect sizes from different meta-analyses, this study used a narrative synthesis approach to summarize and evaluate the evidence. The results for each outcome measure were categorized based on the direction of the effect and its statistical significance: if RAT was significantly superior to conventional therapy, it was classified as “Significant;” if there was no significant difference or if conventional therapy was superior, it was classified as “Non-significant.” The determination of statistical significance was based on whether the 95% CIs reported in the original meta-analyses included the null value.

In the synthesis and interpretation of the evidence, we fully incorporated considerations of heterogeneity and risk of bias. For heterogeneity, in addition to examining statistical indicators such as I², we focused on analyzing its methodological sources, including population characteristics, intervention protocols, and measurement tools. When heterogeneity was high, interpretations were made cautiously, relying on subgroup results. For the risk of bias, we considered both the methodological quality of the included meta-analyses (assessed using AMSTAR 2) and potential biases in the original studies. These factors were integrated to grade and critically discuss the evidence during result interpretation.

During the study screening phase, the publication date range specified in the registration protocol (CRD42024497183) was updated from December 30, 2013-December 30, 2023, to January 1, 2019-December 26, 2025, to obtain the most recent evidence. All other methodological aspects followed the original protocol.

From an initial pool of 1307 records identified, 21 meta-analyses [10,16,19-22,33-47] were included following deduplication and multiple rounds of screening. The study selection flow is detailed in Figure 1. The 21 meta-analyses collectively cited primary studies 591 times. After deduplication, 310 independent primary studies were obtained. Among these, multiple studies were repeatedly included in different systematic reviews. All studies that were cited 2 or more times (≥2) and their citation frequencies are summarized in Table S2 in Multimedia Appendix 5. The CCA was 4.532%, indicating a slight degree of overlap among the primary studies in this research.

This review ultimately included 21 meta-analyses [10,16,19-22,33-47], encompassing a total of 535 randomized controlled trials and involving 27,598 patients with upper limb dysfunction after stroke. The study participants included patients in the acute, subacute, and chronic phases of stroke. The types of robots included are end-effector, exoskeleton, soft robotic gloves, etc. Table 1 provides a detailed overview of all the included meta-analyses.

The methodological quality of the included reviews was assessed using the AMSTAR 2 tool. We synthesized the reported findings from these reviews rather than conducting a new meta-analysis of primary data. According to the AMSTAR 2 appraisal, 17 of the 21 included meta-analyses were of high quality [10,16,20-22,34,35,37,39-47], 3 were of moderate quality [19,33,36], and one was of critically low quality [38] (Table 2). Downgrading was primarily due to factors such as unmet justification for significant statistical heterogeneity [36] or inadequate consideration of potential biases from the primary randomized controlled trials [38]. Despite these limitations, the evidence base is reliable, as the key conclusions are underpinned by a preponderance of medium- to high-quality studies (95.2%), indicating robust primary findings [40].

The Fugl-Meyer Assessment (FMA) is the “gold standard” for assessing upper-extremity sensorimotor function, due to its ability to assess aspects such as movement within synergies, mixing synergies, reflexes, wrist, hand, grip, coordination, and speed movements, thus providing a large amount of information that is very useful for understanding the sensorimotor capacity of the affected upper limb after a stroke [48]. The FMA was adopted as the primary outcome measure in 18 out of the 21 included meta-analyses (85.7%) to evaluate the restoration of upper limb motor function.

However, the specific clinical questions explored by various studies using FMA differ significantly, covering multiple dimensions such as the immediacy and sustainability of therapeutic effects [36], responses in populations at different stroke stages [16,21], and impacts on other functional domains [35]. By synthesizing these studies, the consistent use of FMA provides highly comparable and reliable evidence for a comprehensive, multifaceted evaluation of the efficacy of robot-assisted therapy.

This study synthesizes evidence on the immediate effects of RAT from multiple meta-analyses. As shown in Table 3, various studies used standardized mean difference (SMD), mean difference (MD), weighted mean differences (WMD), and Hedges gas effect size metrics. Eight independent effect size estimates from different meta-analyses demonstrated statistically significant positive effects of RAT [20,33,34,36,42,44,45,47], supporting its immediate benefits. However, one study reported a non-significant result [43], and variations were observed in the magnitude of effects. The meta-analysis demonstrated moderate heterogeneity (I²=35%), with variations among the included studies in terms of population characteristics, device types, and treatment protocols. Moreover, the study’s own assessment indicated potential risks of bias in areas such as intervention implementation and completeness of outcome data.

Regarding the long-term maintenance effects of RAT, this review confirms that RAT demonstrates sustained short-term efficacy in improving upper limb function [20,36,42,45], as shown in Table 4. Meanwhile, our comprehensive analysis reveals a significant time-gradient effect: while significant functional improvements were observed immediately after the intervention and during short-term follow-up (≤12 weeks) [36,45], these advantages generally diminished or even disappeared during long-term follow-up (>12 weeks) [20,42]. This time-dependent pattern of efficacy decay indicates notable limitations in current robotic rehabilitation strategies for maintaining long-term therapeutic effects.

Based on the systematic synthesis of existing meta-analytical evidence conducted in this umbrella review, we are able to examine the value and position of RAT in poststroke upper limb rehabilitation from a broader perspective. Importantly, our analysis is guided by the ICF framework, which provides a structured approach to evaluating outcomes across different dimensions of recovery. The main findings, clinical implications, and future research directions of this study are outlined as follows.

This review confirms that RAT demonstrates certain efficacy in improving upper limb motor function at the ICF body functions level, as measured by the FMA [36]. However, our synthesis reveals a crucial gradient in treatment effects across ICF levels. First, the therapeutic effects show significant stage-dependency, with the most robust and consistent evidence supporting its effectiveness for patients in the chronic phase [16,34,35], while its efficacy for subacute patients remains inconsistent due to variations in intervention protocols and patient characteristics [16]. Second, a prominent finding is the limited translation of benefits to the ICF activity level. Although it effectively improves upper limb function and reduces spasticity [34,41], its effects on higher-level functions such as hand dexterity, ADL, and social participation are limited and inconsistent [22,41].

Specifically, when assessing manual dexterity, studies using the Nine-Hole Peg Test demonstrated that the robot-assisted group completed tasks significantly faster than the conventional therapy group [22,38], whereas studies using the Box and Block Test found no significant advantage [41]. More importantly, although RAT performed no worse than conventional therapist-guided training in improving both ADL and enhancing social participation, neither domain showed statistically significant superiority for the robotic approach [22,41]. This dissociation between body functions and activity-level outcomes represents a key challenge in robotic rehabilitation.

Within the ICF framework, we identified several critical moderators that explain the substantial heterogeneity in treatment effects. Intervention intensity is a critical factor influencing therapeutic efficacy. Zhang et al [20] found that when the total treatment duration exceeded 15 hours, the RAT group showed significant improvements in both motor control and functional activity. However, Yang et al [16] reported that excessively long treatment cycles may lead to diminishing returns. Notably, Carrillo et al [19] evaluated the impact of different training intensities on FMA scores and demonstrated that higher-intensity interventions were most beneficial for functional recovery. Furthermore, multiple dimensions of training intensity (session duration, frequency, total cycle) need to be considered synergistically [45,47]. Johansen et al [21] found that significant therapeutic effects were achieved when the total intervention dose reached 1375.33 minutes, while Ko et al [36] subgroup analysis regarding total training duration indicated no significant improvements, regardless of whether the training period exceeded or fell short of 2 weeks. This finding suggests that we need to consider interactions among dose parameters [36].

Technical characteristics significantly moderate therapeutic efficacy. Wu et al [34] found that end-effector robots demonstrated superior effectiveness in improving upper limb motor impairment compared to conventional rehabilitation therapy (Hedges g=0.22; 95% CI 0.09-0.36, I²=35.4%), potentially attributable to their multijoint coordination training mechanism [52]. Moggio et al [40] further compared device types and revealed that exoskeleton devices showed advantages in enhancing overall hand function and reducing disability, while the end-effector device demonstrated better performance in specific motor control. The selection of training modes is equally crucial for achieving functional gains. Wu et al [34] confirmed that unilateral robot training demonstrated clear efficacy, whereas the advantages of bilateral training and combined unilateral-bilateral training regimens remain unsubstantiated [34,38]. Yang et al [16] emphasized that training modes requiring active patient participation are essential for achieving improvements in ADL.

When comprehensively evaluating the effects of RAT on poststroke upper limb function, changes in muscle strength, as one of the core indicators at the ICF body functions level, hold significant clinical reference value. This umbrella review synthesizing relevant meta-analyses found that robotic therapy demonstrates a positive trend in improving muscle strength, though the strength of evidence remains limited. For instance, the analysis by Zhao et al [41] indicated that the upper limb strength in the RAT group was significantly greater than in the control group (SMD=0.42, 95% CI 0.07-0.78, I²=4%), a finding consistent with the results reported by Johansen et al [21] (SMD=0.43, 95% CI 0.16-0.71, I²=46%). However, evidence in this area shows inconsistency, as some studies found no significant between-group differences (MD=0.51, 95% CI –0.06 to 1.09, I²=0%) [41]. This heterogeneity may stem from differences in assessment tools, insufficient specificity of training protocols for strength enhancement, and variations in patients’ baseline levels.

From a health economic perspective, the cost-effectiveness of rehabilitation robotics still warrants careful evaluation. Currently, the price range for such assistive robotic products varies widely, ranging from US $9000 to US $100,000 [33,53]. Robot-assisted rehabilitation requires not only a high initial investment but also involves relatively high system maintenance and operational costs [54,55]. Notably, the ICF framework helps explain the current cost-effectiveness challenges: the clinical benefits are largely confined to the body functions level, with limited impact on the activity level and virtually no high-quality evidence regarding participation-level outcomes. Furthermore, the implementation of robotic rehabilitation is influenced by health care system disparities and cultural contexts [56]. The high costs may pose significant barriers in resource-limited settings, while cultural differences in technology acceptance can affect patient engagement and treatment adherence [57]. Concurrently, the clinical benefits derived from robotic rehabilitation by both patients and health care institutions remain relatively limited, resulting in a cost-effectiveness ratio that has not yet reached an ideal level [58].

Furthermore, patients’ attitudes toward RAT present a mixed picture. Multiple studies indicate that participants perceive robotic therapy as more precise and less invasive, and hold higher expectations regarding treatment outcomes [59,60]. Such positive perceptions may enhance clinical acceptance of the technology. However, a considerable proportion of patients express concerns about the risk of device malfunctions, with some even mistakenly believing that robotic therapy has higher error rates than conventional methods [60-62]. This crisis of trust could significantly impact treatment adherence and ultimately compromise therapeutic outcomes [63].

Currently, evidence regarding the ability of rehabilitation robots to reduce overall health care costs remains insufficient, and their potential advantages in controlling medical expenditures require further validation [64]. Considering multiple factors, including equipment costs, patient acceptance, health care system differences, and cultural contexts, the pathway for clinical integration of robot-assisted rehabilitation still demands careful planning. Future research should not only focus on more rigorous economic evaluations but also address how to enhance patient trust through technical transparency and improved clinician-patient communication, while developing implementation strategies adapted to diverse health care systems and cultural environments, thereby ultimately improving the cost-benefit profile.

It should be noted that this umbrella review has several limitations. First, although our synthesis approach accounted for heterogeneity and bias, the inability to reanalyze individual participant data from the primary studies limited a deeper exploration of heterogeneity sources. Second, some included meta-analyses contained methodological flaws or had missing outcome data, factors that could affect the accuracy of the conclusions. Additionally, as a secondary synthesis of existing evidence, this review cannot circumvent potential biases inherent in the evidence base, including possible publication bias (where studies with positive results are more likely to be published) and the fact that some primary studies may have received commercial funding or involved conflicts of interest. Such factors could compromise the objectivity of the evidence. Third, the comprehensive cross-checking and deduplication of the included systematic reviews and meta-analyses, which themselves contain extensive references to indexed primary studies, required immense and highly time-consuming effort [65]. This posed a significant methodological challenge in the implementation of this study.

Building on the ICF framework, we recommend several key directions for future research. Future studies should prioritize the investigation of participation level outcomes to fully understand the technology's impact on patients’ lives. Furthermore, research should explicitly target the transfer of gains from body functions to activities and participation levels. Future umbrella reviews could enhance the reliability of their findings by expanding the literature search scope, strengthening the assessment of bias risks in primary studies, and raising the quality thresholds for included studies. In future research, artificial intelligence technologies could be leveraged to further optimize the identification and processing of literature overlap, thereby enhancing both efficiency and accuracy.

These limitations suggest that the conclusions of this review should be interpreted with due caution. However, despite these limitations, our application of the ICF framework provides a robust structure for understanding and advancing the field of robot-assisted upper limb rehabilitation.

This umbrella review demonstrates that RAT serves as an effective intervention for poststroke upper limb rehabilitation, with the most robust and consistent evidence supporting its efficacy in improving motor function, particularly demonstrating clear benefits for patients in the chronic phase. However, its therapeutic advantages are primarily concentrated in short-term functional improvement, while long-term efficacy maintenance remains insufficient, constituting a core bottleneck in current clinical application. The therapeutic effect is significantly modulated by robot type, training mode, and intervention intensity. Future research should focus on developing stratified and individualized treatment protocols based on patient injury characteristics, rehabilitation stage, and device features, and conduct cross–health care system health economic evaluations to advance the clinical translation of precision rehabilitation technologies.