The systematic review employed the PICO (Population, Intervention, Control, Outcomes) framework to establish selection criteria: (1) Population: Adults (≥18 years) with radiologically confirmed unilateral hemispheric stroke and upper limb motor and function impairment; (2) Intervention: Experimental interventions involving robot-assisted upper limb rehabilitation utilizing electromechanical devices, including but not limited to end-effector robotics (EE), exoskeletal robot devices (EXO), or soft robotic gloves (SRG); (3) Control: At least one control group receiving conventional rehabilitation therapy (e.g., physical therapy, occupational therapy, task-oriented therapy, intensive rehabilitation therapy); (4) Outcomes: outcome measures included the Fugl-Meyer Upper Extremity Motor Assessment (FMA-UE), Modified Ashworth Scale (MAS), grip strength (GS), Modified Barthel Index (MBI), and Stroke Impact Scale (SIS), Fugl-Meyer Upper Extremity Motor Assessment -proximal/ shoulder and elbow coordination (FMA-SE), Fugl-Meyer Upper Extremity Motor Assessment -distal/ wrist and hand function (FMA-WH), Fugl-Meyer Upper Extremity Motor Assessment -hand dexterity (FMA-H), Motor Activity Log - amount of use (MAL-AOU), and Motor Activity Log - quality of movement (MAL-QOM); (5) Randomized controlled trials (RCTs).

Studies were excluded if they met any of the following criteria: (1) Interventions involving group-, family-, or self-guided therapy; (2) Bilateral upper limb treatment in either experimental or control groups; (3) Insufficient outcome data for effect size calculation; (4) Robotic-assisted therapy combined with adjunct interventions (e.g., repetitive transcranial magnetic stimulation (rTMS), virtual reality (VR), or brain-computer interfaces (BCI)); (5) Comparative studies using distinct robotic device types across intervention arms; (6) Mixed stroke stage cohorts (acute/subacute/chronic) without stratified analysis; (7) Crossover study designs.

Stroke phases were classified as: (1) Acute: ≤1 month post-onset; (2) Subacute: 1–6 months; (3) Chronic: ≥6 months (10, 11).

Note: This staging system may differ from source studies due to temporal reclassification.

Sample size and attrition rates were calculated exclusively from participants completing final outcome assessments.

Two investigators (WLL and ZEY) conducted dual independent screening of titles/abstracts against predefined eligibility criteria, followed by full-text assessments. Discrepancies were resolved through consensus discussions with a senior researcher (LXL).

A standardized extraction template captured: (1) Study descriptors: Title, first author, publication year, location; (2) Participant characteristics: Age, sex, post-stroke duration (days), baseline motor scores; (3) Intervention parameters: Robotic device type, session length (minutes), total duration length, follow-up period; (4) Outcome metrics: Pre/post-intervention scores with variance measures for all specified scales.

Missing outcome data were addressed by contacting corresponding authors via email. Studies with incomplete datasets unresponsive to data requests were systematically excluded from quantitative synthesis.

Two independent reviewers (LY and LCT) evaluated methodological quality using the Cochrane Risk of Bias Tool for Randomized Trials (RoB 2.0), assessing five domains: randomization process, deviations from intended interventions, missing outcome data, outcome measurement, and selective reporting. Inter-rater discrepancies were resolved through consensus adjudication by a third reviewer (LXL).

To assess the robustness of our findings, a pre-specified sensitivity analysis was performed by excluding studies judged to be at high overall risk of bias. The impact of this exclusion on both baseline characteristics and pooled effect estimates was rigorously evaluated.

Meta-regression analyses were performed to quantify the confounding effects of baseline imbalances on the estimated treatment effects. The results revealed significant associations for several outcome measures:

GS: Baseline imbalance significantly influenced outcomes (β = 0.666, 95% CI (0.11, 1.23), p = 0.020), primarily attributed to studies A20 (51), B4 (44), B7 (52), and B15 (40). After statistical adjustment for this imbalance, the treatment effect was substantially attenuated and became non-significant (Pooled Hedge’s g = 0.224, 95% CI (−0.10, 0.55), p = 0.174), indicating that the unadjusted effect was likely inflated.

GS Subgroups: different phases (β = 0.628, p = 0.022), Chronic phase (β = 0.799, p = 0.001), age ≤60 years (β = 0.848, p = 0.001), and duration <24 sessions (β = 0.700, p = 0.007) showed strong baseline effects (sources: A20 (51), B15 (40)). Adjusted effects remained non-significant (p > 0.05).

MBI (session ≤30 min): Baseline imbalance impacted outcomes (β = 1.36, p = 0.046), though adjusted effects were non-significant (WMD = −1.24, p = 0.493).

MAL Scales (AOU/QOM): Baseline effects were significant (AOU: β = 0.62, p = 0.028; QOM: β = 0.58, p = 0.032), largely driven by study A20 (51). Adjusted effects were non-significant (p > 0.05).

These results indicate that the apparent benefits of robotic therapy on grip strength and functional activity are likely inflated by pre-existing group differences, and definitive conclusions on its efficacy for these specific outcomes cannot be drawn from the present data.

FMA-SE: In contrast to other measures, the significant treatment effect for FMA-SE demonstrated robustness against baseline confounding (β = 0.85, p < 0.001), the adjusted treatment effect remained both statistically and clinically significant (Pooled Hedge’s g = 0.92, 95% CI (0.23, 1.61), p = 0.009).

This confirms that the improvement in proximal upper limb motor function is a reliable and robust finding, independent of baseline confounding.

Complete statistical outputs for all analyses, including subgroup-specific confidence intervals and source study details, are provided in Supplementary Table S3. Sensitivity analyses further corroborating these findings are presented in Supplementary Tables S3-3–S3-13.

Significant publication bias was detected through Egger’s test in critical analyses, with statistically significant effects observed for the FMA-UE chronic phase subgroup (p = 0.044) and age≤60 years subgroup (p = 0.002), as well as across all grip strength (GS) analyses including the overall assessment (p = 0.012), different phases subgroup (p = 0.005), and age≤60 years subgroup (p = 0.014). Trim and fill adjustment confirmed the robustness of these findings, as evidenced by the FMA-UE subgroups demonstrating enhanced significance after adjustment (chronic phase: p = 0.044 → 0.000; age≤60: p = 0.002 → 0.000), while GS subgroups maintained statistical significance post-adjustment (overall: p = 0.017; different phases: p = 0.030; age≤60: p = 0.013). These analyses collectively indicate that although publication bias was statistically evident, trim and fill adjustments consistently demonstrated robust effect estimates across all outcome measures. Complete publication bias statistics, including Egger’s intercept values and trim and fill-adjusted outcomes, are detailed in Supplementary Table S4, with corresponding funnel plots incorporating imputed studies provided in Supplementary Figure S3.

The reclassification protocol was applied to 1,583 participants with available data. The distribution of time from stroke onset yielded a median of 60.5 days (IQR: 28.3–182.7). This empirical distribution was used to calibrate and validate the following predefined phase thresholds: Acute phase: ≤1 month (≤30 days) post-onset, subacute: 1–6 months (31–182 days), and chronic: ≥6 months (≥183 days).

The median (60.5 days) falling within the subacute phase and the IQR upper bound (182.7 days) aligning closely with the chronic threshold supported the validity of this classification for our cohort. Validation confirmed that extreme values (>5 years) exerted negligible influence on the median (weighted simulation error <2%). This reclassification resolves prior definitional heterogeneity, ensuring the validity of subsequent time-stratified analyses. Detailed original and reclassified phases for all studies are provided in Supplementary Table S5.

To assess the robustness of our findings against methodological quality concerns, we performed a sensitivity analysis by excluding the five studies judged to be at high overall risk of bias (A10 (42), A12 (43), A14 (15), B4 (44), B13 (45)). The results for the primary outcome, FMA-UE, remained stable and statistically significant (WMD = 5.91, 95% CI (3.45, 8.37), p < 0.001), confirming the core finding of this meta-analysis. Similarly, no substantial variations were observed in other primary outcomes, including the clinically meaningful improvement in activities of daily living (MBI).

This exclusion, however, refined our understanding of two secondary outcomes:

Crucially, despite these shifts in baseline comparability, the endpoint treatment effects for both GS and FMA-H remained stable and statistically significant (GS: SMD = 0.63, 95% CI (0.30, 0.95), p < 0.001; FMA-H: SMD = 2.03, 95% CI (0.89, 3.18), p = 0.001). This indicates that the efficacy signals for these outcomes are not solely driven by the identified high-risk studies.

Detailed results of this sensitivity analysis are provided in Supplementary Table S6.

Subgroup analyses were performed, and comprehensive statistical analyses were conducted for all primary and secondary outcome measures, with the detailed results presented in Table 2.

The interpretation of our extensive subgroup analyses should be framed within their methodological limitations. These analyses were exploratory and not pre-powered for interaction effects. The multiplicity of statistical tests increases the risk of Type I error, and thus these subgroup findings should be considered hypothesis-generating rather than confirmatory. We advise caution in drawing clinical conclusions from them alone.

An important consideration is the distinction between statistical significance and clinical relevance. Our analysis suggests that robotic-assisted therapy may yield not only statistically superior gains but also clinically meaningful improvements in certain domains. The most compelling evidence emerged for motor recovery in the subacute phase, where the improvement (FMA-UE WMD = 8.82) surpassed the lower bound of the established MCID range (6.6–10 points) (56–61), indicating a change likely perceptible to patients. Similarly, the effect on activities of daily living across the cohort (MBI WMD = 8.00) substantially exceeded its MCID threshold (1.85 points) (21), suggesting a high probability for a meaningful enhancement in functional independence. In contrast, while motor improvements in the chronic phase were statistically significant (FMA-UE WMD = 3.74), the effect size fell below the MCID range for this population (4.25–7.25 points) (10), implying that the average benefit may not constitute a substantial functional shift. Likewise, the improvement in social participation (SIS WMD = 4.19), though statistically significant, did not reach the MCID for the participation domain (5.9 points) (62). This gradation of findings underscores that the intervention’s capacity to elicit patient-important changes is not uniform but is most strongly supported for improving functional independence in subacute and more severely impaired patients.

Our findings address a critical gap in previous meta-analyses that focused predominantly on single-phase stroke rehabilitation, by systematically evaluating robotic interventions across acute, subacute, and chronic post-stroke phases. Our results delineate a distinct efficacy profile: the most substantial benefits were observed during the subacute phase, a period of heightened neuroplastic potential, while a more modest yet statistically significant effect was maintained in the chronic phase, consistent with prior findings (63). Acute-phase analyses revealed no EG-CG differences (p > 0.05), likely attributable to confounding factors including clinical instability, treatment intolerance, heterogeneous spontaneous recovery rates, and limited sample size (7, 24).

Mechanistically, robotic rehabilitation enhances upper limb functional recovery through neuroplastic modulation. Evidence from neuroimaging and electrophysiological studies indicates that robotic training augments proprioceptive-driven cortical excitability in sensorimotor regions (64, 65), strengthens contralateral frontoparietal connectivity (26, 64), and reduces transcallosal inhibition to restore interhemispheric balance (66). These effects align with animal models demonstrating critical post-stroke recovery windows characterized by heightened neural circuit plasticity, particularly within 12 weeks post-onset (67, 68). While maximal functional gains occur in acute (≤7 days) and subacute (7 days–3 months) phases (69–72), our data corroborate Mazzoleni et al. (67) in showing that chronic patients (≥1 year post-stroke) still achieve incremental benefits from intensive robotic therapy, underscoring its utility across disease stages.

Our stratified analysis revealed differential therapeutic responses across injury severities. Patients with severe upper limb impairment demonstrated clinically meaningful gains from robotic rehabilitation, corroborating findings by Wu and Klamroth-Marganska et al. (73–75). This may be attributed to enhanced neuroplasticity mechanisms and greater potential for spontaneous recovery in this population, coupled with higher adherence to repetitive robotic training protocols. Conversely, no significant intergroup differences were observed in mildly impaired patients (EG vs. CG, p > 0.05), likely influenced by methodological limitations: (1) restricted sample size in mild-injury subgroups; (2) ceiling effects of conventional assessment scales; and (3) prioritized rehabilitation goals focusing on fine motor control and participation metrics, which are less targeted by current robotic systems (48).

Robotic rehabilitation systems are classified as exoskeletons (EXO), end-effectors (EE), or soft robotic gloves (SRG), with SRGs categorized separately due to their structural incompatibility with traditional EXO/EE designs (76, 77). Both EXO and EE devices demonstrated statistically and clinically significant improvements in upper limb function compared to traditional therapies (p < 0.01), with no significant difference observed between these two modalities (p = 0.939), consistent with prior comparative studies (6, 78). In contrast, our analysis found SRGs did not exhibit superiority over traditional methods (p = 0.110), diverging from earlier meta-analyses (10, 40). However, this SRG result warrants careful interpretation given the limited included studies (n = 4: B3 (37), B4 (44), B7 (52), B15 (40)), small sample size (n = 128 total), strict inclusion criteria (excluding hybrid devices), and participant characteristics (predominantly subacute/chronic stage, mild/moderate impairment) coupled with relatively short intervention durations (<24 sessions). Furthermore, as the Fugl-Meyer Assessment for Upper Extremity (FMA-UE) evaluates overall upper limb function, the specific hand-focused training provided by the SRGs in the included studies, potentially delivered at lower intensity or duration compared to EXO/EE protocols, may limit the detection of their effects on this scale. Consequently, the efficacy of SRGs requires further investigation with larger, more targeted studies.

Our findings do not provide compelling evidence to support the use of chronological age as a key stratification factor for robotic therapy. The absence of significant subgroup differences indicates that treatment efficacy is largely comparable across age groups. Any observed marginal benefit in younger patients is likely attributable to confounding factors correlated with age—such as overall health status, frailty, or therapy adherence—rather than to age itself (79, 80). This suggests that future research should focus on these specific aging-related factors rather than on chronological age alone to better understand differential treatment responses.

Our analysis demonstrates that robotic rehabilitation therapy elicited significant functional improvements in both proximal (shoulder/elbow) and distal (wrist/hand) joints in stroke patients. The magnitude of improvement did not differ significantly between joint groups (p > 0.05), corroborating findings by Wu et al. (73). This functional equivalence may stem from distinct neural substrates: proximal control primarily engages the corticoreticulospinal tract, whereas distal control relies predominantly on the corticospinal tract (81). Although a marginal advantage for proximal training was observed, this difference was not statistically significant. We hypothesize that the observed functional balance may result from distal training enhancing inter-joint coordination and inducing compensatory proximal stabilization during movement execution (82, 83).

Our subgroup analysis of robot-assisted rehabilitation therapy did not show a significant dose-dependent therapeutic effect. The grouping of training duration (≤ 30 min vs. > 30 min per session) and the classification of cumulative frequency (< 24 sessions vs. ≥ 24 sessions) (10, 84) consistently indicated that the upper limb function improved significantly after the intervention (p < 0.001). Consequently, we provisionally recommend prioritizing high-dose protocols (>30 min/session, ≥24 sessions) for moderate-to-severe impairments to optimize functional gains, while considering the minimal effective dose (e.g., 24 sessions × 30 min) for mild cases to enhance resource utilization efficiency. These clinical interpretations require validation through rigorously designed trials.

Longitudinal follow-up data showed sustained functional gains in patients assessed >3 months post-intervention, with progressive improvements observed between treatment completion and final follow-up (85). This delayed enhancement may reflect residual neuroplastic effects from robotic therapy, potentially mediated by long-term cortical reorganization (86). In contrast, patients evaluated ≤3 months post-treatment maintained but did not significantly exceed their immediate post-training functional levels (p > 0.05). The attenuated recovery trajectory in shorter follow-up cohorts likely results from reduced post-protocol training intensity and physiological plateaus in neural repair processes.

Our analysis demonstrated comparable spasticity modulation between robotic and conventional therapies, with both groups showing modest reductions in MAS scores lacking clinical significance (ΔMAS <1.0 point; p > 0.05) (87–89). This aligns with neuroanatomical evidence implicating reticulospinal tract hyperexcitability as the primary driver of post-stroke spasticity (90, 91). Robotic interventions, while effective in enhancing corticospinal pathway plasticity, may inadequately modulate brainstem-mediated reticulospinal circuits critical for stretch reflex regulation (92). Notably, longitudinal assessments confirmed peak spasticity severity at 1–3 months post-stroke (93), suggesting this phase represents a critical window for targeted antispastic interventions.

Regarding grip strength outcomes, baseline heterogeneity limited definitive conclusions, though chronic-phase patients exhibited statistically meaningful improvements in the experimental group (EG). This aligns with evidence that chronic-stage functional recovery correlates with affected hemisphere sensorimotor cortex reactivation during grasping tasks (36, 49). Subgroup analyses indicated distal-focused robotic training (>30 min/session) produced superior grip strength gains (p = 0.003 vs. proximal training), likely through enhanced proprioceptive feedback to cortical hand representation areas. These findings advocate for task-specific, intensity-optimized robotic protocols in chronic stroke rehabilitation.

A key distinction emerges between the effects of robotic therapy on basic activities of daily living versus broader social participation. The intervention demonstrated a robust and clinically meaningful impact on functional independence (MBI), a finding that underscores its success in restoring foundational self-care capabilities. This benefit is likely mediated by robotic training’s capacity to deliver high-intensity, task-oriented practice, which promotes adaptive neural plasticity in sensorimotor networks and enhances interhemispheric connectivity (26). These neuroplastic mechanisms appear highly effective for recovering the discrete motor functions required for basic activities.

In contrast, the transfer of these gains to enhanced social participation (SIS) was limited and fell below the threshold for clinical importance. This discrepancy suggests that the neuroplastic changes sufficient for improving discrete motor tasks do not automatically translate into the broader cognitive, psychosocial, and environmental adaptations required for successful community reintegration. While a marginal, age-related benefit in participation was observed, its interpretation remains speculative and warrants further investigation to disentangle neurobiological from psychosocial influences. Future interventions may need to explicitly incorporate participation-focused strategies to bridge this gap between motor recovery and real-world involvement.

Although the methodological heterogeneity of kinematic data precluded formal meta-analysis, a narrative synthesis of the included studies offers valuable insights into motor control quality that extend beyond conventional functional scales. The available evidence suggests that robotic devices provide sensitive kinematic measures—such as movement smoothness and reduced muscle co-contraction—that serve as objective biomarkers of neuromuscular improvement (8, 19, 94, 95). These parameters appear to capture the dissolution of pathological synergies and the restoration of inter-joint coordination, changes that may reflect fundamental neurobiological recovery (67, 96). Notably, the observed dissociation between these qualitative motor improvements and functional scores in some studies is particularly instructive; it indicates that kinematics may uniquely quantify the reduction of compensatory strategies, thereby providing a purer measure of true neurological restoration. This perspective is further supported by correlative neurophysiological evidence of normalized interhemispheric connectivity (4, 84, 97). Consequently, while quantitative synthesis was not feasible, the consistent reporting of these kinematic gains across studies underscores their complementary value in elucidating recovery mechanisms and highlights the importance of standardizing these measures in future trials.

The sensitivity analysis excluding five high-risk studies (A10 (42), A12 (43), A14 (15), B4 (44), B13 (45)) confirmed the robustness of our primary findings. Treatment effects for the FMA-UE and activities of daily living (MBI) remained statistically significant and stable. This process, however, revealed a critical distinction in the reliability of secondary outcomes. While the benefits for proximal limb control (FMA-SE) withstood adjustments for baseline confounding, the apparent effects on grip strength (GS) and self-reported arm use (MAL) were substantially attenuated and lost statistical significance after meta-regression. Therefore, we can state with confidence that robotic therapy improves core motor and functional outcomes, but its specific efficacy on grip strength remains uncertain and warrants further investigation in more rigorously designed trials.

The observed baseline heterogeneity across several outcomes necessitates a careful consideration of the internal validity of our meta-analytic findings. Our meta-regression quantified the substantial confounding effect of these imbalances, which has critical implications for the clinical interpretability of the results. For instance, baseline discrepancies significantly inflated the apparent treatment effect for grip strength. After statistical adjustment, the effect was substantially attenuated and became non-significant (p = 0.174). This pattern, consistent across GS subgroups, suggests that unadjusted GS effects are likely overestimates. A similar pattern was observed for MAL-AOU/QOM, where adjustment also nullified significance.

While randomization in the primary RCTs aims to minimize confounding, these analyses indicate that residual imbalances can significantly impact the results for specific outcomes like GS and MAL. For these measures, the stability of the treatment effect is questionable, as statistical significance was contingent upon unadjusted models. This limits the strength of causal inference regarding the robot-assisted therapy’s specific effect on grip strength and self-reported arm use. Clinically, the unadjusted benefits for these outcomes may not be robust and could be attributable, in part, to pre-existing group differences.

In contrast, the treatment effect for FMA-SE demonstrated notable robustness, remaining statistically and clinically significant after adjustment (Hedge’s g = 0.92, 95% CI (0.23, 1.61), p = 0.009), which enhances our confidence in its stability and internal validity. Therefore, we recommend that conclusions regarding intervention efficacy be drawn primarily from outcomes, like FMA-SE, that have demonstrated stability against potential baseline confounding. Complete analytical outputs are provided in Supplementary Table S3.