The systematic review was registered a priori to PROSPERO (CRD42023488270) and reported following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses; checklist provided in Checklist 1).

We systematically searched publications in the following databases in June 2025: PubMed, Cochrane Library, Web of Science, Embase, Proquest, Physiotherapy Evidence Database, and clinical trial registries: ClinicalTrials.gov and ChiCTR. Two researchers (PS and CN) developed and performed the search strategy together. The following keywords were used for the search: “Stroke,” “Motor Imagery,” “Action Observation,” “MI,” “AO,” etc. The search method uses the combination of Medical Subject Headings (MeSH) and free terms, and the search strategy is shown in Table 1.

The studies were selected based on the PICOS verification method (P-participants; I-intervention; C-comparison; O-outcome; and S-study design), as shown in Table 2.

Two reviewers (XL and XZ) carried out the study selection independently in accordance with the eligibility criteria, and in case of disagreement, a third reviewer (PS) was asked to resolve this. Initially, all potential articles were screened according to the title, abstract, and keywords. Then the full text of all selected articles was checked, and the final selection was made based on the predefined selection criteria.

Three reviewers (PS, XL, and XZ) independently extracted data from the included studies using a predefined form. Data abstraction included study characteristics, participant characteristics, sample size, intervention measures and intervention content, duration of experiments, assessment instrument, and outcomes. Pre-post differences data were extracted directly. For studies that did not report the mean and SD of the pre-post differences, we used the Follmann et al [32] method to convert the means and SDs of the baseline and final values into the mean and SD of the pre-post differences, as described in version 3.0.2 of the Cochrane Collaboration Handbook (page 213). In case of missing or unclear data, we tried to obtain additional information from the study authors.

Methodological quality assessment will be conducted independently by 2 researchers, following the Cochrane Risk of Bias tool (RoB 2.0; The Cochrane Collaboration) for assessing the risk of bias in randomized trials [33] recommended in the Cochrane Handbook [34]. If 2 reviewers (XL and XZ) cannot determine the evaluation results in the quality evaluation, a third independent reviewer (PS) makes the decision.

Two independent reviewers (XZ and XL) used the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach to assess the certainty for each meta-analysis outcome. There are 4 classification levels to assess the quality of evidence: high, moderate, low, and very low. We used the GRADEpro GDT (GRADEpro GDT; Evidence Prime, Inc) to create the “Summary of findings” tables.

Statistical analysis was performed using Review Manager software (RevMan 5.3; The Cochrane Collaboration). Measurement data with the same outcome index unit were calculated using mean difference and 95% CI. SMD and 95% CI were used for different outcome index units. The difference was statistically significant at P<.05. For studies reporting median, we converted these values to mean and SD using the method proposed by Hozo et al [35].

Heterogeneity was tested using the chi-square test and I² statistic [36]. The heterogeneity was considered low when P>.1 and I²<50%, and the fixed-effects model estimated the combined effect size. The heterogeneity was considered significant when P<.1 and I²>50%, and the random effects model was adopted. When it was impossible to combine effect sizes, we performed only a descriptive analysis of the results. Subgroup analyses were conducted to explore potential sources of heterogeneity through subgroup analysis.

For interventions that could not be directly compared, a mesh meta-analysis was performed using Stata 14.1 software (StataCorp LLC). SMD and 95% CI were used to calculate the effect size of this study. The surface under the cumulative ranking curve (SUCRA) was used to predict and rank the efficacy of rehabilitation measures [37]. When there was a closed loop in the network map, the inconsistency model was used for testing. If P>.05, the consistency model was used for analysis. If P≤.05, the source of heterogeneity should be identified and eliminated.

We conducted the sensitivity analysis using the one-by-one elimination method and changing the combined model. If more than 10 studies were included, the funnel plot was used to analyze publication bias.

A total of 7563 results were retrieved from 8 databases, and 5850 results remained after removing duplicates. After the preliminary screening, 219 articles remained. After reading the full text and rescreening, 13 articles [38-50] were finally included, and the flowchart is shown in Figure 1.

The 13 included studies [38-50] involved a total of 399 participants, and most of them were small sample trials. The participants were mainly patients with stroke with upper limb paralysis. A total of 12 studies compared AO+MI with their independent use or conventional rehabilitation [38-49]. In addition, one [50] compared synchronous and asynchronous combination models. Most of these studies had an intervention duration of about 4 weeks. Most of the primary outcome indicators were upper extremity function, while only 2 studies [39,45] assessed the lower extremity function and 1 study [49] assessed the overall motor function. The features of literature inclusion are shown in Table 3. In this table, age and stroke course were expressed as mean (SD) or median (IQR) or median (maximum, minimum).

The RoB 2.0 tool was used to evaluate literature quality. All 13 randomized controlled studies were included, of which 6 [39-42,46,47] had a low risk of selection bias, 2 [42,46] used assignment concealment, 3 [39,42,43] blinded participants and implementors, 8 [38,39,42-44,46,47,50] blinded outcome evaluators, only one [48] outcome measure was incomplete, all studies reported full results, and 7 [42,43,45-47,49,50] had a low risk of other bias. The final evaluation results are shown in Figure 2.

We assessed the certainty of evidence for the 7 outcomes in the meta-analysis. The overall certainty of evidence was rated as low to very low. The GRADE results are shown in Table 4.

This study review aimed to synthesize the evidence on the effects of AO+MI on limb functional rehabilitation in patients after stroke through a systematic review and meta-analysis. We included 13 trials with 399 participants in this review.

The meta-analysis showed that compared with routine rehabilitation therapy, AO+MI could improve limb motor function in patients with stroke. And compared with their independent use, AO+MI could improve upper extremity motor function in patients with stroke. These findings are consistent with the results of a meta-analysis by Chye et al [21] that included all populations (encompassing patients with Parkinson disease, older individuals, children, and healthy adults et al). Relevant and experience-dependent practice encourages the brain to create and reorganize functionally appropriate neural connections, which is pivotal for neurorehabilitation following stroke [52]. A related functional magnetic resonance imaging study [53] has revealed that when individuals engage in AO+MI, blood oxygen level-dependent (BOLD) signals increase and become more widespread in brain regions involved in motor execution. Electroencephalography investigation demonstrates that during AO+MI, individuals exhibit significantly decreased power spectral densities in the alpha and beta bands, indicating heightened activity in the primary sensorimotor cortices [54]. For patients with stroke who have difficulty performing actual movements, the increased neural activity during AO+MI may support repeated Hebbian modulation of intracortical and subcortical excitatory mechanisms through synaptic plasticity, producing effects similar to those of physical practice [21]. This may be the reason why AO+MI improves motor function in patients with stroke. However, our meta-analysis showed that AO+MI did not improve ADL in patients after stroke compared with routine rehabilitation. A potential explanation is that poststroke ADL improvement requires a longer intervention duration, while the 2 studies [42,46] included in this meta-analysis featured relatively short intervention periods (2‐4 weeks).

Although the only direct evidence suggests that synchronous may enhance motor function in patients with stroke compared to asynchronous [50], the network meta-analysis results in this review indicate no significant difference between synchronous and asynchronous AO+MI in improving upper extremity function in patients with stroke. The potential reason for the inconsistency between direct evidence and meta-analysis results may stem from differences among participants. As Eaves and colleagues point out [5], synchronous AO+MI may impose greater cognitive load compared to asynchronous AO+MI. This could result in the synchronous mode being overly demanding for certain patient populations or rehabilitation stages. In such cases, asynchronous intervention might be more advantageous. Consequently, variations in key characteristics among participants across included studies—such as severity of neurological deficits and baseline cognitive levels—may modulate the boundaries of effectiveness between synchronous and asynchronous intervention modalities. Therefore, future studies with more direct evidence are needed to compare the effects of these 2 modalities on poststroke recovery. In addition, asynchronous AO+MI does not rule out the possibility of participants spontaneously performing MI during the AO component of AO then MI. This may result in some studies using asynchronous mode inadvertently including elements of synchronous mode, which might also explain why no significant difference in effectiveness has been observed between the 2 combined approaches.

This review evaluated the included studies strictly according to the relevant measures in the Cochrane 5.1.0 manual for randomized controlled trial quality assessment. Among the 13 studies [38-50] included, only one study [42] had a low risk of bias in all 7 aspects of bias analysis, while the risk of bias in the other studies mainly focused on selection bias and implementation bias. Although most studies claim randomization, only half clearly state the correct way to generate random sequences. In addition, most studies did not report how the randomization scheme was assigned to hide, which led to high-risk outcomes in both evaluations. Most of the studies included in this review were not blinded, which may be due to the changes in rehabilitation measures and the limitations of human resources, material resources, and venues, making it challenging to perform double blindness. The overall quality of the evidence presented in this review, as assessed by the GRADE approach, was low or very low for all comparisons. The main limitations were imprecision due to very small sample sizes, risk of bias, and inconsistency.

In addition to the low quality of the included articles, this review has several limitations. First, the studies we included may exhibit a certain degree of heterogeneity in terms of interventions. In some studies, the effects of AO +MI might be influenced by other concurrent therapeutic approaches, such as mirror therapy or graded motor imagery. For instance, in the study by Ietswaart et al [42], in addition to the 30-minute AO+MI intervention, each session also included an additional 10 minutes of active motor imagery and 5 minutes of a certain implicit form of motor imagery activity. This multicomponent intervention approach may mean that the observed therapeutic effects cannot be entirely attributed to pure AO+MI, potentially affecting the results of this study in the overall analysis. Second, there are a few included articles, and most are small sample studies, thus limiting the reliability of the research conclusions. Third, we only searched and reviewed articles in English; publications in other languages may have been missed. Fourth, the absence of standardized intervention protocols and variation in outcome assessment tools across the included studies complicates direct comparisons and meaningful synthesis of the findings. In addition, significant heterogeneity was observed in the meta-analysis results. Consequently, the results should be interpreted cautiously.

Dorsch et al 2024 [55] meta-analysis explored the effect of adding nonstimulation-based priming prior to task-specific practice on activity and motor impairment outcomes compared with task-specific practice alone in stroke rehabilitation, searching for manuscripts up to March 2019. This review included 2 trials investigating the impact of AO+MI on activity outcomes in patients after stroke, and the results did not show a significant improvement effect of AO+MI. The reason for the inconsistency with our findings might be that Dorsch et al [55] compared the postintervention scores between the conventional group and the AO+MI group, whereas our study compared the pre-post differences of the 2 groups.

Lin et al [56] systematic review included 9 studies investigating the effects of AO+MI on upper limb function in patients with stroke, and the results demonstrated that AO+MI significantly improves upper limb function in participants. Our study differs from Lin et al [56] review in several key aspects. First, our study adopted broader inclusion criteria, examining not only upper extremity function but also lower extremity function, overall motor function, and activities of daily living, thereby addressing a wider range of clinical outcomes. As a result, the set of included studies differs somewhat from that in Lin et al [56] review. Second, this study conducted separate meta-analyses according to the type of control group (conventional intervention, MI alone, or AO alone), thereby minimizing heterogeneity introduced by differences in control conditions, enhancing the precision of the findings, and offering more specific guidance for clinical practice. In addition, this study incorporated a network meta-analysis to compare the effectiveness of 2 AO+MI integration modes—synchronous versus asynchronous—on upper extremity function after stroke. Therefore, although there are overlapping findings, this study and Lin et al [56] review can complement each other and reinforce key conclusions, serving the valuable purpose of replication and strengthening the evidence base.

This systematic review and meta-analysis examined the effects of AO+MI on motor function in patients with stroke. Our results found that AO+MI can improve the motor function compared to routine rehabilitation, AO, or MI. Furthermore, no conclusive evidence supports AO+MI’s efficacy for improving ADL, nor demonstrates differential effects between synchronous and asynchronous application modes on poststroke limb function. The quality of GRADE‐based evidence in this review varied from low to very low, due to the high risk of bias and small sample sizes. Consequently, large-scale randomized controlled trials with rigorous methodology are imperative to establish definitive clinical recommendations.