This study strictly followed the Preferred Reporting Items for Systematic Evaluation and Meta-Analysis (PRISMA) guidelines. The study was prospectively registered with the PROSPERO platform (registration number: CRD420251106629).

Literature screening and inclusion were done independently by two researchers. Discrepancies were addressed through discussion, with the option to escalate to an independent arbitrator for final resolution.

The inclusion criteria were shown below:

The exclusion criteria were:

To ensure comprehensive collection of relevant studies, we systematically searched multiple Chinese and English electronic databases. English databases included Embase, PubMed, the Cochrane Library, and Web of Science. Articles published from the inception of each database through July 30, 2025, were included in English language. An updated search was performed on October 19, 2025. The full search strategy is provided in Supplementary Table S1. Literature screening was done independently by two researchers. Duplicate records were initially eliminated through the reference management software, and the remaining records were subsequently screened title by title, abstract by abstract, and full text by full text. All disagreements were settled via team consultation or, failing which, by recourse to an independent adjudicator.

Data extraction was performed in line with the Cochrane Handbook. The process was completed independently by two researchers utilizing a pre-planned standardized data extraction sheet in Microsoft Excel. Disagreements were worked out by jointly reviewing the original article or third-party adjudication. The extracted information mainly includes: basic study characteristics, participant characteristics, intervention details, including specific type of TMBE, session duration, frequency, total intervention period, and post-intervention mean, standard deviations, and sample sizes for cognitive domain tests. Overall cognitive domains mainly include executive function, language ability, attention and memory. In this study, we mainly extracted the mean and standard deviation of each group post-intervention to analyze the intervention effect of TMBE for cognitive function.

The methodological quality including RCTs was evaluated by the Cochrane Risk of Bias Assessment Tool (RoB 2). This tool assesses seven fields (Figure 1). The work was completed independently by two researchers. Each field was assessed as ‘low risk of bias,’ ‘unclear risk of bias,’ or ‘high risk of bias’ according to the RoB 2 criteria. If the two evaluators did not agree, consensus was reached through joint deliberation or arbitration by a third researcher. The overall risk of bias judgment for each study was assigned according to its comprehensive domain scores. Detailed assessment results are shown in Figure 1.

All meta-analysis were conducted by Stata 18.0 and Reviewer manager 5.4. Effect sizes were combined and analyzed using a random effects model to better accommodate heterogeneity among studies regarding exercise type and population characteristics. To investigate potential sources of heterogeneity, Sufficient studies were available (k ≥ 10), meta-regression analyses were performed to explore the influence of continuous variables on effect sizes, using the metareg command in Stata with REML estimation and Knapp-Hartung adjustment. Subgroup analyses were conducted for categorical variables such as intervention type and participant diagnosis.

For continuous outcome measures related to cognitive function, when all studies employed the same assessment tools and units by using MMSE or MoCA scale scores, we pooled the mean differences. When studies assessed the same cognitive construct using different tools such as employing multiple distinct memory tests, we pooled the standardized mean differences. Both are reported with their 95% confidence intervals (CI). Heterogeneity was quantified using the I² statistic and the Cochran’s Q test, with I² > 50% or a Q-test p-value < 0.10 indicating substantial heterogeneity.

Sensitivity analyses were conducted by sequentially the one-by-one exclusion method to assess whether results were unduly influenced by specific studies. If heterogeneity significantly decreased or the pooled effect size changed importantly after removal, it suggested that study might be a major source of heterogeneity. The robustness of the findings was evaluated through iterative exclusion of individual studies and reassessment of heterogeneity in the pooled estimates. To assess potential publication bias, Egger’s test was used and assessed by visual inspection of funnel plots. If data were incompletely reported in the original studies, such as means and standard deviations, corresponding authors were contacted using e-mail to obtain information for the analysis. If it cannot be obtained from the authors, we plan to calculate and convert the data using standard error, confidence intervals, or p-values provided in the original articles, following the guidelines in the Cochrane Handbook.

A systematic literature search initially identified 1,422 potentially relevant records. Before screening, 569 duplicate records and 7 retracted records were removed. Subsequently, titles and abstracts of 846 records were screened, leading to the exclusion of 474 irrelevant records, which included 28 study protocols, 254 reviews/meta-analyses, and 192 non-interventional studies. We attempted to retrieve complete text of the remaining 372 records, of which 1 was unavailable. The full text of 371 articles underwent rigorous assessment for compliance with the study. The main reasons for exclusion include: ineligible study design (e.g., non-RCTs, trial registry entries without resultant publications, total n = 169), and ineligible population (n = 181). Furthermore, a citation search turned up 2 more records, but both were excluded after assessment due to ineligible population. Ultimately, the meta-analysis included 21 studies. The research selection process was shown in the PRISMA flowchart (Figure 2).

Details of the 21 studies (32–52) included in our study are listed in Table 1, which were published across the years 2011 to 2025. All studies had varying sample sizes (22–389), but all covered both male and female populations. The participants included individuals with MCI (K = 12), PD (K = 2), AD (K = 1), and other types of cognitive impairment, including a-MCI (K = 1), MD (K = 2), MCI&SMC (K = 1), SMC (K = 1), and MCI&SCD (K = 1).

Twenty-one RCTs were included for methodological quality assessment. Across the seven bias domains, most studies were judged as low risk concerning random ‘sequence generation’, ‘blinding of participants and personnel’, and ‘completeness of outcome data’. Specifically, regarding random sequence generation, approximately half of the studies were low risk, while the others were rated as having an unclear risk, as a result of insufficient description of the randomization method. Concerning allocation concealment, most studies raised concerns due to inadequate description of the allocation concealment method. In the light of blinding implementation, for blinding of participants and personnel (performance bias), most studies were low risk, with a few being unclear; for blinding of outcome assessment (detection bias), most studies were low risk, while some were unclear. For selective reporting, the vast majority of studies were low risk, with only individual studies rated as high risk. Furthermore, all studies were rated as low risk in the ‘other bias’ domain. The qualitative aspects of the methodology in the studies included was generally acceptable, with the primary potential sources of bias related to inadequate reporting of allocation concealment and incomplete details regarding the implementation of some blinding procedures. More detailed assessment results are shown in Figure 1.

A comprehensive assessment of global cognitive function was conducted through a combination of the MMSE and MoCA. Five studies (total 814 participants, intervention/control: 342/472) assessed the intervention effect on global cognitive function. A random-effects meta-analysis of MMSE scores showed a significant advantage for the intervention group over the control group (MD = 0.65, 95% CI: 0.20 to 1.09, p = 0.004), but there was high heterogeneity between studies (I² = 87.6%, p < 0.001). Assessment of treatment effect heterogeneity by disease type confirmed that significant improvement was preserved within the relevant subgroup (MD = 0.44, 95% CI: 0.30 to 0.58, p = 0.000). Among participants with MCI, MMSE scores showed a trend toward improvement but did not reach statistical significance (MD = 0.32, 95% CI: 0.17 to 0.48, p = 0.195), and heterogeneity remained high (I² = 58.3%, p = 0.091). Notably, a substantial benefit effect was observed in the Mild Dementia (MD) population (MD = 1.76, 95% CI: 1.24 to 2.28, p = 0.000). This high degree of heterogeneity may stem from the diversity of the included populations, ranging from MCI to mild dementia and variations in intervention types, such as Tai Chi and Baduanjin. Subgroup analyses revealed the largest effect size and lowest heterogeneity in the mild dementia cohort, indicating that disease severity is a significant influencing factor.

Ten studies (total 419 participants, intervention/control: 211/207) used the MoCA to assess global cognitive function. Pooled analysis revealed that MoCA scores in the intervention group indicated markedly better versus control group (MD = 0.87, 95% CI: 0.46 to 1.29, p = 0.000), with moderate heterogeneity (I² = 74.3%, p < 0.001). This subgroup analysis also showed remarkable improvement (MD = 0.74, 95% CI: 0.54 to 0.94, p = 0.000). In the MCI population, MoCA scores increased significantly (MD = 0.82, 95% CI: 0.60 to 1.05, p = 0.142), but heterogeneity was high (I² = 77.1%). In PD patients, scores in the intervention group showed an increasing trend, however, the observed differences did not reach statistical significance (MD = 0.36, 95% CI: −0.12 to 0.84, p = 0.000). In summary, TMBE conferred substantial benefits on global cognitive function, the MD subgroup exhibited the most pronounced response. Detailed results are shown in Figures 3–7.

To further explore potential sources of MoCA score heterogeneity, we conducted meta-regression analyses for two continuous variables, the mean age of participants and total intervention duration (Supplementary Figure S1). These variables were selected based on their theoretical association with cognitive outcomes and data availability across studies.

Meta-regression analysis for mean age revealed no significant linear association between age and effect size (β = −0.046, 95% CI: −0.122 to 0.029, p = 0.193). This model explained only limited between-group variance (Adj R² = 12.50%), with substantial residual heterogeneity (I²_res = 71.36%). This indicates that age alone cannot fully account for the variability in MoCA improvement effects.

In contrast, meta-regression analysis for total intervention duration revealed a significant positive correlation with effect size (β = 0.00033, 95% CI: 0.00009 to 0.00058, p = 0.013). This model explained a substantial proportion of variance (Adj R² = 70.08%), with reduced residual heterogeneity (I²_res = 45.55%). This indicates that longer intervention duration correlates with greater MoCA score improvement, and that dose variation partially accounts for inter-study heterogeneity.

Subgroup analysis by TMBE type revealed that all three types significantly improved MoCA scores (Figure 8), with no statistically significant differences between subgroups (p = 0.817). However, substantial heterogeneity persisted within the Baduanjin (I² = 77.0%) and Tai Chi (I² = 82.5%) subgroups, suggesting factors beyond exercise type contribute to variability. The Wuqinxi subgroup included only one study, limiting interpretability.

This meta-analysis confirms that TMBE produces significant benefits across multiple core cognitive domains, including executive function, language ability, memory, and attention in patients with NDDs and those in the prodromal stage of cognitive decline.

These findings provide crucial evidence to inform clinical decision-making for this population and establishes a practical foundation for promoting non-pharmacological intervention strategies. Impaired global cognitive function is a core symptom of NDDs and age-related cognitive decline, profoundly compromising the living quality and ability for independent living in affected individuals. In clinical practice, selecting interventions that effectively and consistently improve cognitive function for this population remains an ongoing challenge. The view is supported by the solid evidence from our study, demonstrating that TMBE making global cognitive function elevated, which aligns with findings from previous systematic reviews (23, 24, 53).

The efficacy of TMBE may stem from its multi-component integrated nature. It combines moderate-intensity physical activity, cognitive focus, breathing regulation, and meditation practice, synergistically supporting cognitive health through both physiological and psychological pathways (54, 55). Its gentle nature also ensures good compliance among older adults and those with cognitive impairment (25).

Mechanistically, this integrated intervention may exert effects through multiple pathways, including enhancing cerebral blood flow, regulating neurotrophic factors, strengthening neural connections, and reducing neuroinflammation and stress responses.

When interpreting the findings of this study, it is crucial to consider their clinical significance. Although TMBE produced statistically significant improvements of 0.87 points on the MoCA and 0.65 points on the MMSE, these magnitude effects may be near or below the threshold for minimal clinically important differences, which are typically considered to be one to two points for the MoCA, and two to three points for the MMSE. This suggests that TMBE exerts a positive directional effect on group cognitive function, but its perceptible benefits for individual patients may be subtle. Achieving unequivocal clinical significance may require longer-term, more intensive interventions.

Based on current evidence, therefore, TMBE can be recommended as a clinical cognitive intervention strategy suitable for delaying neurodegenerative cognitive decline.

This study found that TMBE significantly improved multiple core components of executive function, including working memory, cognitive flexibility, and conflict inhibition. This finding aligns with the consensus that executive function represents the cognitive domain most amenable to training effects (56). Such broad improvements are likely directly driven by the multitasking nature of TMBE, practitioners must simultaneously coordinate complex movement sequences while maintaining cognitive focus and regulating respiration. This process continuously challenges and activates the prefrontal cortex—the critical neural substrate underpinning executive function (57–59).

This study also revealed important negative and heterogeneous findings. TMBE showed no significant effect on immediate recall, potentially because this cognitive dimension relies more on medial temporal lobe structural integrity (60), and is less sensitive to interventions primarily challenging prefrontal cortex and executive function. More notably, the negative effect observed on long-delayed recall in the MCI&SCD subgroup (50), strongly suggests significant biological or psychological heterogeneity within the preclinical cognitive decline population. For some individuals, complex mind–body exercises may introduce cognitive load rather than provide support during early intervention phases. These findings caution that the cognitive benefits of TMBE are not universal, underscoring the urgent need for future research to identify optimal responders versus potential non-responders.

Unlike its effects on executive functions, TMBE demonstrates marked dimensional specificity and population heterogeneity in enhancing memory functions. A key finding is that TMBE exerts no significant impact on immediate recall. This may stem from the greater dependence of immediate recall on structural integrity within the medial temporal lobe, such as the hippocampus (61–63). Interventions targeting the prefrontal cortex show lower sensitivity to major challenges (59). More importantly, the long-delayed recall data present a complex pattern, despite observing significant and consistent benefits in individuals with MCI and SMC, a distinct negative effect emerged in the subgroup with MCI&SCD. This heterogeneous finding strongly suggests that distinct populations at subtle stages of cognitive decline may exhibit fundamentally different physiological and psychological responses to TMBE. For some MCI&SCD individuals, the complex Mind–Body practice may have exceeded their cognitive resources during the early intervention phase, introducing cognitive load rather than providing support. Therefore, treating prodromal cognitive impairment as a homogeneous group is inappropriate; this finding underscores the importance of distinguishing responders from non-responders in future studies and exploring the underlying neural mechanisms. Notably, TMBE demonstrated significant memory improvement in patients with mild dementia, whereas its effects were relatively limited in individuals with mild cognitive impairment. This differential response may stem from the relatively intact memory circuits and higher baseline memory performance of the latter group. This aligns with the view proposed by Liu et al. (64) that individuals with more pronounced cognitive impairment can achieve greater cognitive benefits. Furthermore, extending the intervention cycle may enhance the memory-improving effects of TMBE.

This analysis demonstrates that TMBE consistently enhances language abilities, such as verbal fluency (7). Given that language production heavily relies on executive control for lexical retrieval and memory systems (65, 66). The benefits of TMBE on language function are likely not mediated directly, but rather indirectly through its effects on improving executive function and episodic memory.

The effects of TMBE on attention are the most complex, with its efficacy assessment highly dependent on the tools employed. One of the most significant findings is that TMBE demonstrates a consistent improvement in auditory attention, as indexed by forward digit span (Figure 11), while its benefits for visual processing speed are inconsistent and modest by TMT-A. This discrepancy may reflect the differential effects of TMBE components, exercises involving verbal instructions and rhythmic control appear to enhance auditory processing more strongly (67), while its impact on fundamental visual search efficiency is limited. Thus, TMBE may not constitute a broad-spectrum attention intervention (68), and clinicians should consider its application based on the specific attention deficit dimension of patients, such as auditory or visual.

A primary limitation is the substantial heterogeneity observed across multiple analyses. Although anticipated and addressed through random-effects models and subgroup analyses, its presence cautions against overgeneralizing conclusions. This heterogeneity stems from the study design, which intentionally included a wide-ranging patient population, ranging from SCD to mild dementia, diverse TMBE intervention protocols varying in type, dosage, and guidance quality, and the versions and cultural adaptations of cognitive assessment tools employed. Although our sensitivity analyses successfully identified specific sources such as exergaming-based interventions, residual heterogeneity indicates that unmeasured effect-modifying factors play a role, such as genetic background, specific pathology, and sociocultural context.

Another issue is the lack of detailed raw data, which prevents us from conducting in-depth analyses of specific factors influencing treatment efficacy. For example, the absence of information on patient comorbidities, gender, or genetic background limits our ability to explore personalized treatment approaches. It is important to note that the findings of this study may be more applicable to East Asian populations. Most research has been conducted in this region, and the effectiveness of different mind–body exercise modalities may be influenced by cultural contexts. Furthermore, blinding patients and coaches in such exercise studies is challenging, potentially introducing bias into the findings.

To address these limitations, future research should conduct larger-scale, longer-term clinical trials. We recommend directly comparing the efficacy of different exercise modalities to provide patients with more precise exercise regimens. Additionally, we explored the effects of age, intervention duration, and exercise type on heterogeneity through meta-regression and subgroup analysis, other potential moderating factors could not be fully analyzed due to inconsistencies across studies. Moreover, the limited number of yoga intervention studies and the lack of standardized assessment tools constrained our ability to conduct independent subgroup analyses for this intervention type.