Two reviewers (SYZ and SF) independently conducted a systematic literature search for randomized controlled trials (RCTs) investigating Baduanjin for T2DM. The search encompassed eight electronic databases: China National Knowledge Infrastructure (CNKI), Wanfang Database (WANFANG), VIP Database (VIP), China Biomedical Literature Service System (CBM), Embase, Web of Science, PubMed, and the Cochrane Library. We also examined clinical trial registries and grey literature sources. The search period extended from each database inception to April 30, 2025. Subject terms and keywords included “Baduanjin,” “eight section brocades,” “diabetes,” “prediabetes,” “diabetes patients,” “diabetes sufferers,” “type 2 diabetes,” “randomized controlled trials,” and “clinical trials.” We further manually screened the reference lists of all included studies. After duplicate removal, studies that clearly did not meet the inclusion criteria were excluded based on title and abstract screening. The full texts of the remaining articles were assessed for eligibility by two reviewers (SYZ and LML). Any disagreements were resolved through consultation with the third reviewer (XYW). The complete search strategy is available in Supplementary Table 2.

The inclusion criteria were (1): RCTs design (2); participants diagnosed with T2DM, irrespective of gender, age, or ethnicity (3); Baduanjin as an intervention in at least one group, with no restrictions on frequency, session duration, or program length (4); The control condition could be non-exercise/usual care (e.g., standard care, medication, health education, routine management, or daily activities) or an active exercise comparator (e.g., walking, aerobic exercise, resistance training, or other structured exercise programs). For analysis, control groups were a priori categorized as non-exercise versus active exercise comparators to enable sensitivity analyses assessing the impact of control group types on the pooled estimates (5); reporting of at least one outcome measure from FBG, HbA1c, TG, or TC.

Pharmacological therapy was permitted in the control condition because it represents routine standard care for T2DM and exercise interventions are typically evaluated as an adjunct to usual care. In this review, concomitant medication was considered a background co-intervention, and pooled effects were interpreted as the incremental benefit of adding Baduanjin beyond standard treatment. Medication-related information (e.g., drug use and co-interventions) was extracted when reported. However, medication regimens (drug classes, dosages, and treatment adjustments) were inconsistently reported across trials; therefore, subgroup analyses or meta-regression stratified by medication class/dose could not be performed.

The exclusion criteria were (1): studies not involving human subjects; (2) sample sizes smaller than 10 participants; (3) duplicate publications; and (4) studies from which relevant outcome data could not be obtained after (i) extracting data reported as Mean ± SD, (ii) transforming reported statistics (e.g., Median/IQR, Median/Range, five-number summary, or Mean with SE/CI) into Mean ± SD using established methods (23, 24), and (iii) attempting to contact the corresponding authors for missing numerical information.

Two reviewers (SYZ and SF) screened the literature using EndNote 21. When outcome data were missing, unclear, or not directly extractable, we contacted the corresponding authors of the original studies for clarification or to request additional information. If the required data could not be obtained, we attempted to extract data from figures/tables when feasible; otherwise, the outcome was excluded from quantitative synthesis. Two reviewers (SYZ and XYW) independently extracted data from the included studies, with any discrepancies resolved by the third reviewer. The extracted information encompassed the authors, publication year, participant characteristics (including age, gender distribution, and sample size), intervention group details (such as Baduanjin frequency, session duration, total weekly practice time, and intervention cycle), and outcome measures (including specific indicators and data). After extraction, reviewers cross-verified all entries in the data form for accuracy, and any disagreements were settled through discussion until a consensus was reached. For continuous outcomes, we preferentially extracted change-from-baseline data (mean change and its SD) for each group. If change scores were not reported, the mean change was obtained from baseline and post-intervention means. When the SD of change was unavailable, it was derived from baseline and post-intervention SDs using the within-person correlation coefficient (R) between baseline and post-intervention measurements. If R was not reported, we applied a prespecified plausible value (R = 0.50) and tested alternative plausible values (e.g., 0.25 and 0.75) in sensitivity analyses (25). When continuous outcomes were not reported as Mean ± SD, we attempted to convert the available summary statistics to Mean ± SD prior to exclusion. For studies reporting medians with interquartile ranges, ranges, or five-number summaries, Mean and SD were estimated using quantile-based estimation methods recommended for meta-analysis (23, 26). For studies reporting Mean with standard error (SE), SD was calculated as SD = SE × √n. When 95% confidence intervals (CI) were reported, SE was derived from the CI width using the appropriate t-quantile and then converted to SD. Studies were excluded only if conversion was infeasible due to insufficient reporting and unavailable author clarification (24, 27).

The methodological quality of the included studies was evaluated with the Cochrane Risk of Bias tool for randomized trials (RoB2). Two reviewers (HYW and SYZ) performed the assessments. The third reviewer (KHS) resolved any discrepancies that arose between the initial reviewers.

All statistical analyses were performed using R software (version 4.4.2), utilizing the meta (version 8.0-2) and metafor (version 4.6-0) packagesto assess pooled effect sizes via a random-effects model with the restricted maximum likelihood (REML) estimator. Continuous outcome measures (FBG, HbA1c, TG, and TC) were analyzed using Weighted Mean Difference (WMD) with 95% confidence intervals (CIs). WMD was selected as the effect size metric because all included studies reported objective biochemical indicators in identical units (or were converted to uniform units), thereby offering clinically interpretable results. The primary meta-analysis was conducted using change-from-baseline WMDs between groups for each outcome, with SDs of change handled as described above. To ensure comparability for WMD pooling, we harmonized units across studies before analysis. FBG, TG, and TC were converted to SI units (mmol/L) when necessary (glucose: mg/dL ÷ 18; TG: mg/dL ÷ 88.57; TC: mg/dL ÷ 38.67). HbA1c was analyzed in percent (%) as absolute percentage-point differences. The pooled effects were calculated automatically using the metacont function within the meta package. Heterogeneity among studies was assessed using the I² statistic, where values >50% indicated substantial heterogeneity and ≤50% suggested low heterogeneity.

Subgroup analyses and comparisons—considering intervention frequency, session duration, weekly exercise volume, and intervention period—were performed to identify optimal exercise parameters for T2DM management. The subgroup cut-off values were prespecified a priori in the review protocol based on clinical relevance and commonly reported Baduanjin prescriptions. Subgroup categories were defined as follows (1): intervention frequency: 3 sessions/week, 5–7 sessions/week, and >7 sessions/week (2); session duration: 20 min, 30 min, 40–45 min, and 60–90 min (3); weekly exercise volume: 60–149 min/week, 150–210 min/week, 211–300 min/week, and >300 min/week (4); intervention period: 4–8 weeks, 12–14 weeks, and 24–48 weeks. These cut-offs were not modified after data extraction and no data-driven cut-point optimization was conducted. The cut-off values were determined based on two criteria (1): The natural clustering of intervention protocols identified during data extraction. Specifically, session durations in the included studies predominantly clustered at 40–45 minutes and 60–90 minutes, aligning with the standard Health Qigong Baduanjin protocol promulgated by the General Administration of Sport of China (28), where a complete session typically includes warm-up, one or two rounds of practice, and cool-down (2). International physical activity guidelines. The weekly exercise volume subgroup of 150–210 minutes was defined to align with the American Diabetes Association (ADA) and World Health Organization (WHO) recommendations, which advise a minimum of 150 minutes of moderate-intensity physical activity per week to optimize glycemic control (29, 30). Furthermore, intervention cycles were categorized into medium-term (12–14 weeks) and long-term (24–48 weeks) to assess the temporal stability of metabolic improvements.

We additionally conducted exploratory random-effects meta-regression (REML) to investigate potential sources of heterogeneity. Potential exercise-prescription moderators were prespecified according to the extracted parameters, including frequency, session duration, weekly volume, and intervention period. A moderator was entered into the meta-regression only when it was reported as a single numeric value and showed sufficient completeness across trials. Parameters reported as ranges (e.g., “3–7 sessions/week”) or as categorical descriptions were not converted into single values for regression to avoid introducing measurement error; these were instead examined using prespecified subgroup categories. In practice, meta-regression was feasible for intervention duration (weeks) only, due to incomplete or range-reported data for other parameters.

Because pharmacological therapy may influence glycemic and lipid outcomes, medication was considered a background co-intervention, and pooled estimates were interpreted as the incremental effect of Baduanjin beyond usual care. Owing to inconsistent reporting of medication regimens (e.g., drug class, dosage, and treatment adjustments), subgroup analyses or meta-regression stratified by medication class/dose could not be conducted.

Publication bias was evaluated through funnel plot inspection and Egger’s test, with a significance threshold of P < 0.05. If significant publication bias was detected, the Duval and Tweedie non-parametric “trim and fill” method was employed to estimate the number of missing studies and adjust the effect size to verify the robustness of the results. A leave-one-out sensitivity analysis was performed, where each study was sequentially excluded to assess the robustness of the overall effect. Given the clinical heterogeneity introduced by different comparator types and potential methodological quality concerns, we additionally conducted sensitivity analyses restricting the pooled estimates to two scenarios: (1) trials using non-exercise controls only (i.e., excluding active exercise comparators); and (2) trials excluding those with a high risk of bias. This was performed for each primary outcome (FBG, HbA1c, TG, and TC).

The quality of evidence was assessed using the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach. RCTs are initially considered high-quality evidence (31). The quality rating is then adjusted according to specific criteria: factors such as limitations in study design, inconsistency of results, indirect evidence, imprecision, or publication bias each warrant a one-level downgrade, potentially to very low quality. Conversely, upgrading may occur for reasons such as a large magnitude of effect or a dose-response gradient. Any disagreements during assessment were resolved through discussion among all reviewers.

Our comprehensive search of the eight specified databases initially yielded 1109 articles. Following duplicate removal, 407 records remained. After screening titles and abstracts, we excluded 321 articles that failed to meet the inclusion criteria, resulting in 86 articles for full-text assessment. A detailed review of these full texts led to the exclusion of 38 additional articles. Ultimately, 48 RCTs were included in the final analysis (Figure 1).

After search screening, 48 studies involving 3699 participants met the inclusion criteria and were included in this meta-analysis (32–79). Among these, 42 studies evaluated the effect of Baduanjin on FBG in patients with T2DM (32–49, 51–55, 57–60, 62, 63, 67–75, 77–79). Another 41 studies assessed its effect on HbA1c levels (32–40, 42–47, 50–59, 61, 62, 64–67, 69–77, 79), while 24 studies examined its impact on TG levels (32–35, 38, 39, 42–47, 51–55, 59, 67, 69, 72, 73, 75, 79), and 22 studies investigated its effect on TC levels (32–35, 38, 39, 43, 44, 47–49, 51–55, 67, 69, 72, 73, 76, 79). The age of the patients predominantly ranged from 50 to 70 years, while one study did not report the age of its participants. In most studies, the number of male and female participants was approximately balanced at a ratio of 1:1, eight studies did not explicitly state the male-to-female ratio, and nine included two or more control groups (Table 1).

According to the RoB2 assessment, 28 studies (32–35, 41, 42, 44, 45, 47, 49, 51–53, 55, 56, 58, 59, 61, 63, 66–71, 73, 75, 78) were rated as having some concerns and 20 studies (36–40, 43, 46, 48, 50, 54, 57, 60, 62, 64, 65, 72, 74, 76, 77, 79) as high risk. Due to the nature of the Baduanjin intervention, participant and personnel blinding could not be implemented in any of the trials, thereby representing a major source of bias and contributing to the overall low quality of evidence (Figures 2, 3).

The Table 2 summarizes the subgroup analyses of Baduanjin exercise parameters on glycemic and lipid outcomes in T2DM. For transparency, the number of included studies (S) and participants (N) contributing to each subgroup are reported in Table 2.

A univariable meta-regression was performed to examine whether intervention duration (weeks) moderated the effect sizes of Baduanjin across outcomes. The Supplementary Table 3 and Supplementary Figure 1 indicated that intervention duration was a significant source of heterogeneity for HbA1c (P = 0.016, R² = 16.60%). The regression coefficient was −0.0162 (95% CI: −0.0294 to −0.0031), suggesting that for each additional week of intervention, the weighted mean difference (WMD) in HbA1c between the Baduanjin and control groups decreased by a further 0.0162 percentage points, indicating significantly greater improvement with longer intervention duration. For triglycerides (TG), the meta-regression showed a trend toward statistical significance (P = 0.067), implying that a longer Baduanjin intervention may be associated with greater TG improvement. However, no significant linear association between intervention duration and effect size was observed for fasting blood glucose (FBG; P = 0.173) or total cholesterol (TC; P = 0.283), suggesting that improvements in these outcomes may not be solely dependent on extending the intervention period.

Among the four outcome measures in this study, three (FBG, HbA1c, and TG) were supported by low–quality evidence. One measure (TC) was supported by very low–quality evidence, which limits the reliability of its effect estimate and warrants cautious interpretation (Table 3).

Funnel plots were generated to assess the risk of publication bias for outcomes reported in more than 10 studies (see Figure 8, Supplementary Table 4). The funnel plots for FBG, HbA1c, and TG were largely symmetrical. Egger’s test indicated no significant publication bias for FBG (P = 0.09), HbA1c (P = 0.60), or TG (P = 0.45), a finding consistent with the visual inspection of the funnel plots. In contrast, significant publication bias was detected for TC (P = 0.01) using both Egger’s test and funnel plot inspection. Given this bias, a trim-and-fill analysis was performed. The analysis imputed 13 potentially missing studies, resulting in an adjusted WMD of –0.77 (95% CI: –0.98 to –0.56). The adjusted result remained statistically significant (P < 0.0001), suggesting that the improvement in TC associated with Baduanjin is robust despite the potential publication bias.

Sensitivity analyses were conducted to assess the robustness of the findings using three distinct approaches.

First, the “leave-one-out” method confirmed consistent outcomes across all four measures, indicating that no single study disproportionately influenced the pooled estimates (Supplementary Figure 2).

Second, to address the potential impact of study quality, we performed an analysis excluding studies rated as “high risk of bias.” The results indicated that Baduanjin remained effective in improving metabolic indicators even when analyzing only studies with lower risk of bias. Specifically, significant reductions were maintained for FBG (WMD = –0.73, 95% CI: –0.94 to –0.52), HbA1c (WMD = –0.58, 95% CI: –0.74 to –0.42), TC (WMD = –0.32, 95% CI: –0.54 to –0.10), and TG (WMD = –0.37, 95% CI: –0.60 to –0.13) (Supplementary Figure 3).

Third, to reduce clinical heterogeneity attributable to active exercise comparators, we performed sensitivity analyses restricting comparisons to non-exercise control groups only. The pooled effects remained statistically significant and were directionally consistent with the primary analyses: FBG (MD = −0.88, 95% CI: −1.06 to −0.69, P < 0.001), HbA1c (MD = −0.72, 95% CI: −0.86 to −0.57, P < 0.001), TC (MD = −0.45, 95% CI: −0.62 to −0.28, P < 0.001), and TG (MD = −0.47, 95% CI: −0.71 to −0.23, P < 0.001) (Supplementary Figure 4). These findings collectively suggest that the observed benefits of Baduanjin are robust, regardless of study quality or comparator type.

This meta-analysis synthesizing 48 RCTs (n = 3,699) indicates that Baduanjin yields statistically significant improvements in glycemic and lipid outcomes in T2DM, including FBG, HbA1c, TG, and TC, compared with control conditions. Importantly, subgroup analyses suggest that a dose pattern of 40–45 min per session, ~3 sessions per week, sustained for 24–48 weeks may be associated with the most consistent metabolic improvements.

In exploratory univariable meta-regression, Baduanjin intervention duration (weeks) emerged as a significant moderator for HbA1c, indicating that longer Baduanjin programs were associated with larger HbA1c reductions. This is clinically plausible because HbA1c reflects average glycemic exposure over approximately 8–12 weeks and may therefore be more responsive to the cumulative effects of sustained Baduanjin practice. By contrast, Baduanjin intervention duration showed only a borderline trend for TG and no significant linear association for FBG or TC, suggesting that changes in these outcomes may be influenced by additional factors (e.g., differences in comparators, baseline metabolic profiles, adherence, or medication adjustments) rather than the duration of Baduanjin alone.

Recent syntheses focusing on Baduanjin (or mind–body exercise sets that include Baduanjin) generally support beneficial effects on glycemic control and cardiometabolic markers, but they also highlight substantial limitations, including heterogeneous training prescriptions, variable comparators, and trial-level risks of bias (20). In addition, earlier reviews rarely quantified “dose” in a manner that supports clinical prescription (80). Our findings extend this literature by providing parameterized estimates of training dosage most consistently associated with improvements across glucose and lipid outcomes, thereby refining Baduanjin-specific exercise prescription rather than reiterating generic exercise recommendations.

Mechanistically, Baduanjin is characterized by coordinated low-to-moderate intensity movements integrated with breath regulation, relaxation, and meditative attention, which may engage metabolic benefits through mind–body mechanisms beyond energy expenditure alone. First, mind–body practice may facilitate autonomic nervous system (ANS) rebalancing, often indexed by heart rate variability (HRV). In T2DM, impaired autonomic regulation (e.g., reduced HRV) is clinically relevant, and exercise training has been shown to improve HRV parameters in T2DM populations, supporting ANS modulation as a plausible pathway (81).

Second, the meditative components of Baduanjin (paced breathing, mindful attention, relaxation) may reduce chronic stress and hypothalamic-pituitary-adrenal (HPA) axis activation; cortisol biology is closely linked to insulin resistance and cardiometabolic risk. Recent evidence indicates physical activity interventions can lower cortisol and improve sleep-related outcomes (82, 83), and mind–body exercise reviews propose neuroendocrine and inflammatory modulation (including stress-related pathways) as key contributors to health effects (84).

Third, emerging Baduanjin-related trials in other clinical contexts suggest improvements in HRV and recovery-related autonomic indices, which is consistent with a parasympathetic-enhancing, stress-reducing profile (19).

Nevertheless, we emphasize that many Baduanjin RCTs in T2DM do not routinely measure mechanistic biomarkers (e.g., HRV, cortisol), so the mechanistic discussion should be interpreted as biologically plausible and literature-supported, but not definitive for the included trials.

This study has several strengths. First, to our knowledge, it is among the most comprehensive syntheses aiming to parameterize Baduanjin dosage (frequency, session duration, weekly volume, and intervention cycle) for glycolipid outcomes in T2DM, rather than merely estimating overall efficacy. Second, we used RoB 2 for risk-of-bias assessment and GRADE to rate the certainty of evidence, alongside sensitivity analyses, providing a transparent appraisal of evidence strength. Third, by integrating subgroup analyses across multiple dosage dimensions, we offer clinically interpretable, testable hypotheses for Baduanjin prescription in future high-quality trials.

Our findings support Baduanjin as a feasible adjunct to usual care for adults with T2DM, particularly for individuals who prefer culturally grounded, low-barrier mind–body activity. Given its integration of coordinated movement, breath regulation, and meditative attention, Baduanjin may be especially relevant when clinicians aim to improve metabolic control alongside stress-related barriers to self-management (e.g., poor sleep, anxiety, low exercise adherence).

Importantly, the dosage pattern identified in this review should be interpreted as a preliminary suggestion rather than a definitive clinical prescription. These subgroup definitions were selected to mirror clinically actionable components of a Baduanjin prescription (sessions/week, min/session, min/week, and intervention duration in weeks) and to align with commonly used protocol structures and physical activity guidance, thereby improving the translational interpretability of the subgroup findings. Subgroup analyses across multiple dosage dimensions suggested that approximately 40–45 minutes per session, around three sessions per week, sustained for 24–48 weeks may be associated with the most consistent improvements across glycemic and lipid outcomes. Consistent with this, exploratory univariable meta-regression indicated that longer Baduanjin intervention duration (weeks) was associated with greater HbA1c reduction, supporting the notion that sustained practice may be particularly relevant for longer-term glycemic control. In practice, clinicians should individualize recommendations according to age, comorbidity burden, baseline activity, and safety considerations, with ongoing monitoring of metabolic response and appropriate treatment adjustments as needed.

Several limitations warrant caution. First, overall trial quality was suboptimal: no included RCT was rated as “low risk of bias” under RoB 2, and blinding is inherently difficult for exercise interventions, increasing risks of performance and detection bias. Second, heterogeneity was substantial across outcomes, reflecting differences in participant characteristics, co-interventions, adherence, and outcome measurements. Third, evidence certainty was limited (low for FBG/HbA1c/TG and very low for TC), and publication bias was detected for TC, reducing confidence in the pooled estimates.

Fourth, ‘optimal parameters’ were inferred from prespecified subgroup analyses and exploratory univariable meta-regression. Given the potential for false positives arising from multiple comparisons, these findings should be interpreted with caution as hypothesis-generating rather than definitive prescriptions, especially where subgroup cells were supported by few studies. Notably, meta-regression was limited to Baduanjin intervention duration (weeks) because training frequency, session duration, and weekly practice time were frequently reported as ranges (e.g., “3–7 sessions/week”) or categorical/qualitative descriptions and were incompletely reported across trials; therefore, these parameters were not converted into single continuous values for regression and were primarily evaluated using prespecified subgroup categories. Consequently, residual heterogeneity is likely influenced by other unmeasured or incompletely reported factors, and our ability to examine multiple Baduanjin dose moderators simultaneously was limited.

Fifth, our reliance on prespecified subgroup categories to summarize Baduanjin “dose” may not fully capture the complexity of continuous exposure–outcome relationships. Methodological guidance highlights that continuous variables often have non-linear associations, and categorization can lead to loss of information and potentially misleading inferences about dose–response patterns (85, 86). Therefore, our subgroup-derived “optimal parameters” should be interpreted as hypothesis-generating rather than precision prescriptions. Future trials and meta-analyses with more granular reporting (or individual participant data) are needed to evaluate potential non-linear dose–response relationships using flexible approaches (e.g., splines or fractional polynomials).

Sixth, reporting of key training characteristics (notably intensity, supervision, and adherence) was inconsistent, limiting dose–response inference. Finally, background pharmacotherapy (drug class/dose changes) may confound metabolic outcomes; although many trials aimed for stable medication regimens, reporting was incomplete and residual confounding cannot be ruled out. In addition, comparator types varied across trials (non-exercise/usual care vs. active exercise), which may contribute to clinical heterogeneity; therefore, we provided comparator-restricted sensitivity analyses to improve interpretability. Future large, rigorously designed trials should standardize Baduanjin protocols, report intensity/adherence comprehensively, and incorporate mechanistic endpoints (e.g., HRV, cortisol, inflammatory markers) to clarify pathways of effect.