This systematic review and network meta-analysis of RCTs followed PRISMA 2020 and its NMA extension. The protocol was registered in PROSPERO (CRD420251133289), and all procedures adhered to Cochrane Handbook standards (22).

We systematically searched four major electronic databases—PubMed, Embase, Web of Science, and the Cochrane Library—from their inception to August 7, 2025. The search strategy combined controlled vocabulary (MeSH/Emtree) and free-text terms related to type 2 diabetes mellitus, mind–body exercise (e.g., Yoga, Tai chi, Qigong, Pilates, and Mindfulness-based interventions), and randomized controlled trials, and was adapted to each database. The search was limited to English-language publications, as these databases predominantly index peer-reviewed biomedical literature in English and the majority of randomized controlled trials in this field are published in English. Trial registries and grey literature were not systematically searched to ensure data completeness and methodological consistency for quantitative synthesis; however, reference lists of relevant reviews and included studies were manually screened to minimize the risk of missing eligible trials. A total of 2,809 records were identified and managed using EndNote software. The full search strategies are provided in Supplementary 1A–D.

Studies were included if they met the PICOS framework: patients diagnosed with type 2 diabetes mellitus (T2DM) without restrictions on age, sex, disease duration, or comorbidities; interventions involving at least one mind–body exercise (e.g., Pilates, Yoga, Tai chi, Qigong, Mindfulness-based interventions, Biofeedback-Assisted Relaxation Training, Guided Imagery Training); comparators such as usual care, diabetes education, lifestyle modification (defined as non–exercise-based interventions focusing on dietary advice, diabetes education, or general lifestyle counseling, and explicitly excluding any structured mind–body exercise components), waitlist control, or other non–mind–body approaches; outcomes reporting at least one psychological (depression, anxiety, stress) or physical function measure (balance, walking ability, muscular strength, aerobic capacity) with sufficient data for effect size calculation; and randomized controlled trials (RCTs) published in English.

Exclusion criteria were non-randomized designs (e.g., cohort, case–control, pre–post studies); studies not specific to T2DM or with mixed populations lacking separate T2DM data; interventions outside the scope of mind–body exercise (e.g., pharmacological, dietary, conventional aerobic/resistance training); studies without relevant outcomes or adequate data; and conference abstracts, case reports, reviews, protocols, or duplicate publications.

Two reviewers independently screened the literature and extracted data, with discrepancies resolved through discussion. During screening, titles and abstracts were first reviewed to exclude irrelevant studies, followed by full-text assessment to determine eligibility. Extracted information included the first author, country and year of publication, participants’ baseline characteristics, sample sizes of intervention and control groups, intervention frequency and duration, details of the intervention protocols, outcome measures, and corresponding results. All references were managed using EndNote X9, and the full texts of eligible studies were retrieved for data extraction. Data were then organized systematically to ensure accuracy and completeness.

The risk of bias of the included studies was assessed using the Cochrane Risk of Bias 2.0 (RoB 2.0) tool, which evaluates five key domains: (1) the randomization process, (2) deviations from intended interventions, (3) missing outcome data (4), outcome measurement, and (5) selective reporting of results. Each domain was judged as “low risk,” “some concerns,” or “high risk,” and an overall risk-of-bias judgment was subsequently derived for each study (23).

A frequentist network meta-analysis was conducted using Stata 17.0. Network plots were constructed to illustrate the geometry of available direct comparisons, followed by network meta-analyses for each outcome. Global inconsistency was first assessed; a consistency model was applied when no significant inconsistency was detected (P > 0.05), whereas an inconsistency model was used when inconsistency was significant (P ≤ 0.05) (24, 25). When closed loops with both direct and indirect evidence were present, local inconsistency was further explored using the node-splitting approach (26). All outcomes were continuous variables and summarized using mean differences (MD) for outcomes measured on the same scale or standardized mean differences (SMD) for outcomes measured on different scales. For multi-arm trials, shared comparator groups were appropriately split to avoid double counting (24). When both baseline and endpoint data were available, change scores were preferentially used; otherwise, endpoint values were analyzed. Standard deviations missing but reported as standard errors, 95% confidence intervals, or interquartile ranges were converted according to the Cochrane Handbook (27). Effect directions were harmonized so that positive SMD values consistently indicated beneficial intervention effects. Psychological outcomes (depression, anxiety, and stress) were interpreted according to their original scoring direction, with lower scores reflecting symptom improvement. For physical function outcomes, measures in which lower values indicated better performance were inverted prior to analysis to ensure directional consistency. Intervention ranking was estimated using the surface under the cumulative ranking curve (SUCRA), with higher values indicating a greater probability of being the most effective intervention (28). Results were presented as forest plots, league tables, and ranking plots, and comparison-adjusted funnel plots were used to assess potential publication bias and small-study effects (25, 29).

Leave-one-out sensitivity analyses were performed by sequentially excluding each study and recalculating the pooled effect estimate using a random-effects model. The results indicated that the overall effect was largely driven by one influential study, while exclusion of other studies did not materially change the direction of the pooled estimate (30).

To ensure objectivity and consistency, two reviewers independently conducted the assessments and documented their judgments in the data extraction form. Any discrepancies were resolved through discussion or consultation with a third reviewer. The results were summarized in graphical and tabular formats and served as the basis for subsequent quality of evidence evaluations using the CINeMA frameworks (31, 32).

A total of 2,809 records were identified through database searching (Cochrane Library, n = 874; Embase, n = 859; Web of Science, n = 753; PubMed, n = 323). After removal of duplicate records (n = 683), review and meta-analysis articles (n = 360), conference abstracts (n = 223), trial registry records (n = 169), and letters or correspondence (n = 14), 1,360 records remained for title and abstract screening. Following initial screening, 1,297 records were excluded for not meeting the eligibility criteria, leaving 63 articles for full-text assessment. Of these, 46 studies were further excluded due to irrelevant interventions or comparators (n = 23), absence of extractable outcome data (n = 19), non-randomized study design (n = 5), or inclusion of participants without type 2 diabetes mellitus (n = 3). Ultimately, 13 randomized controlled trials (RCTs) (33–45) were included in the systematic review and network meta-analysis. The study selection process is illustrated in Figure 1.

A total of 13 randomized controlled trials (RCTs), published between 2005 and 2024, were ultimately included, comprising approximately more than 500 participants with type 2 diabetes mellitus (T2DM). The studies were conducted across multiple countries, including the United States, China, Iran, Japan, Thailand, Brazil, Germany, and Australia, providing reasonable international representation. Sample sizes of individual trials ranged from 15 to 120 participants. The age of participants varied widely, spanning from adolescents aged 14–17 years at high risk of T2DM to older adults aged over 78 years, with most studies focusing on middle-aged and older populations. Female participants predominated across studies, and several trials exclusively enrolled women. T2DM diagnosis was primarily based on American Diabetes Association (ADA) criteria or clinical diagnosis by physicians, with some studies additionally applying thresholds for glycated hemoglobin (HbA1c) or fasting plasma glucose. Diabetes duration ranged from newly diagnosed or not explicitly reported to more than 10 years. Baseline HbA1c levels generally indicated suboptimal glycaemic control, ranging approximately from 6.1% to 9.0%. Interventions mainly consisted of distinct mind–body and exercise-related approaches, including yoga (2 trials), mindfulness-based interventions (3 trials), Tai chi (2 trials), Pilates (1 trial), Yi Ren Medical Qigong–based mind–body exercise (YRMQ; 1 trial), Walking Meditation (1 trial), Laughter yoga (1 trial), Biofeedback-Assisted Relaxation Training (BFRT; 1 trial), Progressive Resistance Training (PRT; 1 trial), and Guided Imagery Training (GI; 1 trial). Intervention durations most commonly ranged from 8 to 16 weeks, with the longest lasting up to 6 months. Intervention frequency was typically 1–3 sessions per week, with session durations of 30–90 minutes. Control conditions included usual medical care, diabetes education, lifestyle advice, wait-list control, or sham exercise/low-intensity stretching. Outcome measures covered both physical function and psychological domains. Physical outcomes included walking capacity, balance, aerobic fitness, and muscle strength (e.g., the 6-minute walk test, 10-m walk test, balance assessments, and strength measures), while psychological outcomes primarily focused on depression, anxiety, and perceived stress, commonly assessed using validated instruments such as the BDI/BDI-II, STAI, and PSS. Detailed study characteristics are presented in Supplementary 2.

According to the Risk of Bias 2.0 (RoB 2.0) assessment, most included studies were judged to be at low risk of bias in the domains of the randomization process, missing outcome data, and selection of the reported results. However, some concerns were commonly identified in the domains of bias due to deviations from intended interventions and bias in measurement of outcomes. Overall, the majority of studies were rated as having some concerns in terms of overall risk of bias, with only a small number of trials classified as low risk. The detailed RoB 2.0 assessments are presented in Supplementary 4A.

Leave-one-out sensitivity analyses indicated variable robustness across outcomes. Muscle strength and balance outcomes remained stable after exclusion of any single study, whereas depression and stress outcomes were sensitive to influential trials, particularly study 9 (and study 23 for stress), with effect estimates becoming non-significant after exclusion. Anxiety and walking ability showed consistent directions of effect, although the precision of the estimates was influenced by the removal of individual studies. In contrast, aerobic capacity results were highly unstable due to the inclusion of only two trials. Overall, several outcomes were sensitive to individual studies and should therefore be interpreted with caution. Detailed results are presented in Supplementary 6.

Funnel plot inspection suggested that the overall distributions of studies across outcomes were largely symmetrical, indicating an acceptable risk of publication bias. For balance ability and muscle strength outcomes, the distribution of study points showed slight asymmetry, suggesting the possible presence of small-study effects. In contrast, the funnel plots for aerobic capacity and walking ability appeared relatively regular and symmetrical, with no clear evidence of publication bias. Among psychological outcomes, the distribution of studies for depression was relatively scattered with noticeable asymmetry, indicating a potential risk of publication bias, while the funnel plot for anxiety showed mild asymmetry, possibly attributable to limited sample sizes. The funnel plot for stress outcomes demonstrated an overall symmetrical distribution with evenly dispersed study points, suggesting no obvious publication bias. Detailed funnel plots are provided in Supplementary 5D.

Quality of evidence assessment using the CINeMA framework showed marked variability in the credibility of network comparisons across outcomes. Overall, the certainty of evidence was predominantly low to very low, although some direct comparisons reached moderate or high confidence. For balance ability, the comparison between control and Tai chi (CON vs. Tai chi) was rated as high confidence, while that between control and Pilates (CON vs. Pilates) was rated as moderate confidence; the indirect comparison between Pilates and Tai chi was rated as very low confidence due to serious heterogeneity and inconsistency. For aerobic capacity, evidence was mainly derived from single studies, with the control versus Tai chi comparison rated as moderate confidence and most other comparisons rated as low to very low confidence because of imprecision and network inconsistency. For walking ability and muscle strength, most comparisons were rated as low or very low confidence, largely owing to limited study numbers, imprecision, and heterogeneity. For psychological outcomes (depression, anxiety, and stress), overall certainty was low, with direct comparisons mostly rated as low confidence and indirect comparisons almost uniformly rated as very low confidence due to serious inconsistency and imprecision. Detailed assessments are provided in Supplementary 4B.

This study compared the relative effects of multiple mind–body exercise interventions on physical function and psychological outcomes using network meta-analysis. The findings indicated that potential advantages varied across outcomes, with Pilates ranking higher for balance and walking ability, Walking Meditation for aerobic capacity, and Yoga for muscle strength, whereas differences among interventions were generally limited for psychological outcomes (depression, anxiety, and stress). Importantly, most comparisons were informed by a small number of randomized controlled trials, and several outcomes were affected by inconsistency or imprecision. Moreover, CINeMA assessments indicated that the overall certainty of evidence ranged from low to very low. Therefore, the present findings should be interpreted primarily as exploratory, intended to highlight potential intervention signals rather than to establish definitive comparative effectiveness.

Compared with previous studies, the findings of the present analysis show both consistency and incremental contributions across several domains. A 12-week Pilates program has been reported to significantly improve glycaemic control, muscle strength, gait speed, balance, and flexibility in older adults with type 2 diabetes mellitus (T2DM) (46). Building on this evidence, the present network meta-analysis further supports the relative advantage of Pilates in improving balance and walking ability. With regard to mindfulness-based interventions, prior systematic reviews have predominantly focused on populations with general chronic conditions or psychological disorders, including university students, and have consistently demonstrated significant effects in reducing anxiety and stress (47–49). In contrast, our findings indicate that Mindfulness-based interventions produced significantly greater improvements in anxiety among patients with diabetes when compared with both BFRT and control conditions, supporting the potential value of mindfulness approaches in diabetes-related mental health management. For Tai chi, previous literature has reported benefits in enhancing balance and reducing fall risk (50). However, in the present analysis, Tai Chi demonstrated effects on balance ability that were comparable to those of conventional care, with no statistically significant differences observed. Regarding YRMQ, Laughter yoga, and BFRT, existing evidence is largely derived from small, single trials with inconsistent findings. Although our results suggested potential improvement trends for these interventions, the overall quality of evidence remained low, highlighting the need for further large-scale, high-quality randomized controlled trials to clarify their effectiveness.

The present findings suggest that mind–body exercise may confer dual benefits for patients with type 2 diabetes mellitus (T2DM) through a coordinated neuro–muscular–immune–metabolic axis. At a mechanistic level, these benefits appear to arise from the simultaneous modulation of skeletal muscle metabolism and central neural–emotional regulatory circuits. On the one hand, regular low- to moderate-intensity exercise modalities such as Pilates and Tai chi can activate the PGC-1α/AMPK signaling pathway (51), thereby improving mitochondrial function and glucose utilization efficiency, These adaptations may subsequently enhance muscle strength and overall physical performance (52). On the other hand, interventions represented by mindfulness training and yoga may attenuate hyperactivation of the hypothalamic–pituitary–adrenal (HPA) axis and sympathetic nervous system (53), leading to reduced cortisol secretion and enhanced inhibitory control of limbic regions by the prefrontal cortex. These neural adaptations contribute to alleviation of anxiety and depressive symptoms, while concurrently downregulating pro-inflammatory cytokines such as IL-6 and TNF-α and upregulating anti-inflammatory mediators including IL-10 and regulatory T-cell (Treg) activity. Together, these processes promote neuroplasticity and suppress chronic low-grade inflammation (54).

Differential effects across interventions may reflect their distinct physiological emphases. Pilates places greater emphasis on core stability and proprioceptive feedback (10), which may explain its superior performance in improving balance and walking ability. Walking Meditation ranked relatively higher for aerobic capacity, likely because sustained walking-based aerobic activity activates muscle energy homeostasis pathways (e.g., AMPK/PGC-1α signaling), enhances mitochondrial biogenesis, and improves oxidative metabolic efficiency, thereby facilitating cardiorespiratory endurance adaptations (55). Yoga may show greater benefits for muscle strength, as the isometric contractions and anti-gravity postures commonly incorporated in yoga practice improve neuromuscular control and muscle fiber coordination, while simultaneously promoting exercise-induced muscular adaptations that enhance strength output (56). Mindfulness-based interventions primarily alleviate anxiety through cognitive–emotional regulation mechanisms (57). Overall, these interventions appear to disrupt the vicious cycle of stress, metabolic dysregulation, and functional impairment through multiple interacting pathways, thereby supporting integrated improvements in both psychological and physical function.

Despite the overall low certainty of evidence, the present findings provide exploratory clinical insights for the management of patients with type 2 diabetes mellitus (T2DM). Different mind–body exercise interventions showed outcome-specific potential advantages across physical function domains, suggesting that exercise prescriptions may be tailored to patients’ predominant functional limitations rather than relying on a single uniform modality. Interventions emphasizing postural control and core stability may be more suitable for individuals with functional decline or elevated fall risk, whereas rhythmic or sustained activities may better support improvements in cardiorespiratory endurance.

For psychological outcomes, between-intervention differences were generally modest and uncertain; however, relatively favorable effects observed for anxiety- and stress-related outcomes suggest that integrating psychological regulation components into exercise programs may help address the mental health burden commonly experienced by patients with T2DM. Overall, these findings underscore the potential value of individualized, multidisciplinary, mind–body–integrated approaches, while emphasizing that the results should be interpreted as supportive, exploratory evidence rather than definitive recommendations.

Several limitations should be acknowledged. Only 13 RCTs with fewer than 800 participants were included, limiting statistical power and precluding meta-regression or subgroup analyses to explore potential effect modifiers (e.g., age, sex, intervention type, or disease duration). In addition, the included trials encompassed a wide age range and heterogeneous clinical profiles, spanning from adolescents at high risk of type 2 diabetes mellitus to older adults with long-standing disease and diabetes-related complications. Such demographic and clinical heterogeneity may limit the generalizability of the findings and could partly explain the inconsistency or imprecision observed across certain outcomes. Methodological quality was generally modest, with limited blinding and substantial heterogeneity in intervention frequency, duration, intensity, and adherence reporting. Psychological outcomes were assessed using diverse instruments, reducing cross-study comparability, and most interventions were short term (8–24 weeks), with no long-term follow-up to assess sustained effects.

Future studies should focus on large, multicenter, high-quality RCTs with improved methodological rigor, standardized intervention parameters and outcome measures, and longer follow-up periods. Mechanistic investigations and adequately powered subgroup analyses are also needed to clarify differential benefits across populations and to support more precise, individualized intervention strategies.