Exercise modalities and dosages on glycated haemoglobin and exerkines in type 2 diabetes: a systematic review with pairwise, network, and dose-response meta-analyses Authors: Li-bing Liang, Zhi-hao Zhao, Wen-Xuan Li, Shan Wang, Qun-song Shen, Cai-qin Wu, Kun-peng Li Published: 2026 ABSTRACT BackgroundExercise is a cornerstone of type 2 diabetes mellitus (T2DM) management; however, the comparative efficacy of specific modalities, optimal dosages, and associated exerkine responses remains incompletely defined. This study aimed to synthesise evidence on the effects of different exercise types and doses on glycated haemoglobin (HbA1c) levels and to explore modality-specific patterns of exerkine regulation in adults with T2DM.MethodsA systematic review and Bayesian evidence synthesis of randomised controlled trials was conducted up to October 2025. Pairwise, network, and model-based dose-response meta-analyses were employed to evaluate the effects of eight exercise modalities on HbA1c levels. Exploratory analyses were performed to evaluate changes in circulating adiponectin, brain-derived neurotrophic factor (BDNF), interleukin-6, leptin, resistin, and tumour necrosis factor-α levels.ResultsA total of 127 trials involving 8,744 participants were included. Network meta-analysis ranked combined mind-body resistance training (tai chi + elastic bands) and high-intensity interval training (HIIT) as the most effective modalities for reducing the level of HbA1c (surface under the cumulative ranking curve: 85.9% and 78.6%, respectively). An L-shaped nonlinear dose-response relationship was observed, with the majority of the clinical benefit being achieved at 600 MET-min/week (mean difference: -0.57%; 95% credible interval: -0.72, -0.42) and a plateau beyond approximately 1,000 MET-min/week. Exploratory exerkine analyses suggested modality-specific signatures; specifically, HIIT was associated with greater increases in adiponectin levels, whereas resistance-based interventions trended towards greater reductions in interleukin-6 and tumour necrosis factor-α levels. In pairwise comparisons versus the control, no significant overall changes were detected for adiponectin, BDNF, interleukin-6, or resistin.ConclusionIn adults with T2DM, exercise efficacy varies by modality and follows a dose-plateau pattern. The achievement of 600 MET-min/week produces a meaningful reduction in HbA1c levels, with diminishing returns being observed beyond approximately 1,000 MET-min/week. MBRT and HIIT are the most promising modalities, although the evidence concerning MBRT remains preliminary in nature. Modality-specific exerkine patterns were observed; specifically, HIIT ranked highest for adiponectin, whereas RT ranked highest for IL-6 and TNF-α. These findings inform evidence-based, personalised exercise prescription in T2DM patients.Graphical abstract Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12933-026-03236-7. Research insights What is currently known about this topic? Exercise benefits T2D; however, the best form and amount of exercise remain uncertain.Exerkines may mediate benefits; however, clear causal links are lacking.Guidelines suggest general volume targets (not modality-specific doses). Exercise benefits T2D; however, the best form and amount of exercise remain uncertain. Exercise benefits T2D; however, the best form and amount of exercise remain uncertain. Exerkines may mediate benefits; however, clear causal links are lacking. Exerkines may mediate benefits; however, clear causal links are lacking. Guidelines suggest general volume targets (not modality-specific doses). Guidelines suggest general volume targets (not modality-specific doses). What is the key research question? What are the comparative efficacy and dose-response relationships of different exercise modalities for improving HbA1c in T2D, and how are they associated with exerkine profiles? What are the comparative efficacy and dose-response relationships of different exercise modalities for improving HbA1c in T2D, and how are they associated with exerkine profiles? What are the comparative efficacy and dose-response relationships of different exercise modalities for improving HbA1c in T2D, and how are they associated with exerkine profiles? What is new? An L-shaped dose-response curve was identified, thereby indicating that the majority of the HbA1c reduction is achieved at 600 MET-min/week, with diminishing returns being observed beyond 1000 MET-min/week.A combined modality of tai chi and elastic band resistance training demonstrated potential but requires further validation due to limited data.Exerkine responses appeared to be modality-specific (e.g., HIIT was associated with adiponectin, and RT was associated with IL-6/TNF-α). An L-shaped dose-response curve was identified, thereby indicating that the majority of the HbA1c reduction is achieved at 600 MET-min/week, with diminishing returns being observed beyond 1000 MET-min/week. An L-shaped dose-response curve was identified, thereby indicating that the majority of the HbA1c reduction is achieved at 600 MET-min/week, with diminishing returns being observed beyond 1000 MET-min/week. A combined modality of tai chi and elastic band resistance training demonstrated potential but requires further validation due to limited data. A combined modality of tai chi and elastic band resistance training demonstrated potential but requires further validation due to limited data. Exerkine responses appeared to be modality-specific (e.g., HIIT was associated with adiponectin, and RT was associated with IL-6/TNF-α). Exerkine responses appeared to be modality-specific (e.g., HIIT was associated with adiponectin, and RT was associated with IL-6/TNF-α). How might this study influence clinical practice? The prioritisation of the minimum effective dose (600 MET-min/week) of efficient modalities like tai chi + elastic bands can enhance feasibility and long-term glycaemic control. The prioritisation of the minimum effective dose (600 MET-min/week) of efficient modalities like tai chi + elastic bands can enhance feasibility and long-term glycaemic control. The prioritisation of the minimum effective dose (600 MET-min/week) of efficient modalities like tai chi + elastic bands can enhance feasibility and long-term glycaemic control. Introduction Type 2 diabetes mellitus (T2DM) is a chronic metabolic disorder characterised by persistent hyperglycaemia and represents a major contributor to global morbidity and mortality [1]. T2DM currently affects 10.5% of the worldwide population [2], and its prevalence is projected to increase to 12.2% by 2045, thereby further exacerbating societal and health-care burdens [2, 3]. Glycated haemoglobin (HbA1c) reflects average blood glucose concentrations over the preceding 2–3 months and serves as the principal clinical indicator of long-term glycaemic control [4]. Each 1% increase in HbA1c is associated with a 1.13-fold higher risk of cardiovascular disease and a 1.26-fold increased risk of stroke [5, 6], thus underscoring its robust association with long-term diabetic complications. Regular physical exercise is universally recommended as a first-line intervention for T2DM management [7]; however, the optimal modality and dosage for reducing HbA1c levels remain uncertain. Current guidelines rely on syntheses of heterogeneous evidence [8, 9], and previous systematic reviews have used broad exercise categories that may obscure modality-specific physiological effects [10–12]. Emerging hybrid regimens (such as high-intensity interval training combined with resistance exercise or mind-body practices integrated with resistance training) have not been fully incorporated into existing evidence syntheses [13, 14]. Moreover, the categorisation of mind-body exercises such as yoga and tai chi under conventional exercise types may introduce classification bias [15, 16]. Dose parameters (including intensity, duration, and frequency) are central factors to the glycaemic response; however, precise dose-response relationships across modalities remain ill-defined [10]. Therefore, the accurate elucidation of how specific exercise modalities and their doses influence HbA1c is essential for developing evidence-based exercise prescriptions. Moreover, the elucidation of the biological mediators that transduce the benefits of physical activity has emerged as a research priority with translational potential. A central factor involved in this inquiry is the concept of exerkines. Exerkines are bioactive molecules released into the circulation from multiple tissues (including the skeletal muscle, adipose tissue, and liver) in response to acute or chronic exercise [17]. Initially defined by contraction-induced myokines such as interleukin-6 (IL-6), the exerkine repertoire has expanded to encompass adipokines (such as adiponectin and leptin), hepatokines, and inflammatory cytokines, in which circulating levels are modulated by sustained exercise training [18]. In patients with T2DM, accumulating evidence indicates that exercise training alters the circulating profile of exerkines implicated in glycaemic regulation. A meta-analysis of 40 randomised controlled trials demonstrated that structured exercise significantly increases adiponectin, fibroblast growth factor-21, and interleukin-10 levels while reducing fetuin-A, IL-6, leptin, resistin, and tumour necrosis factor-α (TNF-α) levels. These favourable changes in pooled exerkines were associated with improvements in HbA1c, fasting glucose, and waist circumference [19]. Furthermore, mechanistic studies in humans with T2DM have revealed that an acute bout of aerobic exercise elicits a distinct inflammatory and cytokine response in skeletal muscle that is potentiated relative to that in individuals with normal glucose tolerance [20]. This process includes enhanced induction of chemokines such as CXCL12, thus suggesting that metabolic disease status may prime the exerkine milieu in a manner relevant to tissue remodelling and metabolic adaptation [20]. Importantly, exerkine responses appear to be highly mode-dependent [21]. High-intensity interval training (HIIT) preferentially elevates plasma IL-6 levels with modest effects on IL-10 levels, whereas resistance-based protocols may exert more pronounced effects on inflammatory cytokines such as TNF-α [22]. These observations support the notion that different exercise modalities engage distinct molecular pathways and exerkine networks [23]. Nevertheless, systematic comparisons across the full spectrum of contemporary exercise modalities are lacking. It is also unclear as to whether modality-specific exerkine signatures are correlated with improvements in HbA1c in a causal manner or merely as associative biomarkers [18]. The pleiotropic and sometimes contradictory metabolic effects of individual exerkines further underscore the complexity of exerkine biology and the need for integrated, network-level analyses. For example, the role of IL-6 in lipolysis and glucose homeostasis remains disputed [24–27]. Currently, the preponderance of exerkine research has centred on aerobic exercise. The specific regulatory effects of other modalities (such as resistance training, HIIT, and combined mind‑body regimens) on exerkine profiles and their associations with improvements in HbA1c have not been systematically compared or synthesised [28, 29]. The elucidation of how distinct exercise modalities influence glycaemic homeostasis through divergent exerkine regulatory pathways remains an important objective for informing individualised exercise therapy. Therefore, this study aimed to synthesise the available evidence on the comparative effects of eight distinct exercise types and their dosages on HbA1c in patients with T2DM through an integrated framework of pairwise, network, and dose-response network meta-analyses. As an exploratory objective, it further compared the modulation of key circulating exerkines across exercise regimens and examined associations between changes in HbA1c and alterations in these biomarkers. This integrated approach offers a novel perspective by simultaneously examining modality efficacy, dose-response relationships, and modality-associated exerkine signatures within a unified analytical framework. The findings are intended to inform the development of evidence‑based, personalised exercise prescriptions for T2DM. Methods Protocol registration This meta-analysis followed the PRISMA extension for Network Meta-Analysis [30] and the Cochrane Handbook for Systematic Reviews of Interventions [31]. The protocol was registered in PROSPERO (CRD420251166357). Ethical approval and patient consent were not needed. Search strategy The PubMed, Embase, Web of Science, Scopus, SPORTDiscus, CINAHL, and Cochrane Library databases were searched from inception to October 24, 2025. The reference lists of relevant reviews were manually screened. The search terms included “Diabetes Mellitus, Type 2”, “Exercise”, “Exercise Therapy”, “Resistance Training”, and “Glycated Hemoglobin”. The full strategy is provided in the Supplementary Materials (pp. 1–5). Eligibility criteria Studies included in this analysis were required to adhere to the following criteria, which were structured according to the PICOS framework. (1) Participants: adults aged 18 years or older who were diagnosed with T2DM. (2) Interventions: comparative studies of different exercise modalities or intensities or comparisons between an exercise intervention group and a control group. Exercise was defined as planned, structured, and repetitive physical activity [32]. Furthermore, as the formation of persistent physiological adaptations typically requires an initial period of patient acclimatisation, an intervention duration of at least 4 weeks was stipulated to establish a stable therapeutic effect [33]. (3) Comparators: control groups receiving usual care, placed on a waitlist, or assigned to a different exercise regimen. (4) Outcomes: reported changes in HbA1c as a result of the intervention, which were expressed in traditional percentage units aligned with the National Glycohemoglobin Standardization Program (https://ngsp.org/). (5) Study design: randomised controlled trials (RCTs) (Table 1). Table 1Definitions of exercise training interventionsModalityDefinitionAerobic exercise (AE)Activities such as brisk walking, cycling, and jogging—sustained physical exercises designed to enhance cardiovascular efficiency and endurance[34].Resistance exercise (RT)Resistance training of all types, including structured circuit training programs that utilise modalities such as resistance bands, free weights, and weight machines [34].Combined aerobic and resistance exercise (AERT)Integrated modality of exercise that combines aerobic exercise and resistance training.High-intensity interval training (HIIT)High-intensity interval training consisting of repeated, relatively brief bursts of all-out effort performed at intensities exceeding 90% of peak oxygen consumption or heart rate, as exemplified by sprint interval training protocols[35, 36].Combined high-intensity interval training and resistance exercise (HIITRT)Integrated modality of exercise that combines high-intensity interval training and resistance training.Mind-body exercise (MB)Mind-body exercises encompassing disciplines such as yoga, tai chi, and Qigong that are defined by their integrative focus on breath control, postural alignment, and movement precision, with practices such as pilates additionally emphasising core strength, stability, and muscular control[37, 38].Combined mind-body and resistance exercise (MBRT)Integrated modality of exercise that combines mind-body exercise and resistance training (the two included MBRT studies combined simplified 24-form tai chi with elastic band resistance training; variations in training dose and band resistance existed across the studies). Table 1Definitions of exercise training interventionsModalityDefinitionAerobic exercise (AE)Activities such as brisk walking, cycling, and jogging—sustained physical exercises designed to enhance cardiovascular efficiency and endurance[34].Resistance exercise (RT)Resistance training of all types, including structured circuit training programs that utilise modalities such as resistance bands, free weights, and weight machines [34].Combined aerobic and resistance exercise (AERT)Integrated modality of exercise that combines aerobic exercise and resistance training.High-intensity interval training (HIIT)High-intensity interval training consisting of repeated, relatively brief bursts of all-out effort performed at intensities exceeding 90% of peak oxygen consumption or heart rate, as exemplified by sprint interval training protocols[35, 36].Combined high-intensity interval training and resistance exercise (HIITRT)Integrated modality of exercise that combines high-intensity interval training and resistance training.Mind-body exercise (MB)Mind-body exercises encompassing disciplines such as yoga, tai chi, and Qigong that are defined by their integrative focus on breath control, postural alignment, and movement precision, with practices such as pilates additionally emphasising core strength, stability, and muscular control[37, 38].Combined mind-body and resistance exercise (MBRT)Integrated modality of exercise that combines mind-body exercise and resistance training (the two included MBRT studies combined simplified 24-form tai chi with elastic band resistance training; variations in training dose and band resistance existed across the studies). Definitions of exercise training interventions Studies were excluded according to the following criteria: (1) acute studies with an intervention duration of less than 4 weeks; (2) studies employing a crossover design; (3) conference abstracts, study protocols, and systematic reviews; (4) studies from which appropriate data could not be obtained; and (5) studies involving patients with conditions such as impaired glucose tolerance, insulin resistance, or prediabetes (rather than those diagnosed with type 2 diabetes). Data collection process Two reviewers independently screened the titles and abstracts and then assessed the full texts by using EndNote 20. Disagreements were resolved by a third reviewer. The extracted data included author, year, participant characteristics (including age, sex, BMI, waist circumference, diabetes duration, and comorbidities), intervention details (such as modality, duration, frequency, and intensity), exerkines, and HbA1c. Missing data were requested from corresponding authors up to three times over the course of three weeks. Data coding and management The exercise dose was expressed as MET-min/week (MET value × session duration × weekly frequency). MET values were assigned by using the 2011 Compendium of Physical Activities [39]; modality-specific codes are provided in the Supplementary Materials. With respect to resistance training, MET values reflect the aerobic metabolic cost (rather than the internal muscle load) and may lead to an underestimation of the true physiological stimulus. With respect to mind-body exercises, METs do not capture key therapeutic elements such as neuromuscular activation, core stability demands, or neuroendocrine regulation. These limitations are particularly relevant for combined modalities such as MBRT, wherein the linear summation of MET-min/week from mechanistically distinct components may not accurately reflect biological doses. Therefore, MBRT dose estimates (especially at the lower end of the spectrum) should be interpreted with caution. Exercise frequency was recorded as sessions per week. Only primary exercise duration was extracted; moreover, warm-up and cool-down exercises were excluded. Missing session durations were imputed by using the modality-specific mean. For progressive protocols, the average duration was used. Weekly MET-min values were stratified into seven prespecified categories: 0 (control), 250, 500, 750, 1000, 1250, and 1500 MET-min/week. This categorisation was based on three considerations: (1) consistency with prior systematic reviews to facilitate comparability [40]; (2) a 250 MET-min/week interval providing adequate data density while maintaining model stability, which is consistent with dose-response network meta-analysis guidance [41]; and (3) a 1500 MET-min/week interval, which represents the upper limit that is commonly observed in clinical practice. Categories were prespecified and not data-driven. Risk of bias and certainty of evidence Two reviewers assessed the risk of bias by using RoB 2 [42], with discrepancies being resolved by a third reviewer. HbA1c comparison credibility was evaluated by using CINeMA [43] across the following six domains: within-study bias, reporting bias, indirectness, imprecision, heterogeneity, and incoherence. Imprecision was defined using Hedges’ g = 0.5. Evidence quality was graded as high, moderate, low, or very low according to GRADE [44]. Measures of treatment effect Treatment effects were analysed by using the mean difference (MD) and change standard deviation (SD). Missing SDs were derived from standard errors, 95% CIs, p values, or t statistics [31]. A correlation coefficient of 0.5 was applied to calculate the SD of change (assuming moderate repeatability) [31]. Statistical analysis Pairwise meta-analyses and publication bias For modalities with > 3 studies, random-effects pairwise meta-analyses were performed. Control group sample sizes were halved for multiarm studies to avoid double-counting [31]. For pooled exerkine analysis, the effect directions were reversed for IL-6, leptin, resistin, fetuin-A, visfatin, and TNF-α to ensure directional consistency [45]. This exploratory analysis provided an overview of exercise-induced exerkine changes. When a study reported multiple exerkines, only the largest effect size was retained. Heterogeneity was assessed by using the I² statistic (> 50% indicated substantial heterogeneity) [46]. Publication bias was evaluated via funnel plots, the Egger’s test [47], and the Begg’s test [48]. NMAs Transitivity was evaluated by comparing baseline HbA1c levels, age, diabetes duration, comorbidities, and BMI across the treatment groups (Supplementary Tables 1–2) [49]. Bayesian network meta-analysis was performed by using the “Gemtc” R package with Markov chain Monte Carlo methods. Four chains ran for 25,000 iterations, with the first 5,000 iterations being discarded as burn-in [50]. Convergence was assessed with the Brooks-Gelman-Rubin diagnostic [51]. Fixed effects, random effects consistency, and inconsistency models were fitted. Global inconsistency was evaluated by comparing the DIC and variance parameters [52]. Between-study heterogeneity (τ) was estimated. Local inconsistency was examined via node splitting (p < 0.05 indicated significant disagreement). SUCRA values, rankograms, forest plots, and league tables were generated from the consistency model [53]. Dose-response NMAs Model-based network meta-analysis (MBNMA) with a random-effects Bayesian framework was used to examine the dose-response relationship between exercise volume and HbA1c [54]. Transitivity, consistency, and connectivity assumptions were satisfied (Supplementary Materials, pp. 22–26) [41, 55, 56]. Effect sizes are expressed as the MD with 95% CrI. Dose-response functions (including Emax, restricted cubic splines, quadratic, and nonparametric functions) were compared [57]. The Emax model demonstrated an optimal fit (Supplementary Table 25) and was selected for nonlinear assessment [58, 59]. Statistical significance was defined as a 95% CrI excluding zero. The analyses used the “MBNMAdose” R package; additionally, visualisations used “ggplot2”. Additional analyses Meta-regression explored moderators using a Bayesian linear mixed model with multiple imputation. Associations between changes in HbA1c levels and age, intervention duration, BMI change, and changes in four exerkines (adiponectin, TNF-α, IL-6, and leptin) were assessed. Diabetes duration, background medication type, and medication changes during the intervention were not systematically extractable and were omitted. Continuous variables were standardised. Missing values were imputed by using the “mice” R package (10 datasets). The models included a study-level random intercept and fixed effects for exercise type and covariates. Priors were observed to be normal (0,1) for fixed effects and exponential (1) for standard deviations. The results were pooled by using Rubin’s rules. A positive coefficient indicated an association with a reduction in the level of HbA1c. Sensitivity analyses excluded studies with small sample sizes (n < 15) and a high risk of bias. An additional sensitivity analysis excluded the two MBRT studies to evaluate the robustness of the network meta-analysis rankings and dose-response estimates given the limited direct evidence for this modality. Results Study selection and characteristics A total of 10,486 records were identified. After duplicate removal and title/abstract screening, 324 full-text articles were assessed, and 127 RCTs involving 8,744 participants were included (Fig. 1). The mean age ranged from 35.0 to 71.5 years, with a maximum mean diabetes duration of 21 years. AE was the most frequent intervention, followed by AERT. The exercise frequency ranged from 1 to 7 weekly sessions. All of the interventions were supervised. Only two studies were available for HIITRT and MBRT. Both MBRT studies combined tai chi with elastic band resistance training. The detailed study characteristics and sample sizes per modality are provided in Supplementary Tables 1 and 2. Fig. 1Preferred reporting items for systematic reviews and meta-analyses flow diagram Preferred reporting items for systematic reviews and meta-analyses flow diagram Risk of bias Among the 127 studies included in this meta-analysis, 14 were rated as having a high risk of bias, 84 elicited some concerns, and 29 were assessed as having a low risk of bias. The detailed results of the risk of bias assessment are presented in Supplementary Figure 1. Certainty of evidence CINeMA assessment indicated that evidence certainty for HbA1c comparisons ranged from low to moderate; moreover, no comparison achieved high certainty. Moderate certainty was assigned to direct comparisons of AE, AERT, HIIT, MB, MBRT, and RT versus control, as well as to several direct and indirect comparisons involving MBRT. All of the other comparisons exhibited low certainty. The main factors leading to downgrading included within‑study bias (due to the inherent lack of blinding in exercise trials), imprecision (due to limited sample sizes for certain comparisons), and heterogeneity (due to differences in intervention protocols across the studies). The full assessments are shown in Fig. 2. Fig. 2Summary table for credibility assessment using confidence in network meta-analysis (CINeMA). Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise. CON, control group Summary table for credibility assessment using confidence in network meta-analysis (CINeMA). Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise. CON, control group Pairwise meta-analyses and publication bias A pairwise meta-analysis demonstrated that all of the exercise modalities with at least three available studies significantly reduced HbA1c levels compared with the control (Supplementary Figure 2). Heterogeneity was substantial for most comparisons (I² > 50%). The results of funnel plot asymmetry and the Egger’s test (p = 0.0086) suggested potential publication bias, whereas the Begg’s test was nonsignificant (p = 0.2955) (Supplementary Materials, pp. 16–19). Exercise significantly decreased TNF-α (MD = −2.01, 95% CI [−3.83, −0.19], p < 0.001) and leptin levels (MD = −1.29, 95% CI [−2.44, −0.15], p < 0.001) relative to the control. No significant changes were observed for adiponectin, BDNF, IL-6, or resistin levels (Fig. 3). An exploratory pooled analysis combining multiple exerkines revealed an overall favourable directional change (MD = 1.15, 95% CI [0.67, 1.62], p < 0.001; Fig. 4). Given the exploratory nature and limited sample sizes, these findings should be interpreted with caution. Fig. 3Effect of exercise interventions on exerkines in patients with type 2 diabetes. a: Forest plot for IL-6. b: Forest plot for adiponectin levels. c: Forest plot for leptin. d: Forest plot for BDNF expression. e: Forest plot for TNF-α. f: Forest plot for resistin. Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise Effect of exercise interventions on exerkines in patients with type 2 diabetes. a: Forest plot for IL-6. b: Forest plot for adiponectin levels. c: Forest plot for leptin. d: Forest plot for BDNF expression. e: Forest plot for TNF-α. f: Forest plot for resistin. Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise Fig. 4Effects of exercise interventions on the pooled exerkine profile in patients with type 2 diabetes. Note: TNF-a, tumour necrosis factor-alpha. IL-6, interleukin-6. BDNF, brain-derived neurotrophic factor. FGF-21, fibroblast growth factor 21. AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise Effects of exercise interventions on the pooled exerkine profile in patients with type 2 diabetes. Note: TNF-a, tumour necrosis factor-alpha. IL-6, interleukin-6. BDNF, brain-derived neurotrophic factor. FGF-21, fibroblast growth factor 21. AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise NMAs The network geometry and SUCRA rankings are shown in Fig. 5. MBRT was ranked highest (SUCRA: 85.9%), followed by HIIT (78.6%), AERT (75.0%), MB (55.1%), AE (43.2%), RT (32.4%), and HIITRT (23.8%). MBRT included two RCTs that used tai chi with elastic bands; the protocols varied in frequency, duration, and resistance. The high SUCRA ranking reflects limited direct evidence and should be interpreted with caution. Compared with the control, significant decreases in HbA1c were observed for MBRT (MD = −0.96, 95% CrI [−1.73, −0.21]), HIIT (MD = −0.71 [−0.94, −0.50]), AERT (MD = −0.68 [−0.83, −0.54]), MB (MD = −0.57 [−0.84, −0.29]), AE (MD = −0.50 [−0.62, −0.38]), and RT (MD = −0.44 [−0.59, −0.29]) (Supplementary Table 17). Fig. 5Network diagram (a) and SUCRA rankings (b) for HbA1c. Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise. CON, control group Network diagram (a) and SUCRA rankings (b) for HbA1c. Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise. CON, control group The network diagrams and SUCRA rankings for adiponectin, IL-6, leptin, and TNF-α are shown in Fig. 6. HIIT was ranked highest for adiponectin (SUCRA: 65.4%); RT was ranked highest for IL-6 (79.9%) and TNF-α (70.9%); and AE was ranked highest for leptin (84.1%). These rankings reflect relative probabilities and do not imply statistically significant differences versus the control. No exercise modality significantly affected adiponectin, leptin, IL‑6, or TNF‑α levels compared with those in the control group. Detailed results are provided in the Supplementary Materials (pp. 29–30). Fig. 6Network diagram (a) and SUCRA rankings (b) for TNF-a; network diagram (c) and SUCRA rankings (d) for leptin; network diagram (e) and SUCRA rankings (f) for IL-6; network diagram (g) and SUCRA rankings (h) for adiponectin. Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise. CON, control group Network diagram (a) and SUCRA rankings (b) for TNF-a; network diagram (c) and SUCRA rankings (d) for leptin; network diagram (e) and SUCRA rankings (f) for IL-6; network diagram (g) and SUCRA rankings (h) for adiponectin. Note: AE, aerobic exercise. RT, resistance exercise. AERT, combined aerobic and resistance exercise. HIIT, high-intensity interval training. HIITRT, combined high-intensity interval training and resistance exercise. MB, mind-body exercise. MBRT, combined mind-body exercise and resistance exercise. CON, control group The model fitting results and findings from the node-splitting analysis are reported in the Supplementary Materials (pp. 23–26). No significant global inconsistency was detected. Node-splitting revealed no local inconsistency for HbA1c or TNF-α (p > 0.05) but did reveal a significant inconsistency between AE and RT for adiponectin and IL-6 (p < 0.05). Leptin analysis was not feasible because of the insufficient number of studies. Dose-response NMAs An L-shaped nonlinear dose-response relationship was observed between total exercise volume and HbA1c reduction (Fig. 7). Although this pattern is consistent with that of prior meta-analyses using similar methodologies, it should be interpreted with caution because discretising continuous doses into predefined categories may influence the fitted curve shape. A significant HbA1c reduction was evident beginning at 0 MET-min/week. The minimal clinically important difference (MCID) threshold of ≥ 0.5% was achieved at 239 MET-min/week. At the lower WHO recommendation (600 MET-min/week), the predicted reduction was MD = −0.57% (95% CrI [−0.72, −0.42]). At the upper WHO recommendation (1200 MET-min/week), the predicted reduction was MD = −0.59% (95% CrI [−0.73, −0.43]). At 1500 MET-min/week, the predicted reduction was MD = −0.59% (95% CrI [−0.74, −0.44]). The marginal benefit markedly decelerated after approximately 101 MET-min/week and plateaued beyond 1000 MET-min/week. Notably, fewer studies and participants were available in the high-dose range (> 1000 MET-min/week) than in the low-to-moderate range. Fig. 7Dose-response relationship between total weekly exercise volume and HbA1c. Note: Shaded band, 95% credible interval. Blue shaded area, WHO-recommended weekly energy expenditure range. Red dashed line, ADA MCID threshold (≥ 0.5%). Purple dashed line, exercise dosage required to achieve the MCID. MD, mean difference. METs, metabolic equivalents Dose-response relationship between total weekly exercise volume and HbA1c. Note: Shaded band, 95% credible interval. Blue shaded area, WHO-recommended weekly energy expenditure range. Red dashed line, ADA MCID threshold (≥ 0.5%). Purple dashed line, exercise dosage required to achieve the MCID. MD, mean difference. METs, metabolic equivalents Modality-specific dose-response curves are shown in Fig. 8. No significant relationship was observed for HIITRT. Distinct L-shaped nonlinear relationships were identified for AE, AERT, RT, HIIT, MB, and MBRT. The dosage required to achieve an MCID of ≥ 0.5% was highest for RT (1351 MET-min/week; MD = −0.50%, 95% CrI [−0.69, −0.32]), followed by AE (800 MET-min/week; MD = −0.50%, 95% CrI [−0.62, −0.37]), MB (239 MET-min/week; MD = −0.51%, 95% CrI [−0.77, −0.25]), ACERT (83 MET-min/week; MD = −0.50%, 95% CrI [−0.69, −0.24]), HIIT (40 MET-min/week; MD = −0.50%, 95% CrI [−0.73, −0.15]), and MBRT (33 MET-min/week; MD = −0.50%, 95% CrI [−1.11, −0.06]). The estimated minimal effective dose for MBRT (33 MET-min/week) lacks physiological plausibility and was not corroborated by the sensitivity analysis excluding MBRT. This finding likely reflects limited data, protocol heterogeneity, and model extrapolation at the lower dose spectrum (rather than a true biological phenomenon). Table 2 presents the minimum exercise recommendations. Fig. 8Dose-response relationships between specific exercise modalities and HbA1c. Note: Purple-shaded bands, 95% credible intervals where the relationship is statistically significant. Red dashed line, ADA MCID threshold (≥ 0.5%). Purple dashed lines, dosage required to achieve the MCID for each modality Dose-response relationships between specific exercise modalities and HbA1c. Note: Purple-shaded bands, 95% credible intervals where the relationship is statistically significant. Red dashed line, ADA MCID threshold (≥ 0.5%). Purple dashed lines, dosage required to achieve the MCID for each modality Table 2Exercise recommendations for patients with type 2 diabetesType of exerciseMinimal clinically effective dose(METs-min/week)IntensityEnergy expenditureb(METs-min)Minimum recommendations for exercise prescriptionc(sessions × min/per week)ACE800Moderate4 (code 17270)、 4.8 (code 17260)5 × 405 × 30Vigorous7.0 (code 12020)、 8.0 (code 01030)3 × 403 × 30RT1351Moderate3.5 (code 02054)5 × 45ACERT83Moderate7.0 (code 12020) and 3.5 (code 02054)2 × 10HIIT40Vigorous11 (code 02014\12070)2 × 10MB239Moderate3 (code 15670/02105)、 2.5 (code 02150)6 × 303 × 40MBRT33Moderate3 (code 15670/02105) and 3.5 (code 02054)1 × 10MET codes from Ainsworth et al. [39]. AE, aerobic exercise; RT, resistance exercise; AERT, combined aerobic and resistance exercise; HIIT, high-intensity interval training; MB, mind-body exercise; MBRT, combined mind-body and resistance exercise. The MBRT estimate (33 MET-min/week) is based on limited data and MET-based quantification; moreover, it lacks physiological plausibility and was not confirmed in sensitivity analyses excluding MBRT Table 2Exercise recommendations for patients with type 2 diabetesType of exerciseMinimal clinically effective dose(METs-min/week)IntensityEnergy expenditureb(METs-min)Minimum recommendations for exercise prescriptionc(sessions × min/per week)ACE800Moderate4 (code 17270)、 4.8 (code 17260)5 × 405 × 30Vigorous7.0 (code 12020)、 8.0 (code 01030)3 × 403 × 30RT1351Moderate3.5 (code 02054)5 × 45ACERT83Moderate7.0 (code 12020) and 3.5 (code 02054)2 × 10HIIT40Vigorous11 (code 02014\12070)2 × 10MB239Moderate3 (code 15670/02105)、 2.5 (code 02150)6 × 303 × 40MBRT33Moderate3 (code 15670/02105) and 3.5 (code 02054)1 × 10MET codes from Ainsworth et al. [39]. AE, aerobic exercise; RT, resistance exercise; AERT, combined aerobic and resistance exercise; HIIT, high-intensity interval training; MB, mind-body exercise; MBRT, combined mind-body and resistance exercise. The MBRT estimate (33 MET-min/week) is based on limited data and MET-based quantification; moreover, it lacks physiological plausibility and was not confirmed in sensitivity analyses excluding MBRT Exercise recommendations for patients with type 2 diabetes 5 × 40 5 × 30 3 × 40 3 × 30 6 × 30 3 × 40 MET codes from Ainsworth et al. [39]. AE, aerobic exercise; RT, resistance exercise; AERT, combined aerobic and resistance exercise; HIIT, high-intensity interval training; MB, mind-body exercise; MBRT, combined mind-body and resistance exercise. The MBRT estimate (33 MET-min/week) is based on limited data and MET-based quantification; moreover, it lacks physiological plausibility and was not confirmed in sensitivity analyses excluding MBRT Sensitivity analyses, subgroup analyses, and network meta-regression The exclusion of studies with small sample sizes (n < 15) and a high risk of bias did not alter the intervention rankings. An additional sensitivity analysis excluding the two MBRT studies ranked HIIT as the most effective strategy (SUCRA: 84.2%), followed by ACERT (79.1%). The L-shaped dose-response curve and plateau above 1000 MET-min/week remained unchanged, although MBRT-specific estimates could not be verified. Detailed results are provided in the Supplementary Materials (pp. 31–34). Subgroup analyses were conducted by intervention duration (< 8 vs. ≥8 weeks, given that HbA1c reflects 8–12-week glycaemia) and geographic region (reflecting modality distribution: aerobic/HIIT being more common in Western studies, and mind-body being more common in Asian studies). No significant subgroup differences were observed (p = 0.30 and p = 0.07, respectively). The detailed results are shown in Supplementary Figs. 7–8. Meta-regression revealed no significant associations between changes in HbA1c and changes in TNF-α, adiponectin, IL-6, leptin, BMI, baseline age, or intervention duration. The detailed results are shown in Supplementary Fig. 16. Discussion Principal findings and novelty This systematic review synthesises evidence from 127 RCTs (8,744 participants) to compare eight exercise modalities, quantify dose-response relationships, and explore exerkine profiles for glycaemic control in T2DM. Two primary contributions emerged. First, a network meta-analysis revealed that MBRT (tai chi plus elastic band resistance) and HIIT were the highest-ranked modalities for reducing HbA1c, whereas a dose-response analysis revealed an L-shaped relationship, with a substantial benefit at 600 MET-min/week and a plateau beyond approximately 1,000 MET-min/week being observed. Second, exploratory analyses revealed distinct modality-associated exerkine patterns, with HIIT ranking highest for adiponectin and RT ranking highest for IL-6 and TNF-α. The study population predominantly included middle-aged and older adults without overt cardiovascular disease; therefore, the findings of this study are most directly applicable to this demographic. Optimal modality and dose-response: synthesis and clinical significance Network meta-analysis ranked MBRT highest for HbA1c reduction (SUCRA: 85.9%), followed by HIIT (78.6%) and AERT (75.0%). Compared with the control, MBRT reduced HbA1c by 0.96% (95% CrI: −1.73 to −0.21). This observation aligns with evidence that combined modalities offer advantages over single-component training [10, 11] but extends prior syntheses by incorporating mind-body integration. MBRT combines two mechanistically distinct stimuli; specifically, resistance exercise enhances glucose uptake via AMPK activation, increased muscle mass, and improved insulin signalling [60–62], whereas mind-body practices reduce sympathetic overactivation and cortisol levels, thereby improving autonomic balance and systemic inflammation [63, 64]. Whether these pathways interact synergistically remains a testable hypothesis that requires direct empirical validation. The MBRT findings are derived from two RCTs that combined simplified 24-form tai chi with elastic band resistance. Sensitivity analysis excluding these studies ranked HIIT highest (SUCRA: 84.2%), thus indicating that the top ranking of MBRT is preliminary and reflects limited direct evidence. Moreover, mind-body evidence predominantly originates from Asian populations, whereas aerobic and resistance training data are derived largely from Western cohorts. Future trials should evaluate MBRT in diverse populations and directly compare it with established efficacious modalities. Clinically, MBRT represents a promising option among several effective modalities, with selection being guided by patient preference, physical capacity, and program accessibility. The L-shaped dose-response relationship provides quantitative guidance for exercise prescription. The achievement of the WHO minimum of 600 MET-min/week yielded an HbA1c reduction of 0.57% (95% CrI: −0.72 to −0.42), with a diminishing incremental benefit beyond approximately 1,000 MET-min/week being observed. This plateau effect, which is consistent with recent meta-analyses [10, 65], supports a pragmatic emphasis on consistently attaining the minimum effective dose rather than pursuing the maximal volume. Such an approach may enhance long-term adherence while preserving therapeutic efficacy. Combined and high-intensity modalities (including MBRT, HIIT, and AERT) offer the greatest reduction in HbA1c per unit of exercise volume. The lower minimum effective dose observed in this study compared with some prior reports [10, 65] likely reflects differences in the included modalities and modelling methods. The MBRT dose estimate (33 MET-min/week) lacks physiological plausibility and reflects MET-based quantification limitations. For combined modalities, linear MET summation from distinct components assumes unvalidated additive effects, thereby underscoring the need for alternative dose frameworks. Exerkine patterns: exploratory observations and interpretive considerations Network meta-analysis revealed modest differences in exerkine-associated rankings. HIIT was ranked highest for adiponectin (SUCRA: 65.4%); RT was ranked highest for IL-6 (79.9%) and TNF-α (70.9%); and AE was ranked highest for leptin (84.1%). A pairwise meta-analysis revealed that overall exercise interventions significantly reduced leptin and TNF-α levels. However, for adiponectin, BDNF, IL-6, and resistin, the pooled effects did not achieve statistical significance compared with those of the control group, and substantial heterogeneity (I² > 50%) was observed across studies for these outcomes. Meta-regression revealed no significant associations between exerkine changes and HbA1c improvement, which is likely due to limited sample sizes, assay heterogeneity, and temporal mismatch between acute biomarker fluctuations and chronic glycaemic measurements. Consequently, these exerkine signatures are exploratory and hypothesis-generating. The observed modality-associated patterns are directionally consistent with known acute exercise physiology. HIIT imposes repeated, intense metabolic stress on skeletal and cardiac muscle, thereby activating AMPK and PPAR-δ pathways that enhance glycolipid metabolism and may promote functional remodelling of white adipose tissue [66]. This remodelling is characterised by increased mitochondrial biogenesis and an improved adipokine secretory profile, including elevated levels of adiponectin (which is an insulin-sensitising and anti-inflammatory hormone) [18]. The association between HIIT and higher adiponectin rankings aligns with evidence that HIIT effectively improves cardiorespiratory fitness in individuals with T2DM [67], and this factor is closely linked to systemic metabolic enhancement. In contrast, RT exerts metabolic benefits primarily through the endocrine and immunomodulatory functions of skeletal muscle. The contraction-induced release of IL-6 stimulates anti-inflammatory cascades, including those involving IL-1RA and IL-10, which systemically suppress proinflammatory pathways such as TNF-α [68]. Long-term RT increases muscle mass and regularly induces a moderate IL-6 response, thereby ameliorating low-grade inflammation and optimising the microenvironment for insulin signalling. The higher rankings of RT for IL-6 and TNF-α reduction are consistent with this anti-inflammatory profile. AE (which is a sustained moderate-intensity activity) promotes lipolysis and fatty acid oxidation, which may contribute to reductions in leptin (an adipokine that is closely linked to fat mass) [19]. The favourable leptin ranking for AE aligns with its established role in reducing adiposity and improving energy homeostasis. Two important interpretive caveats merit emphasis. First, IL-6 is defined as an exerkine based on its acute secretion from contracting skeletal muscle (not its chronic decline after training). Transient exercise-induced IL-6 increases trigger downstream anti-inflammatory cascades, whereas lower resting levels reflect broader long-term adaptation. Second, mechanistic explanations derived from acute exercise physiology may not be directly extrapolated to chronic interventions. None of the included trials directly measured the mechanistic endpoints described above (e.g., adipose tissue UCP1 expression or circulating IL-1RA or IL-10 levels) after chronic training. Thus, although the observed modality-associated patterns are directionally consistent with acute mechanistic hypotheses, they do not constitute direct pathway evidence. Future studies incorporating serial mechanistic biomarkers during chronic training are needed to bridge acute insights with long-term outcomes. Limitations Several limitations warrant consideration. First, the certainty of the evidence ranged from low to moderate across the comparisons, which is mainly due to lack of blinding, imprecision, and heterogeneity. Second, the MBRT findings were derived from only two RCTs with protocol heterogeneity. The high SUCRA ranking reflects the probability within a limited evidence network, and sensitivity analysis excluding MBRT altered the ranking order. Third, MET-based dose quantification has inherent limitations. Specifically, METs capture aerobic cost but not internal muscle loading or neuroendocrine regulation. For MBRT, linear MET summation from distinct components lacks biological validation. The 33 MET-min/week estimate should be regarded as a methodological artefact. Dose-response analyses used discrete categories, and high-dose data were sparse, which may influence curve shape and plateau identification. Fourth, exerkine analyses were exploratory, underpowered, and unadjusted for multiple comparisons. Incomplete reporting of medications, diet, and diabetes duration precluded comprehensive confounding adjustments. Fifth, most of the trials excluded patients with overt cardiovascular disease and possessed short-to-moderate duration without concrete clinical endpoints. Mind-body studies predominantly included Asian studies, whereas aerobic and resistance trials largely included Western studies. Age- and sex-specific responses could not be evaluated because of study-level reporting limitations. Conclusions In adults with T2DM, exercise efficacy exhibits clear mode-dependency and a dose-plateau effect. The achievement of 600 MET-min/week yields a substantial HbA1c reduction. MBRT (tai chi plus elastic band resistance) and HIIT emerged as promising modalities, although the evidence concerning MBRT remains preliminary and requires confirmatory trials. Exploratory exerkine analyses revealed the following modality-associated patterns: HIIT was ranked highest for adiponectin, whereas RT was ranked highest for IL-6 and TNF-α. These findings provide a quantitative foundation for evidence-based, personalised exercise prescription in T2DM patients. Supplementary Information Supplementary Material 1 Supplementary Material 1