Structured Exercise Interventions Improve Glycemic Control in Adults With Type 2 Diabetes

A comprehensive three-level meta-analysis of fifty randomized controlled trials demonstrates that structured exercise interventions deliver statistically significant improvements in glycemic control for adults with type 2 diabetes mellitus. Synthesizing data across 123 effect sizes, the investigation revealed an overall post-intervention standardized mean difference of minus 0.738, confirming a medium-to-large beneficial reduction in circulating glucose compared to control conditions. Translated into clinical metrics across individual outcomes, structured exercise yielded average reductions of 0.52 percent in glycated hemoglobin, 1.13 millimoles per liter in fasting blood glucose, and 1.87 millimoles per liter in two-hour postprandial glucose. While these pooled figures highlight therapeutic efficacy, the analysis uncovered high between-trial heterogeneity alongside wide 95 percent prediction intervals spanning from minus 2.98 to 1.50, demonstrating that the aggregate summary reflects an average population benefit rather than a uniform individual response.

The findings address a burgeoning global metabolic disease crisis characterized by steep escalations in diabetes prevalence and chronic vascular complications. Citing epidemiological surveillance from the eleventh edition of the International Diabetes Federation Diabetes Atlas published in 2025, the researchers noted that approximately 589 million adults aged twenty to seventy-nine were living with diabetes worldwide in 2024, representing about one in every nine adults. Over ninety percent of these individuals have type 2 diabetes mellitus, and international projections indicate this burden will reach 852 million people by 2045, with 252 million adults currently living undiagnosed. Within China, epidemiological research published in 2025 by the Chinese Center for Disease Control and Prevention documented 233 million people with diabetes in 2023, reflecting an age-standardized prevalence increase from 7.53 percent in 2005 to 13.67 percent in 2023. This disease burden highlights the critical need for effective non-pharmacological management strategies.

To synthesize existing evidence, the research team searched five electronic databases without language restrictions: Web of Science, PubMed, Cochrane Library, Embase, and the China National Knowledge Infrastructure through July 1, 2026. The search combined controlled vocabulary terms and keywords covering diabetes mellitus, exercise modalities, and glycemic outcomes. After eliminating 1,178 duplicate citations, two independent reviewers screened 11,662 unique records, assessing 2,525 full-text articles against pre-specified criteria. Included trials were randomized controlled designs evaluating adults aged eighteen or older with type 2 diabetes who underwent structured exercise compared to a non-exercise control group. Following full-text retrieval, one trial published in 2017 was excluded because it enrolled prediabetic individuals rather than participants with diagnosed type 2 diabetes. All statistical models were subsequently recalculated without this study, establishing a final analytical dataset of fifty randomized controlled trials contributing 123 effect sizes.

A primary methodological focus involved accounting for the statistical dependence of multiple effect sizes contributed by trials reporting several glycemic measures or testing multiple exercise arms. The researchers implemented a three-level random-effects meta-analysis in R using the metafor package, assigning sampling variance to level one, within-study variance to level two, and between-study variance to level three. Restricted maximum likelihood estimation yielded a level-two variance component of zero at the parameter boundary, reducing the three-level framework to a standard two-level study model. An empirical study-aggregated re-analysis verified this equivalence, yielding identical pooled effects of minus 0.738 and an identical between-study variance of 1.125. While the study-aggregated model attributed within-trial variance to sampling error, producing an I-squared of 93.8 percent, the three-level model reported an I-squared of 86.0 percent. The substantial heterogeneity yielded a 95 percent prediction interval from minus 2.98 to 1.50, demonstrating broad dispersion across settings.

The fifty analyzed trials exhibited diverse clinical characteristics, participant demographics, and training protocols published between 1986 and 2024. Across reporting trials, participant cohort mean ages spanned from 41.5 to 78.9 years, while body mass index averages extended from 23.4 to 37.1 kilograms per square meter. Sample sizes ranged from 7 to 77 participants in intervention arms and 4 to 77 in control arms. Intervention durations ranged from one to twelve months, with all trials satisfying operational minimums of at least four weeks and two sessions weekly. Prescribed intensity spanned 2 to 8 metabolic equivalents of task, and weekly exercise doses ranged between 180 and 2,700 MET-minutes per week. Evaluated modalities encompassed aerobic training across sixty effect sizes from twenty-three trials, combined aerobic and resistance training across nineteen effect sizes, resistance training across sixteen effect sizes, and high-intensity interval training across six effect sizes from three trials.

Analysis of specific glycemic indicators confirmed that structured exercise produced statistically significant improvements across all three primary laboratory measures. For glycated hemoglobin, pooling fifty-six effect sizes across forty-two trials yielded a standardized mean difference of minus 0.640, corresponding to an absolute reduction of 0.52 percent. For fasting blood glucose, pooling forty-six effect sizes across thirty-six trials produced a standardized mean difference of minus 0.730, reflecting an absolute reduction of 1.13 millimoles per liter. For two-hour postprandial blood glucose, pooling twenty-one effect sizes across sixteen trials yielded a standardized mean difference of minus 0.596, representing an absolute reduction of 1.87 millimoles per liter. An omnibus subgroup test revealed no statistically significant differences across these three endpoints, yielding a p-value of 0.740. This consistency indicates that exercise exerts broad glycemic regulation, improving both basal glucose production and postprandial disposal capacity.

Subgroup analyses comparing exercise modalities demonstrated differences in point estimates, although between-modality variations were not statistically significant. High-intensity interval training generated the largest point estimate at minus 1.897, followed by combined training at minus 1.156, continuous aerobic training at minus 0.678, and resistance training at minus 0.487. Mind-body exercise showed a negative direction of effect at minus 0.456, but did not attain statistical significance, registering a p-value of 0.060 across nine effect sizes. Crucially, the omnibus test for subgroup differences across modalities was not significant, yielding a Q statistic of 7.91 and a p-value of 0.544. High-intensity interval training comprised only six effect sizes from three trials, resulting in very wide confidence and prediction intervals. Consequently, the authors emphasized that higher point estimates cannot be interpreted as empirical evidence of clinical superiority over foundational aerobic and resistance exercise regimens.

Univariate meta-regression analyses evaluated continuous moderating factors to clarify which prescription variables explain differences in glycemic outcomes. Weekly exercise dose showed a statistically significant association with greater glucose-lowering effects, yielding a regression coefficient beta of minus 0.101 per MET-hour per week, which corresponds to approximately minus 0.168 per 100 MET-minutes per week and accounted for 26.03 percent of between-study variance. Intervention duration was likewise significantly associated with greater glycemic benefit, with a beta of minus 0.152 per month explaining 8.48 percent of variance. Conversely, exercise intensity alone showed no statistically significant moderating effect, explaining only 0.13 percent of variance. Baseline diabetes duration, patient age, and body mass index also failed to demonstrate significant linear relationships. Because exercise dose was calculated from protocol-assigned MET values rather than direct energy expenditure, the researchers cautioned that these findings reflect study-level protocol associations rather than individual targets.

To explore non-linear moderator interactions, the research team applied MetaForest, a random-forest implementation adapted for meta-analysis, across eighty-nine effect sizes from thirty-one trials with complete moderator profiles. Nine candidate moderators were evaluated: diabetes duration, age, female participant proportion, body mass index, exercise duration, exercise intensity, weekly exercise dose, baseline glucose, and overall study risk of bias. Recursive pre-screening over one hundred replications retained all nine candidate variables with a selection rate of 1.00, reflecting the limited sample size relative to candidate parameters. Hyperparameters selected through cross-validation used two split variables per node, a minimum node size of five, and one thousand regression trees. Permutation importance metrics ranked diabetes duration as the most influential moderator, followed by age, body mass index, and female participant proportion, while weekly exercise dose and intervention duration ranked fifth and sixth.

Complementing permutation importance, KernelSHAP analysis was employed to examine variable contributions, ranking diabetes duration, age, and body mass index as the top three contributors, followed by exercise intensity and intervention duration. Crucially, the researchers emphasized that the machine learning results lacked out-of-sample predictive validity. When cross-validation folds were structured at the study level to prevent data leakage between correlated effect sizes, the cross-validated R-squared dropped to minus 0.11. By contrast, an ungrounded effect-size-level split yielded an inflated R-squared of 0.54, illustrating how dividing correlated estimates from single trials across training and test sets creates substantial optimism. Because the model demonstrated no capacity to forecast outcomes in unseen trials, the authors strictly framed the machine learning findings as exploratory and hypothesis-generating for future stratified clinical investigations rather than an empirical foundation for personalized exercise prescriptions.

Methodological quality assessments and sensitivity analyses reinforced the overall stability of the primary findings while outlining essential caveats. Risk of bias evaluations using the Revised Cochrane RoB 2 tool rated eighty-two effect sizes as low risk, eight as having some concerns, and thirty-three as high risk. Excluding high-risk effect sizes produced an essentially unchanged pooled standardized mean difference of minus 0.766 across ninety effect sizes from forty trials. Publication bias testing detected no statistically significant small-study asymmetry via Egger regression or Begg rank correlation, while leave-one-out sensitivity analyses yielded consistently significant pooled estimates ranging between minus 0.761 and minus 0.617. Trim-and-fill modeling imputed sixteen missing studies, shifting the adjusted pooled estimate to minus 1.095. Acknowledging substantial statistical heterogeneity, wide prediction intervals crossing zero, and reliance on assigned MET proxies, the authors concluded that structured exercise is an essential cornerstone of diabetes care, where cumulative volume, adherence, and patient safety should take precedence over pursuing a single optimal modality.

Subject of Research: Exercise interventions and glycemic control in type 2 diabetes mellitus

Article Title: Exercise interventions for glycemic control in adults with type 2 diabetes mellitus: a three-level meta-analysis with exploratory machine-learning analysis

Article References: Jiang, T., Jia, Y., & Zhang, H. (2026). Exercise interventions for glycemic control in adults with type 2 diabetes mellitus: a three-level meta-analysis with exploratory machine-learning analysis. Frontiers in Endocrinology, 17, Article 1961425. https://doi.org/10.3389/fendo.2026.1961425

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1961425

Keywords: type 2 diabetes mellitus, glycemic control, structured exercise, three-level meta-analysis, glycated hemoglobin, fasting blood glucose, 2-hour postprandial glucose, aerobic training, resistance training, high-intensity interval training, combined exercise training, meta-regression, MetaForest, random-effects model, weekly exercise dose, Shapley additive explanations, prediction intervals, RoB 2 assessment