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

Background Exercise is established as a non-pharmacological strategy for type 2 diabetes mellitus (T2DM), but heterogeneity across glycemic outcomes, modalities and doses, and the statistical dependence induced by multiple effect sizes per trial, limit the clinical interpretability of conventional pooled estimates. Objective To quantify the effect of structured exercise on glycemic control in T2DM while handling dependent effect sizes, and to separate dose-response evidence from exploratory analyses of heterogeneity. Methods Fifty randomized controlled trials contributing 123 effect sizes were pooled in a three-level random-effects meta-analysis, with subgroup analyses, univariate meta-regression and 95% prediction intervals. Effects were expressed as post-intervention standardized mean differences (Hedges’ g), and absolute mean differences were pooled for each outcome. Exploratory MetaForest with recursive pre-screening, study-grouped cross-validation and KernelSHAP was applied to 89 effect sizes with complete moderator records. Results Exercise significantly improved glycemic control (g = -0.738, 95% CI -1.046 to -0.430, p < 0.001), corresponding to reductions of 0.52% in HbA1c, 1.13 mmol/L in fasting glucose and 1.87 mmol/L in 2-hour postprandial glucose. Heterogeneity was substantial (I2 = 86.0%, tau2 = 1.125; 95% prediction interval -2.98 to 1.50). All three outcomes improved significantly, whereas between-modality differences were not significant (p = 0.544). Intervention duration (beta = -0.152 per month, p = 0.018) and weekly exercise dose (beta = -0.101 per MET-h/week, p < 0.001) were associated with larger effects; intensity was not (p = 0.282). MetaForest ranked diabetes duration, age and BMI highest, but study-grouped cross-validation showed no out-of-sample accuracy (R2 = -0.11). Conclusion Structured exercise meaningfully improves glycemic control in T2DM. Duration and total weekly volume, rather than intensity alone, were associated with larger effects, but heterogeneity was high and prediction intervals were wide, so the pooled estimate describes an average effect rather than an individual response. HIIT and combined training showed the largest point estimates without significant between-modality differences. Machine-learning findings are hypothesis-generating only. Systematic review registration https://www.crd.york.ac.uk/prospero/ , identifier CRD420261416860.