The review, registered prospectively with PROSPERO under number CRD420261380631, set out to address two methodological gaps that the authors argue have hampered earlier syntheses. Most previous meta-analyses of exercise interventions in ADHD used traditional two-level models, which struggle to handle the non-independence of effect sizes that arises when a single trial reports multiple outcomes, multiple intervention arms, or several follow-up time points. This new analysis instead employed a three-level meta-analytic model, which nests correlated effect sizes within their parent studies and partitions variance into sampling, within-study, and between-study components. The team also chose to quantify internalizing and externalizing symptoms within the same analytical framework, something few prior reviews had done, while treating the familiar three-way split of executive function into inhibitory control, working memory, and cognitive flexibility as a secondary analysis rather than a core innovation.

The search covered six databases, including PubMed, Web of Science, Embase, PsycNET, the Cochrane Library, and SPORTDiscus, from database inception through April 2026. From 3,335 initially identified records, the researchers screened down to 25 randomized controlled trials. Participants were children and adolescents aged 6 to 18 with ADHD identified through formal clinical diagnosis, validated diagnostic interviews, symptom screening scales, or combined parent and clinician assessment. Trials were excluded if participants had comorbid mental health conditions such as anxiety disorders, depression, autism spectrum disorder, or post-traumatic stress disorder, and only English-language publications were considered. Study sizes ranged from 12 to 94 participants, with reported ages spanning 6 to 15 years, and the interventions included aerobic exercise programs, cognitive combined exercise, and comprehensive exercise regimens.

The statistical machinery was correspondingly detailed. Analyses were run in R using the metafor, clubSandwich, and ggplot2 packages, with parameters estimated by restricted maximum likelihood. Because many of the included trials had small samples, the team used the bias-corrected Hedges’ g as the effect size measure, applying cluster-robust variance estimation to correct standard errors, together with Satterthwaite’s degrees-of-freedom correction and a t-distribution test to control Type I error. Heterogeneity was assessed with I-squared statistics, and both 95 percent confidence intervals and prediction intervals were reported, the latter reflecting the range within which a true effect might fall in a new setting. Sensitivity checks included leave-one-out analyses and influence diagnostics flagging effect sizes with a Cook’s distance more than three times the mean or an absolute studentised residual above 3.

The central result concerned inhibitory control, one of the three core executive function components widely recognized in the literature alongside working memory and cognitive flexibility. Pooling 26 effect sizes from 16 studies, the analysis yielded a statistically significant effect of g = 0.45, with a 95 percent confidence interval of 0.16 to 0.74 and a p-value below 0.01. By conventional benchmarks that is a small-to-moderate effect. Yet the prediction interval stretched from −0.76 to 1.65, meaning that in a future study the true effect could plausibly be negligible or even negative. Heterogeneity was also substantial, with 72 percent of the variance attributable to within-study differences, and the GRADE assessment downgraded this outcome to low certainty because of serious risk of bias and serious inconsistency across the trials.

Internalizing symptoms, encompassing depression, anxiety, and somatisation, produced the largest pooled estimate: g = 0.97 from 22 effect sizes across 8 studies, statistically significant at p = 0.03 with a confidence interval of 0.16 to 1.80. But this result carries the heaviest caveats in the paper. The certainty of evidence was rated very low, downgraded for high risk of bias, serious inconsistency, and serious publication bias, with both the level 2 and level 3 multilevel Egger tests indicating significant funnel plot asymmetry. The prediction interval ran from −1.17 to 3.11, an extraordinarily wide range that underscores how unstable this estimate is. The authors state plainly that although the pooled effect is statistically significant, the effect cannot yet be reliably established.

The remaining outcomes did not clear the significance threshold. Cognitive flexibility, pooled across 18 effect sizes from 10 studies, produced g = 0.69 with a confidence interval of −0.01 to 1.40 and p = 0.054, hovering just above the conventional cutoff, with 90 percent of variance at the between-study level. Working memory yielded g = 0.28 across 19 effect sizes from 10 studies, p = 0.10. Externalizing symptoms, assessed in 17 effect sizes from 8 studies, produced g = 0.64 with a confidence interval spanning −0.69 to 2.00 and p = 0.29, alongside enormous heterogeneity. Sensitivity analyses revealed that the conclusions for cognitive flexibility and working memory were highly dependent on individual studies: removing certain trials flipped the working memory result to significance, while removing others rendered the cognitive flexibility estimate clearly non-significant, indicating poor robustness for both.

One of the more intriguing findings concerned exercise duration. In subgroup analyses for inhibitory control, the overall test across duration categories was statistically significant, and both short-duration and long-duration subgroups showed significant pooled effects, g = 0.58 and g = 0.83 respectively, while the moderate-duration subgroup did not, g = 0.13. Adjusted pairwise comparisons found a significant difference only between the short- and moderate-duration groups. The authors propose, drawing on prior literature, a possible non-linear pattern: brief sessions may acutely elevate monoamine neurotransmitters such as dopamine and noradrenaline, transiently boosting prefrontal excitability, while sustained long-term programs may upregulate brain-derived neurotrophic factor and promote plasticity in prefrontal-striatal circuits. Moderate durations may fall into a transitional zone where neither mechanism is fully engaged, or may simply contain highly heterogeneous studies that dilute the pooled effect. They stress that these explanations are hypotheses the meta-analysis cannot directly verify.

Meta-regression added another exploratory signal: at the study level, mean body mass index was negatively associated with the effect of exercise on inhibitory control, with a beta of −0.15, meaning studies with higher average participant BMI tended to show smaller cognitive benefits. The authors suggest that systemic inflammation associated with higher BMI might suppress BDNF secretion and limit exercise-induced neuroplasticity, but they caution that this is an aggregated, ecological association that cannot be extrapolated to individuals and may be confounded by other study characteristics. For internalizing symptoms, subgroup analyses found significant pooled effects within the long-duration and aerobic exercise strata, g = 1.35 and g = 1.41 respectively, though the overall moderation tests were not significant, so no firm conclusions about which exercise types work best can be drawn. More broadly, the authors suggest that exercise may act selectively on executive function subcomponents, and that externalizing problems, which are shaped by parenting, school environments, and basal ganglia dopamine pathways as well as cognition, may resist exercise interventions delivered alone.

The authors are candid about the limitations that temper their conclusions. Most of the 25 trials raised some concerns in the Cochrane RoB 2 assessment, with 18 rated as having some concerns and 6 at high risk of bias, primarily due to inadequately described randomization, unblinded outcome assessment with subjective scales, loss to follow-up, and missing trial registration. Subgroup samples were small, limiting statistical power, and the regression models did not account for medication use or baseline symptom severity. Diagnostic methods for ADHD varied across studies, participants were overwhelmingly boys, and only published English-language literature was included. The authors conclude that physical exercise may be positively associated with inhibitory control in children and adolescents with ADHD, but that intervention effects show high heterogeneity, wide confidence intervals, and limited stability, and they call for larger, lower-risk randomized trials to confirm which forms of exercise, at what doses, genuinely help.

Subject of Research: Effects of physical exercise on executive function and emotional and behavioral symptoms in children and adolescents with ADHD

Article Title: The effects of physical exercise on executive function, internalizing symptoms and externalizing symptoms in children and adolescents with attention-deficit/hyperactivity disorder: a systematic review and meta-analysis

Article References: Shi, J., Zheng, X., & Wang, W. (2026). The effects of physical exercise on executive function, internalizing symptoms and externalizing symptoms in children and adolescents with attention-deficit/hyperactivity disorder: a systematic review and meta-analysis. Frontiers in Psychology, 17, 1984575. https://doi.org/10.3389/fpsyg.2026.1984575

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1984575

Keywords: ADHD, physical exercise, executive function, inhibitory control, meta-analysis, children and adolescents, internalizing symptoms, externalizing symptoms, randomized controlled trials, GRADE evidence, working memory, cognitive flexibility