High-intensity interval training in cancer patients: a bayesian meta-analysis of effects on VO₂max, dose–response, and adherence

Abstract

Background

Cancer treatment impairs cardiorespiratory fitness (CRF), an independent prognostic indicator in cancer populations. High-intensity interval training (HIIT) improves maximal oxygen uptake (VO₂max), but existing meta-analyses have relied on frequentist approaches without probabilistic expression of clinical relevance and have analyzed dose–response and adherence in isolation.

Methods

Four databases were searched through 1 April 2026 for randomized controlled trials (RCTs) of HIIT versus non-exercise controls reporting VO₂max. Mean differences (MDs) were pooled using a Bayesian three-level random-effects model, with the posterior median, 95% credible interval (CrI), probability of direction (pd), Bayes factor (BF10), and the posterior probability of exceeding a prespecified threshold of 1.0 mL·kg⁻¹·min⁻¹ reported. Functional-form comparison was applied to 14 dose moderators, and certainty of evidence was rated using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach.

Results

Sixteen RCTs contributing 35 effect sizes were included. HIIT increased VO₂max by 2.61 mL·kg^⁻1·min^⁻1 versus non-exercise controls (95% CrI: 1.68 to 3.51), with pd = 1.00 and a 99.96% posterior probability of exceeding the threshold. The primary effect remained stable across all subgroup and sensitivity analyses. Dose–response curves showed minimal variation, and only the number of intervals reached pd = 0.95 among the 14 moderators. Mean adherence was 85.7%, with no prescription parameter showing a statistically significant association. Certainty of evidence was rated moderate.

Conclusions

HIIT reliably improves VO₂max in cancer patients, with near-certain posterior probability of exceeding clinical relevance. Dose–response and adherence analyses were exploratory and underpowered and do not support dose equivalence. Interval number was the only directional signal and may warrant further investigation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13102-026-01876-6.

Introduction

Impaired cardiorespiratory fitness (CRF) is among the most clinically meaningful functional consequences of cancer diagnosis and treatment in survivorship care. Cardiovascular deconditioning, declines in skeletal muscle mass and function, and chronic inflammatory burden induced by chemotherapy, radiotherapy, and surgical stress can reduce post-treatment maximal oxygen uptake (VO₂max) by 20% to 30% relative to age-matched healthy populations [1, 2]. Beyond its descriptive role, CRF is an independent prognostic indicator in cancer populations. Large prospective cohorts and dose–response meta-analyses have consistently shown an inverse association between CRF and cancer mortality. In the general population, each 1-MET increment corresponds to an approximately 7% reduction in cancer mortality risk [3], and in patients already diagnosed with cancer the association is stronger, with each unit increase in CRF linked to up to an 18% reduction in cancer-specific mortality risk [4]. Together, these findings position VO₂max not merely as an exercise-physiology endpoint but as a central variable linking treatment burden, rehabilitative intervention, and long-term survival.

Among the exercise modalities capable of improving CRF, high-intensity interval training (HIIT) has emerged over the past decade as one of the most actively investigated approaches in exercise oncology, owing to its efficiency in delivering a high-intensity training stimulus within short sessions. Multiple meta-analyses have shown that HIIT yields clinically meaningful improvements in VO₂max compared with usual care in cancer patients [5–7], and subgroup analyses indicate consistent direction of effect across breast cancer, lung cancer, preoperative prehabilitation, and post-treatment rehabilitation settings [8–10]. Current cancer survivorship guidelines uniformly list regular moderate-to-vigorous aerobic activity as a core recommendation [11–14]. However, the prescriptive frameworks they offer remain largely confined to weekly volume targets and do not provide evidence-based guidance on the finer prescription dimensions that are specific to HIIT, including interval structure, work-to-recovery ratio, and single-session dose configuration.

Although the overall efficacy of HIIT has been repeatedly confirmed, the existing evidence base contains systematic gaps at both the methodological and clinical-decision levels. Most published HIIT-cancer meta-analyses use frequentist random-effects models reporting pooled mean differences or standardized mean differences with 95% confidence intervals [5–7, 9, 10, 15]. Although confidence and prediction intervals convey clinically interpretable uncertainty, they do not directly express the probability that the effect exceeds a clinically relevant threshold. As a result, readers cannot tell whether this probability is 50%, 80%, or close to certainty. At the dose–response level, prior studies have generally relied on binary subgroup comparisons, such as dichotomies based on an 8-week threshold or treatment phase [6, 16], and have not incorporated continuous HIIT dose variables into a unified framework for functional dose–response modeling. As a result, the central clinical question of which prescription configuration is more effective remains unanswered. A further limitation is that adherence and dose have typically been analyzed in isolation [17, 18], leaving unclear whether the prescription flexibility implied by these analyses is feasible in real-world practice, which is precisely the information clinicians need when prescribing HIIT for patients at different treatment stages and with varying symptom burdens.

To address these gaps, the present study applied a Bayesian three-level random-effects model to evaluate the effect of HIIT on VO₂max, its dose–response, and its link to adherence in cancer patients. Mean difference was used as the effect size, with the native clinical unit of mL·kg⁻1·min⁻1 retained for direct interpretation of clinical relevance. Statistical significance was reframed as a probabilistic statement of clinical relevance through posterior probabilities, Bayes factors, and the posterior probability that the effect exceeds the minimal clinically important difference (MCID). Functional-form comparison and dose–response modeling were then applied to prescription dose modifiers, and prescription parameters were integrated with adherence within a single explanatory framework, aiming to provide directly applicable decision support for clinical practice in exercise oncology.

Methods

Registration and reporting

The protocol of this systematic review and meta-analysis was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO; registration number: CRD420261351910) before the start of the study. The conduct and reporting of the review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement, and the corresponding checklist is provided in Supplementary File 1 [19]. As the study was a secondary analysis of previously published data, no additional ethical approval was required.

Search strategy

Two reviewers independently searched four electronic databases (PubMed, Embase, the Cochrane Central Register of Controlled Trials, and Web of Science Core Collection), with searches conducted up to 1 April 2026. The search framework was constructed based on previously published meta-analyses in the field and refined through team discussion to finalize the keyword combinations [6]. The search terms were organized around three core concepts and combined using the Boolean operator AND. The first concept covered HIIT and included “high-intensity interval training”, “high-intensity intermittent exercise”, “sprint interval training”, “aerobic interval training”, and “Wingate”. The second concept covered cancer and included the Medical Subject Headings (MeSH) term “neoplasms” together with free-text terms for individual cancer types. The third concept covered CRF and included “VO₂max”, “VO₂peak”, “peak oxygen uptake”, “cardiorespiratory fitness”, and “cardiopulmonary exercise testing”. The search was limited to published randomized controlled trials (RCTs), with no language restriction. A total of 753 records were retrieved from the four databases. An additional 26 records were identified by hand-searching the reference lists of retrieved studies and relevant reviews to ensure comprehensive coverage. The full search strings are provided in Supplementary File 1.

Eligibility criteria

Eligibility criteria were specified according to the PICOS framework. The population (P) comprised adults (≥ 18 years) with a clinical diagnosis of cancer, including survivors, with no restriction on cancer type or treatment phase. The intervention (I) was HIIT. Drawing on exercise prescription guidelines and previous HIIT reviews, HIIT was operationally defined in this study as a regimen that alternates high-intensity work intervals with lower-intensity recovery intervals, in which the high-intensity bouts typically correspond to ≥ 85% of maximal heart rate, ≥ 80% of maximal oxygen uptake, or an equivalent intensity index [6, 20]. Comparators (C) were non-exercise controls, including usual care, waitlist, and attention controls. Low-intensity attention activities, such as stretching below the training threshold or simulated cycling at < 3 METs, were classified within this comparator category because they generally fall below the moderate-intensity aerobic training threshold [21, 22]. The outcome (O) was maximal or peak oxygen uptake (VO₂max or VO₂peak), expressed in mL·kg⁻1·min⁻1, with priority given to results obtained by direct cardiopulmonary exercise testing (CPET) [23]. Eligible study designs (S) were RCTs.

Studies were excluded if the regimen was not pure HIIT, for example, combined HIIT and resistance training; if VO₂ was reported only in absolute units (L/min) or estimated from submaximal testing in a way that could not be combined with mL·kg⁻¹·min⁻¹ data obtained by direct CPET; or if the study was non-randomized, a review, a conference abstract, or otherwise lacked the quantitative data required for meta-analysis. Body-mass-normalized VO₂ (mL·kg⁻¹·min⁻¹) was used because it is the clinically interpretable measure of cardiorespiratory fitness and the basis for the prespecified threshold, and absolute values (L/min) cannot be converted to it without individual body-mass data, which the reports in question did not provide.

Study selection and data extraction

Two reviewers independently performed title and abstract screening followed by full-text assessment, with disagreements resolved by a third reviewer. Data were independently extracted by two reviewers into a predefined Excel template. Extracted items included study identifiers (first author, year of publication, country), population characteristics (cancer type, stage, treatment phase, anticancer therapy, mean age, proportion of women, body mass index [BMI]), intervention characteristics (exercise modality, work intensity, work and recovery durations, number of intervals, session duration, weekly frequency, intervention duration in weeks, supervision, and exercise setting), outcome measurement modality and time points, and arm-level pre- and post-test means, standard deviations (SDs), and sample sizes for VO₂max or VO₂peak. For each trial, a metabolic-equivalent (MET) value was assigned from the 2024 Adult Compendium of Physical Activities [21] according to the exercise modality and the prescribed work intensity. Weekly MET-minutes were then calculated as this MET value multiplied by the session duration in minutes and by the weekly session frequency, and MET-minutes per session as the MET value multiplied by the session duration. Because the Compendium provides modality-specific values, the assigned MET reflected the modality of each trial (for example, treadmill versus cycle-ergometer protocols), and the value was based on the high-intensity work component of the HIIT protocol.

Risk of bias assessment

Risk of bias was independently assessed by two reviewers using version 2 of the Cochrane Risk of Bias tool for randomized trials (RoB 2), with disagreements resolved by a third reviewer [24]. Five domains were evaluated, comprising the randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result, together with an overall judgment. Detailed results are presented in Supplementary File 3.

Statistical analysis

The mean difference (MD) was used as the effect size, expressed in mL·kg⁻1·min⁻1. For each study arm, the pre-to-post change was computed as Δ = post-test mean − baseline mean. In the Bayesian model, arm-level change served as the dependent variable, and group was specified as an indicator distinguishing the HIIT and control arms. The group coefficient therefore represented the MD in change between HIIT and control [25]. The standard error (SE) of the change was calculated as:

  

where SD_pooled was the pooled SD of baseline and post-test SDs, and r was the pre–post correlation coefficient [25]. As none of the included trials reported this correlation, r was set to 0.5 in the primary analysis, following the conservative default recommended in the Cochrane Handbook [25]. Because the true correlation between repeated VO₂ measurements is generally higher than 0.5, this choice yields comparatively larger standard errors and therefore wider, more conservative credible intervals. When the post-test SD was not reported, the baseline SD was used as a conservative substitute, and this handling strategy, together with the assumed value of r, was incorporated into the sensitivity analyses.

The primary analysis was a Bayesian three-level random-effects model specified as:

  

This specification accounted for dependence among effect sizes arising from multiple arms or multiple time points within the same study [26, 27].

Weakly informative priors were applied. The intercept and regression coefficients used Normal(0, 5), and variance components used Half-Cauchy(0, 2) constrained to positive values [28]. All models were fitted in brms (v2.21.0) and Stan [29, 30], with four parallel chains and 4000 iterations per chain, of which the first 2000 served as warm-up, and with adapt_delta set to 0.95. Chain convergence was verified by trace plots and by R-hat < 1.01 for all parameters [31], and adequacy of model fit was confirmed by graphical posterior predictive checks (Supplementary Figure S4d).

Pooled effects were reported as the posterior median with 95% credible intervals (CrIs), together with the probability of direction (pd), the region of practical equivalence (ROPE), and Bayes factors (BF10) [32]. The ROPE was defined as ± 0.5 mL·kg⁻¹·min⁻¹, that is, half of the 1.0 mL·kg⁻¹·min⁻¹ clinical threshold. Practical equivalence was assessed from the proportion of the 95% highest density interval (HDI) falling inside this region, with an effect regarded as not practically equivalent to zero when no part of its HDI overlapped the ROPE [33]. Parameter-level BF10 was obtained via the Savage–Dickey density ratio, and model-level BF10 was obtained via bridge sampling [34]. The posterior probability that the pooled effect exceeded a prespecified clinically relevant threshold of 1.0 mL·kg⁻¹·min⁻¹ was also reported. Consistent with the prognostic importance of cardiorespiratory fitness in cancer described above [4], this value was adopted as a deliberately conservative lower bound, below the conventional 1-MET increment (≈3.5 mL·kg⁻¹·min⁻¹). This threshold was used solely for interpretation and was not treated as an established universal MCID for VO₂max or VO₂peak in cancer patients, for which no validated value currently exists. Heterogeneity was reported as τ at both levels with the corresponding I2 proportions [35].

Subgroup analyses were performed separately by cancer type, treatment phase, anticancer therapy, exercise modality, supervision, exercise setting, and overall risk of bias. For each variable, models with and without the interaction term were fitted and compared using BF10. Following common interpretation thresholds, BF10 > 3 was considered positive Bayesian evidence for a subgroup difference [36]. Subgroup-level pooled effects were obtained from posterior marginal predictions of the interaction model.

Meta-regression and dose–response analyses were conducted for 14 continuous moderators. These included nine HIIT prescription parameters (intervention duration in weeks, weekly frequency, session duration, work duration, recovery duration, number of intervals, intensity in METs, MET·min per session, and weekly MET·min), four population characteristics (age, proportion of women, BMI, and baseline VO₂max), and adherence rate (analyzed only in the 11 trials that reported adherence). Each moderator was analyzed in a separate univariate model. To limit overfitting given the small number of studies, the candidate functional forms were restricted to a linear model and two low-complexity thin-plate splines (k = 3 and k = 4) [37], and the form for each moderator was selected by leave-one-out cross-validation (elpd_loo) [38], which penalizes overfitting by evaluating out-of-sample predictive accuracy. The resulting curves were flat with wide credible intervals and are therefore interpreted descriptively rather than as evidence of nonlinearity.

Three sets of sensitivity analyses were undertaken. The first addressed prior sensitivity, in which the primary prior was replaced with a narrow prior, a wide prior, a Student-t prior, and an exponential prior on τ. The second addressed data-subset sensitivity, restricted in turn to studies excluding high-risk-of-bias trials, studies measuring VO₂ by direct CPET only, studies using only immediate post-intervention data, and studies using passive controls only. The third was a leave-one-out influence analysis, in which the primary model was refitted with each study removed once. The same leave-one-out procedure was also applied to the interval-number meta-regression to assess the robustness of that exploratory finding. Publication bias was evaluated through visual inspection of funnel plots, Egger regression, and the Begg rank correlation test [39, 40]. Given that some included studies contributed multiple effect sizes, the publication bias tests were interpreted as exploratory.

For the 11 studies that reported HIIT adherence rates, Spearman correlation tests were used to explore associations between prescription parameters and adherence, and adherence distributions were compared across treatment phase, cancer type, and exercise modality. Given the limited sample size, this analysis was framed as exploratory.

The certainty of evidence for the primary effect and for each subgroup was rated according to the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) working group guidelines, considering the five downgrading domains of risk of bias, inconsistency, indirectness, imprecision, and publication bias [41]. The specific reasons for downgrading in each row are listed in Supplementary Table S3. All analyses were performed in R 4.5.2, primarily using the brms, Stan, tidybayes, bayestestR, metafor, loo, and ggplot2 packages.

Results

Study selection

A total of 753 records were retrieved from the four databases (PubMed, n = 162; Embase, n = 174; Cochrane CENTRAL, n = 297; Web of Science, n = 120), and 26 additional records were identified by hand-searching the reference lists of included studies and relevant reviews, yielding 779 records in total. After removal of duplicates, 419 records were retained. Title and abstract screening reduced this set to 80 records assessed at full text, of which 64 were excluded because the intervention, comparator, outcome, design, or data did not meet the eligibility criteria. The final dataset comprised 16 RCTs contributing 35 effect sizes to the primary analysis. The full screening flow is presented in Fig. 1.

Fig. 1.

PRISMA 2020 flow diagram. The PRISMA 2020 flow diagram illustrates the study-selection process. A systematic search of four electronic databases (PubMed, Embase, the Cochrane Central Register of Controlled Trials, and Web of Science Core Collection) yielded 753 records (PubMed, n = 162; Embase, n = 174; Cochrane CENTRAL, n = 297; Web of Science, n = 120). An additional 26 records were identified by hand-searching the reference lists of included studies and relevant reviews, yielding a total of 779 records. After removal of duplicates, 419 records entered title-and-abstract screening, of which 339 were excluded for thematic irrelevance and 80 were retained for full-text assessment. At full-text assessment, 64 records were excluded because the intervention, comparator, outcome, study design, or data did not meet the eligibility criteria. Sixteen RCTs were included in the qualitative synthesis and quantitative meta-analysis

Characteristics of included studies

The 16 included trials were published between 2012 and 2025 and covered seven cancer types and three treatment phases. Breast cancer was the most frequent (7 trials), followed by prostate cancer (3 trials) and lung cancer (2 trials), with one trial each in testicular cancer, bladder cancer, colorectal cancer, and mixed cancer cohorts. The HIIT regimens had a median duration of 10 weeks (range 4–12 weeks), a median frequency of 3 sessions per week, and a median session duration of 30 min (range 14–40 min). Cycle ergometer was the most commonly used modality (8 trials), followed by treadmill (4 trials), mixed modalities (2 trials), with one trial each using other modalities and walking. Across the 16 trials, the outcome was assessed by direct CPET. One trial reported VO₂max and the remaining 15 reported VO₂peak. Detailed study characteristics are provided in Supplementary File 2.

Primary effect

The Bayesian three-level random-effects model showed that HIIT increased VO₂max by 2.61 mL·kg⁻1·min⁻1 relative to non-exercise controls (95% CrI: 1.68 to 3.51; Fig. 2). The pd of the pooled effect was 1.00, and the entire 95% HDI lay outside the ROPE of ± 0.5 mL·kg⁻1·min⁻1 (ROPE overlap = 0%). The model-level BF10 exceeded 2500. The posterior probability that the HIIT effect exceeded the prespecified clinically relevant threshold of 1.0 mL·kg⁻1·min⁻1 was 99.96%.

Fig. 2.

Forest plot of the effect of HIIT on VO₂max in cancer patients. The forest plot shows the MD and corresponding 95% CrI for each of the 16 included RCTs, expressed in mL·kg⁻1·min⁻1. Each point represents the effect estimate of a single study, the horizontal line indicates its 95% CrI, and the diamond represents the pooled estimate from the Bayesian three-level random-effects model. Studies are ordered from the smallest to the largest effect. The vertical dashed line denotes the line of no effect (MD = 0), with points to its right favoring HIIT over control. The pooled effect was MD = 2.61 mL·kg⁻1·min⁻1 (95% CrI: 1.68 to 3.51), with pd = 1.00 and a posterior probability of 99.96% that the effect exceeded the prespecified clinically relevant threshold of 1.0 mL·kg⁻1·min⁻1. The horizontal axis is labeled “Favors Control” and “Favors HIIT” to aid directional interpretation

Total heterogeneity was I² = 60.4% (53.7% between-study and 6.7% within-study), with τ_study = 1.45 and τ_es = 0.39 mL·kg⁻¹·min⁻¹.

Subgroup analyses

Across all seven categorical variables, the Bayes factors provided no evidence for between-subgroup differences (all BF10 < 1, Fig. 3), indicating that the pooled effect did not vary credibly by subgroup. The subgroup-level estimates are therefore exploratory and imprecise. Among the larger strata, the credible interval lay entirely above zero for breast cancer (MD = 3.10, 95% CrI 1.79 to 4.42, k = 7) and for both treatment-phase subgroups, during treatment (MD = 2.56, 95% CrI 0.86 to 4.23, k = 6) and after treatment (MD = 3.04, 95% CrI 1.77 to 4.24, k = 7). Estimates for the remaining strata, including the single-trial testicular-cancer subgroup, were based on one to three studies with wide credible intervals and are not interpreted further.

Fig. 3.

Combined forest plot of the subgroup analyses. The combined forest plot presents subgroup-analysis results across seven categorical variables: A cancer type, B treatment phase, C anticancer therapy, D exercise modality, E overall risk of bias, F supervision, and G exercise setting. In each panel, points represent the pooled MD at each subgroup level with 95% CrIs, expressed in mL·kg⁻1·min⁻1. The "Overall" row at the bottom of each panel represents the overall pooled effect (MD = 2.61, 95% CrI 1.68 to 3.51) for reference. For each categorical variable, subgroup differences were evaluated by comparing the BF10 of the models with and without the interaction term. All seven BF10 values fell below 1 (range 0.008 to 0.54), and those for cancer type, treatment phase, and anticancer therapy were below 0.1, indicating no credible between-subgroup differences

Meta-regression and dose–response analyses for moderators

Meta-regression results for the 14 continuous moderators are shown in Fig. 4, and the forest plot of regression coefficients (β) is provided in Supplementary Figure S3. Only the number of intervals reached the prespecified pd threshold of 0.95 (β = − 0.522, 95% CrI: − 1.090 to 0.113), with the coefficient pointing in a negative direction. The 95% CrI nonetheless crossed zero, indicating directional evidence rather than a credible interval excluding the null. Three further variables showed directional signals below this threshold: work interval duration (β = + 0.006, pd = 0.88, positive), BMI (β = + 0.152, pd = 0.86, positive), and intervention duration in weeks (β = − 0.066, pd = 0.84, negative). The remaining 10 variables had pd values below 0.80, including the conventional total dose metric of weekly MET·min (β = 0.001, pd = 0.68). Across the prescription range examined, the dose–response curves showed minimal variation.

Fig. 4.

Meta-regression and dose–response curves for 14 continuous moderators. The panels display the dose–response relationships between 14 continuous moderators and the effect on VO₂max (MD, mL·kg⁻1·min⁻1). Ten panels show HIIT prescription parameters and baseline fitness: A weekly MET·min, B MET·min per session, C intensity in METs, D baseline VO₂max, E intervention duration in weeks, F weekly frequency, G session duration, H work interval duration, I recovery interval duration, and J number of intervals. The remaining four panels show population characteristics and adherence: K mean age, L proportion of women, M BMI, and N adherence rate. In each panel, the solid line shows the best-fitting curve, selected from a linear model, a thin-plate spline with k = 3, and a thin-plate spline with k = 4 using elpd_loo. The shaded band indicates the 95% CrI, and points represent the observed effect sizes. The best-fitting model type is indicated in each panel title. Apart from the number of intervals (β = − 0.522, pd = 0.95), pd did not reach the prespecified threshold of 0.95 for the remaining 13 moderators, and the dose–response curves showed minimal variation across the prescription range examined

Sensitivity and influence analyses

The primary effect remained stable across all sensitivity analyses (summarized in Supplementary Figure S4a–c). Replacing the primary prior with a narrow prior (Normal(0,1)/Cauchy(0,1)), a wide prior (Normal(0,10)/Cauchy(0,5)), a Student-t prior, or an exponential prior on τ produced pooled effects ranging from 2.18 to 2.69 mL·kg⁻1·min⁻1, with all CrIs excluding zero (Supplementary Figure S4a). The data-subset analyses, which restricted the dataset in turn to trials excluding high-risk-of-bias studies, trials measuring VO₂ by direct CPET only, trials reporting immediate post-intervention data only, and trials using passive controls only, yielded pooled effects ranging from 2.60 to 2.69 mL·kg⁻1·min⁻1 (Supplementary Figure S4b). The leave-one-out influence analysis showed that the pooled effect remained within a narrow range when any single trial was removed (Supplementary Figure S4c). Varying the assumed pre-post correlation between r = 0.3 and r = 0.7 produced pooled effects of 2.58 to 2.62 mL·kg⁻¹·min⁻¹, all with 95% credible intervals excluding zero (Supplementary Table S2). A leave-one-out analysis of the interval-number meta-regression confirmed that its negative association was not driven by any single trial: across all 14 removals the coefficient remained negative (β from − 0.71 to − 0.44), and the probability of direction stayed at 0.92 or above in 13 of the 14 cases (Supplementary Table S4). The signal was weakest after removing one trial (Egegaard 2019, pd = 0.82) but remained negative in direction.

Publication bias

Visual inspection of the funnel plot did not indicate substantial asymmetry (Supplementary Figure S2). Neither Egger regression (z = − 1.39, p = 0.184) nor the Begg rank correlation test (τ = − 0.21, p = 0.271) reached statistical significance.

Risk of bias

The RoB 2 assessments of the 16 trials are presented in Supplementary Figure S1. In the deviations-from-intended-interventions domain (D2), 62% of trials were rated as some concerns, 31% as low risk, and 6% as high risk, reflecting the inherent difficulty of blinding participants and providers in exercise interventions. For the remaining four domains, the proportions of trials judged at low risk were 88% for D1 (randomization process), 69% for D3 (missing outcome data), 81% for D4 (measurement of the outcome), and 94% for D5 (selection of the reported result). The overall judgments were low risk in 6% (1 trial), some concerns in 88% (14 trials), and high risk in 6% (1 trial).

Feasibility and adherence

Eleven trials reported HIIT adherence rates. The mean adherence was 85.7% (SD ≈ 14%, range 55% to 100%), and 55% of the trials met the clinically acceptable threshold of ≥ 85%. Spearman correlations between prescription parameters and adherence are shown in Fig. 5. None of the six core prescription parameters (weekly MET·min, intervention duration in weeks, weekly frequency, session duration, work intensity, and work interval duration) showed a statistically significant association with adherence (all p ≥ 0.33). The strongest correlations were observed for work interval duration (ρ = + 0.37, p = 0.33) and weekly frequency (ρ = − 0.31, p = 0.35).

Fig. 5.

Correlation analysis between HIIT prescription parameters and adherence. The scatter plots show the relationships between six core HIIT prescription parameters and study-level adherence rate (%) across the 11 trials that reported adherence: A weekly MET·min, B intervention duration in weeks, C weekly frequency, D session duration, E work intensity (% of maximum), and F work interval duration. Each point corresponds to one included trial. The horizontal dashed line indicates the 85% clinically acceptable adherence threshold. The Spearman rank correlation coefficient (ρ) and the corresponding p value are annotated at the bottom of each panel. None of the six correlations reached statistical significance (all p ≥ 0.33). The strongest correlations were observed for work interval duration (ρ = + 0.37, p = 0.33) and weekly frequency (ρ = − 0.31, p = 0.35)

When stratified by clinical context (Supplementary Figure S5), mean adherence was 91.0% in the post-treatment phase (k = 5), 83.2% during treatment (k = 4), and 77.5% in preoperative prehabilitation (k = 2). By modality, mean adherence was 97.5% on the treadmill (k = 2) and 84.4% on the cycle ergometer (k = 7).

Certainty of evidence (GRADE)

The certainty of evidence for the primary effect was rated as moderate, downgraded by one level from the high starting point for randomized trials because of risk of bias. The subgroup and dose–response analyses were further downgraded for imprecision and rated low to moderate. The full ratings and downgrading rationales are summarized in Table 1.

Table 1.

Summary of pooled effects, key subgroups, and GRADE certainty assessment

Analysis	k	MD [95% CrI] (mL/kg/min)	Heterogeneity/key statistics	GRADE
Primary effect (HIIT vs control)	16	2.61 [1.68, 3.51]	I2 = 60.4%; pd = 1.00; BF10 > 2500; P(MD ≥ threshold) = 99.96%	moderate
By cancer type — Breast	7	3.10 [1.79, 4.42]	Largest evidence base; consistent positive effect	moderate
By cancer type — Testicular	1	3.47 [0.42, 6.38]	Single trial; wide CrI	low
By cancer type — others (combined)	8	1.12 to 3.13 (CrI crosses 0)	Prostate, lung, CRC, bladder, mixed; estimates imprecise	low
By treatment phase — Pre-surgery	3	1.63 [− 0.36, 3.61]	CrI marginally crosses zero; small evidence base	low
By treatment phase — Post-treatment	7	3.04 [1.77, 4.24]	Largest effect among phases; CrI excludes zero	moderate
By treatment phase — During treatment	6	2.56 [0.86, 4.23]	Effective during active treatment; CrI excludes zero	moderate
By risk of bias — across strata	16	2.62 to 2.97	Effect consistent across RoB strata; BF10 = 0.30 (no interaction)	moderate
Dose–response (14 candidate moderators)	11–16	—	Only number of intervals reached pd = 0.95 (β = − 0.522); no moderator had a CrI excluding zero	low
Sensitivity (4 priors + 4 data subsets)	14–16	2.18 to 2.69	All retained significance; result stable	—
Publication bias	16	—	Egger p = 0.184; Begg p = 0.271; funnel symmetric	—
Adherence (feasibility analysis)	11	—	Mean 85.7% (range 55–100%); no parameter strongly predicted adherence	low

Abbreviations: BF10 Bayes factor for the alternative hypothesis, CrI credible interval, CRC colorectal cancer, HIIT high-intensity interval training, k number of trials, LOO leave-one-out cross-validation, MD mean difference, pd probability of direction, RoB risk of bias. The clinically relevant threshold of 1.0 mL·kg⁻¹·min⁻¹ was prespecified for interpretation of clinical relevance and is not a cancer-specific minimal clinically important difference. GRADE certainty: high, moderate, low, or very low; downgrading domains are risk of bias, inconsistency, indirectness, imprecision, and publication bias. The detailed rationale for each GRADE rating is provided in Supplementary Table S3

Discussion

Principal findings

HIIT increased VO₂max by 2.61 mL·kg⁻1·min⁻1 relative to non-exercise controls, with pd = 1.00, a posterior probability of 99.96% that the effect exceeded the prespecified clinically relevant threshold of 1.0 mL·kg⁻1·min⁻1. No credible differences were detected between subgroups, and the primary effect remained stable across all prior specifications, data subsets, and leave-one-out analyses. Among the 14 continuous moderators examined, only the number of intervals reached the prespecified pd threshold of 0.95, with a negative direction (see Fig. 4 and Supplementary Figure S3). Regression coefficients for conventional total-dose metrics, including weekly MET·min, were close to zero, and the dose–response curves showed minimal variation across the prescription range examined. Across the 11 trials reporting adherence, none of the Spearman correlations between prescription parameters and adherence reached statistical significance.

Comparison with previous studies

The pooled effect of the present study aligns closely with several recent HIIT–cancer reviews. A lung-cancer-specific meta-analysis reported a pooled effect of 2.62 mL·kg⁻1·min⁻1 [9], a prehabilitation-focused review reported 2.76 [8], and a Bayesian network meta-analysis in breast cancer survivors estimated the marginal effect of HIIT versus usual care at 2.9 [42]. The present pooled estimate of 2.61 mL·kg⁻1·min⁻1 falls in the middle of this range. Earlier reviews based on smaller samples reported somewhat larger effects [5], whereas more recent reviews using broader inclusion criteria reported slightly smaller effects [6]. This modest dispersion around a common central value reflects methodological differences in inclusion criteria, conventions for effect-size calculation, and the handling of within-study correlation among multiple effect sizes, rather than instability in the underlying HIIT effect.

The physiological basis for HIIT-induced improvements in VO₂max has been characterized in detail in non-cancer populations and includes increases in maximal stroke volume and cardiac output, expansion of capillary density, enhancement of skeletal muscle oxidative enzyme activity, and upregulation of mitochondrial biogenesis. These adaptations are induced jointly through repeated cycles of brief high-intensity stress and recovery [43, 44]. In cancer populations, these adaptive pathways remain available in principle but are modified by the physiological burdens associated with treatment. Anthracycline-related cardiac dysfunction, chemotherapy-associated anemia, adverse skeletal-muscle and body-composition changes, autonomic dysregulation, and inflammation-related fatigue may constrain patients’ tolerance of and adaptation to HIIT, with the magnitude and direction of response varying with treatment exposure, baseline physiological reserve, and symptom burden [45, 46]. This context helps interpret the limited dose–response variation observed in the present analysis. In healthy adults and in overweight and obese populations, intervention duration in weeks, session duration, training intensity, and baseline fitness display identifiable dose gradients [47, 48], whereas in cancer populations such gradients appear compressed by the treatment context.

This dose–response signal should be interpreted with caution. The number of intervals was the only one of the 14 moderators to reach the prespecified probability of direction of 0.95, and even so its 95% credible interval crossed zero (− 1.09 to 0.11). A leave-one-out analysis showed that the negative direction was not driven by any single trial (Supplementary Table S4), but the interval still included the null, so the finding is hypothesis-generating rather than a demonstrated dose–response relationship. One tentative explanation, drawn from general HIIT training science, is that beyond a certain point accumulated fatigue could lower the intensity of later work intervals, so that additional intervals add training time without a proportional increase in effective high-intensity exposure [49, 50]. The included trials, however, reported prescribed rather than achieved intensity and did not document intensity or heart-rate attainment across individual intervals, so we found no direct evidence for this mechanism within the reviewed studies. It therefore remains a hypothesis to be tested in trials that record per-interval intensity and compare different interval counts at matched weekly volume.

With respect to comparative effectiveness, the relative advantage of HIIT over moderate-intensity continuous training has been inconsistent in prior evidence. Early reviews did not detect a significant difference between the two modalities [5, 6]. More recent meta-analyses have reported small-to-moderate advantages [51, 52], and a breast-cancer-specific Bayesian network analysis suggested that combined training (moderate-to-vigorous aerobic plus resistance) may rank above HIIT alone in the surface under the cumulative ranking curve (SUCRA) ordering [42]. This pattern indicates that whether HIIT is effective and whether HIIT is superior to other exercise modalities are two clinically distinct questions that cannot substitute for each other. The present analysis addressed the first question, comparing HIIT with non-exercise controls. The comparative-effectiveness question is better addressed through a dedicated network meta-analysis.

Clinical implications

The most direct clinical implication is that HIIT can serve as a routine option for CRF intervention in cancer rehabilitation and survivorship care. An increase of 2.61 mL·kg⁻¹·min⁻¹ corresponds to approximately 0.75 MET, exceeding the threshold for clinically perceptible improvement. Because the dose–response curves varied little across the prescription range examined, clinicians likely have some latitude to adapt HIIT to the patient's treatment phase, symptom burden, baseline fitness, and capacity for sustained participation. Across the 11 trials reporting adherence, no prescription parameter was significantly associated with adherence. However, with only 11 trials, the statistical power was low, so this absence of significant associations does not mean that no association exists, and it should not be interpreted as evidence that dose configuration is unrelated to adherence. On this basis, Table 2 summarizes preliminary, evidence-informed prescription considerations by domain, covering eligible populations, treatment timing, frequency and duration, session structure and intensity, exercise modality and setting, and feasibility, with the certainty of evidence indicated for each. Of the moderators examined, only the number of intervals showed a directional signal (pd = 0.95), and its credible interval still crossed zero. We therefore regard this as a hypothesis to be tested in dedicated dose-comparison trials rather than a basis for current prescription decisions.

Table 2.

Preliminary HIIT prescription considerations based on current evidence

Domain	Consideration	Confidence	Reasoning for confidence rating
Indication	HIIT is recommended to improve VO₂max in adult cancer survivors. The pooled effect of 2.61 mL·kg⁻1·min⁻1 (95% CrI: 1.68 to 3.51) exceeds a 1.0 mL·kg⁻1·min⁻1 clinically relevant threshold with 99.96% posterior probability	moderate	Supported by 16 RCTs with consistent direction of effect, large model-level Bayes factor (BF10 > 2500), and stable estimates across all subgroup and sensitivity analyses. GRADE certainty was rated moderate after downgrading once for risk of bias (the unblindable nature of exercise interventions)
Eligible populations	Strongest evidence in breast cancer (k = 7); testicular cancer showed a positive estimate but is based on a single trial. Effect direction was also positive in prostate, lung, colorectal, bladder, and mixed cancer cohorts, with limited evidence per subgroup	moderate	Pooled effect is directionally consistent across cancer types (BF10 < 0.01 for interaction). Most non-breast subgroups had k ≤ 3 with CrIs crossing zero; recommendations beyond breast cancer should be regarded as supported by the overall pooled effect rather than by subgroup-specific evidence
Treatment timing	Effective during active treatment (k = 6, MD = 2.56 mL·kg⁻1·min⁻1, 95% CrI: 0.86 to 4.23) and after treatment (k = 7, MD = 3.04, 95% CrI: 1.77 to 4.24). Pre-surgical prehabilitation evidence is more limited (k = 3, MD = 1.63, 95% CrI: − 0.36 to 3.61)	moderate	During-treatment and post-treatment subgroups both showed pooled effects with CrIs excluding zero. The pre-surgery estimate is positive but its CrI marginally crosses zero, reflecting the small evidence base (k = 3) rather than absence of effect
Frequency × duration	2–3 sessions per week for 8–12 weeks. No clear additional VO₂max gain was detected at higher frequencies or longer programs within the studied range, although wide CrIs limit precise dose recommendations	moderate	Meta-regression on weekly frequency and intervention duration showed pd < 0.95, with regression coefficients close to zero. Within-range flexibility is supported, but extrapolation beyond the prescription range examined (4–12 weeks; 2–5 sessions per week) is not
Session structure and intensity	20–40 min per session including warm-up and cool-down, with work intervals at ≥ 85% HRmax or ≥ 80% VO₂max. The number of intervals showed a directional negative association with effect (β = − 0.522, pd = 0.95, with a credible interval that crossed zero) and may warrant focused evaluation in future dose-comparison trials	moderate	Effect was consistent across the full intensity range studied (no significant interaction). Number of intervals was the only moderator reaching the prespecified pd threshold of 0.95; this exploratory finding should be confirmed in dose-comparison trials randomizing patients to different interval counts at constant weekly volume
Modality and setting	Cycle ergometer is the most extensively studied option (k = 7, mean adherence 84.4%); treadmill protocols showed higher mean adherence (97.5%) but in only two trials. Both supervised and home-based settings were used; modality and setting showed no statistical interaction with the effect	moderate	Subgroup analysis by modality and setting showed no statistical interaction (BF10 < 1). Modality choice can therefore be guided by patient preference, available equipment, and clinical context
Feasibility	Mean adherence was 85.7% across 11 trials (range 55–100%). HIIT is broadly tolerated when adequately supervised and adapted to the patient's symptom burden, with adherence highest in post-treatment settings (91.0%) and lower in pre-surgical prehabilitation (77.5%)	low	Based on only 11 trials reporting adherence, with non-standardized definitions across studies. No prescription parameter showed a statistically significant association with adherence, but exploratory power was limited by the small number of trials
Caution/evidence gaps	Limited evidence in hematologic malignancies, advanced or end-stage disease, and severe treatment-related cardiotoxicity. Standard pre-exercise screening and individualized intensity progression are recommended	expert consensus	No included RCTs systematically addressed these populations. Recommendation reflects established exercise oncology safety guidelines (ACSM, COSA) rather than direct evidence from this meta-analysis

Abbreviations: ACSM American College of Sports Medicine, BF10 Bayes factor, COSA Clinical Oncology Society of Australia, CrI credible interval, HIIT high-intensity interval training, HRmax maximum heart rate, k number of trials, MD mean difference, pd probability of direction, RCT randomized controlled trial, VO₂max maximal oxygen uptake. The 1.0 mL·kg⁻¹·min⁻¹ clinically relevant threshold is a pragmatic interpretive anchor based on associations between cardiorespiratory fitness and mortality and is not a cancer-specific minimal clinically important difference. GRADE certainty: high, moderate, low, or very low. These considerations are preliminary and evidence-informed rather than definitive prescriptions. They derive primarily from the pooled comparison of HIIT versus non-exercise controls. The dose–response, subgroup, and adherence analyses that inform the specific parameters were exploratory and underpowered, moderate between-study heterogeneity was present, and no head-to-head comparison with other exercise modalities was performed

Strengths

To our knowledge, few prior meta-analyses on HIIT, cancer, and VO₂max have combined a Bayesian three-level model with simultaneous functional-form comparison across 14 dose moderators and reported the posterior probability of exceeding a prespecified clinically relevant threshold. This framework complements, rather than replaces, existing frequentist syntheses by adding a direct probabilistic interpretation.

Limitations

Several limitations warrant cautious interpretation. The included sample of 16 trials supports a stable estimate of the primary effect, but several subgroups contained few trials and wide credible intervals, and the corresponding conclusions should be regarded as exploratory. The primary analysis also showed moderate between-study heterogeneity (I² = 60.4%) that was not explained by any measured subgroup or moderator, so the pooled effect represents an average across diverse populations and protocols, and the guidance in Table 2 should be regarded as general rather than setting-specific. In particular, the testicular-cancer subgroup was based on a single trial, and its positive estimate is driven entirely by that study and should not be generalized. Although the funnel plot, Egger regression, and Begg test did not indicate asymmetry, these formal tests have low power when the number of trials is small and effect sizes are dependent, so publication bias cannot be excluded. The outcome combined VO₂max and VO₂peak. Although both were obtained by direct cardiopulmonary exercise testing and are closely related, they are not identical constructs, and their pooling may have introduced minor measurement heterogeneity.

The moderator analyses carry additional limitations. All moderators were analyzed at the study level using the prescribed dose rather than the dose actually completed by individual patients, so the meta-regression associations are ecological and may not hold at the individual-patient level. The reporting of heart-rate target attainment, ratings of perceived exertion, and the number of intervals actually completed also varied across trials, indicating uncertain implementation fidelity that may have added heterogeneity and diluted any true dose–response signal. In addition, with only 16 trials and 35 effect sizes, the dose-moderator analyses had limited power and were susceptible to overfitting, so the functional-form results are exploratory and hypothesis-generating, and the negative association with the number of intervals, whose credible interval crossed zero, is a hypothesis for future dose-comparison trials rather than a basis for prescription. Finally, the threshold of 1.0 mL·kg⁻¹·min⁻¹ used for interpretation is a pragmatic reference rather than a formally validated cancer-specific MCID, and validation in future cancer-specific studies is needed.

Conclusion

HIIT improves VO₂max in cancer patients, with a pooled effect that exceeds the prespecified clinically relevant threshold with near-certain posterior probability, although moderate between-study heterogeneity indicates that the magnitude of benefit may vary across settings. Within the dose range examined, no prescription parameter was credibly associated with either effect size or adherence, although these dose and adherence analyses were exploratory and underpowered. The number of intervals was the only moderator to show a directional signal and may warrant focused evaluation in future dose-comparison trials. If these findings are confirmed, HIIT could serve as a routine prescription option in cancer rehabilitation and survivorship care, adapted to the patient's treatment phase, symptom status, and capacity for sustained participation.

Authors’ contributions

All authors had full access to all the data in this systematic review and take responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Jianbo Zhang; Jingjing Tao. Acquisition, analysis, or interpretation of data: all authors. Drafting of the manuscript: all authors. Critical revision of the manuscript for important intellectual content: Jianbo Zhang. Statistical analysis: Jianbo Zhang, Jingjing Tao. Administrative, technical, or material support: Jianbo Zhang. Study supervision: Jianbo Zhang. All authors read and approved the final manuscript.

Funding

Not applicable.

Data availability

The datasets used and/or analyzed in the current study can be obtained from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Human Ethics and Consent to Participate declarations: not applicable. This study is a secondary analysis of previously published data and did not involve any new collection of human data. Therefore, ethics approval and informed consent were not required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.