Introduction

Stroke is the leading cause of disability and the third-leading cause of death worldwide1 accounting for over 6.5 million deaths and 143 million disability-adjusted life-years lost in 2019.2 The burden of stroke-related death and disability can be attributed to a rise in the prevalence of modifiable cardiovascular disease risk factors,3 low cardiorespiratory fitness (V̇O2peak),4–6 and lack of physical activity7 and exercise participation among stroke survivors.8 These factors contribute to a physiological environment rich in oxidative stress, inflammation and vascular dysfunction, which are associated with vascular cognitive impairment and stroke recurrence.9

Aerobic exercise (AE) training, commonly prescribed as moderate-intensity continuous training (MICT),10 11 is recommended for secondary stroke prevention and cardiovascular risk factor management. This prescription is individualised, but typically involves targeted exercise at 40%–60% of heart rate reserve (HRR) for 20–60 min, 3–5 times per week.10 11 Previous meta-analyses, primarily comprised MICT interventions, have reported significant benefits of AE training for improving V̇O2peak,12 mobility (6 min Walk Test distance (6MWT), 10 m Usual Gait Speed),12 and other cardiovascular risk factors such as systolic blood pressure (SBP),13–15 when compared with usual care or non-exercise control interventions. As such, MICT remains the primarily recommended exercise intervention for stroke.11 However, mounting evidence suggests that novel interventions, such as high-intensity interval training (HIIT), may stimulate distinct physiological mechanisms, leading to superior improvements in cardiovascular health outcomes and mobility among individuals poststroke.16–18

A limitation of previous meta-analyses is that prescriptions of AE interventions are often overlapping and combined to facilitate one direct pairwise comparison between exercise and control.12–15 18 However, in the context of exercise, there are many competing AE interventions that are likely to lead to differing treatment effects.16 Thus, a single effect estimate derived from pooled exercise interventions may lack precision and lead to inadequate dosing of AE after stroke.19 Network meta-analyses (NMA) can bridge this knowledge-implementation gap by examining the comparative effectiveness of competing AE interventions, developing precise effect estimates, and establishing a rank order of treatments from least to most effective.20 This will help healthcare professionals and community exercise specialists effectively determine which AE intervention provides the greatest treatment effects to guide AE prescription for individuals poststroke.

To date, no large-scale synthesis has examined the comparative effectiveness and superiority of different AE interventions for improving cardiovascular health or mobility outcomes following stroke. The purpose of this systematic review and NMA was to establish the effects and superiority of different AE interventions for improving V̇O2peak (primary outcome) and SBP (secondary outcome) as a key cardiovascular risk factor,3 as well as mobility outcomes (6MWT and 10 m Usual Gait Speed, tertiary outcomes) in individuals poststroke. Given the anticipated intensity-dependent response,16 it is hypothesised that higher-intensity exercise interventions, whether performed in intervals or continuously, will lead to superior health outcomes after stroke.

MethodologyStudy design

This systematic review with Bayesian NMA was reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses for NMA21 and was registered with PROSPERO (Protocol #303297). The established protocol plan was followed without significant deviations.

Search strategy

An electronic literature search was conducted in eight databases, MEDLINE, EMBASE, Web of Science, CINAHL, CENTRAL, SPORTDiscus, PsycINFO and AMED Allied and Complementary Medicine, from journal inception to 27 February 2023, with restriction to articles published in English. The search strategy was developed, piloted and refined in collaboration with a research librarian and used keywords related to stroke, AE and randomised controlled trials (RCTs). Grey literature, trial registries (eg, https://clinicaltrials.gov/), reference lists of included articles and relevant systematic reviews were handsearched to identify any additional potentially relevant literature. All search results were imported into Covidence software for study selection procedures. The full search strategy is provided in online supplemental appendix 1.0.

Inclusion and exclusion criteria

Studies were included in this review if they were a published RCT that examined the effect of any AE intervention on cardiorespiratory fitness (V̇O2peak, primary outcome), SBP as a key cardiovascular risk factor (SBP, secondary outcome), other modifiable cardiovascular risk factors (eg, diastolic blood pressure (DBP), fasting glucose, fasting insulin, glycosylated haemoglobin, total cholesterol, low-density lipoprotein (LDL) cholesterol, high-density lipoprotein (HDL) cholesterol; secondary outcomes), anthropometric measurements (body mass index, hip-waist ratio, waist circumference; secondary outcomes) or mobility outcomes (eg, 6MWT distance and fast and usual 10-metre gait speed, tertiary outcomes) in individuals with stroke. Several control interventions were eligible for inclusion in the network of evidence including no exercise treatment (eg, no prescribed intervention, relaxation exercises, passive treatment including passive range of motion or mobilisations), usual care (eg, usual stroke rehabilitation typically includes gait training, active range of motion, stretching, light resistance training, balance or functional training), and AE interventions featuring different intensity parameters were compared (eg, low vs moderate intensity AE).

Participants were adults over 18 years old, with no restrictions on the type (eg, ischaemic, haemorrhagic, transient ischaemic attack) or timing poststroke. Studies were excluded from this review if they (1) were editorials, theses, poster or conference abstracts or presentations, or opinion pieces; (2) did not report exercise prescription parameters according to the FITT principle: frequency, intensity, time, and type of AE; (3) included robot-assisted, exoskeleton, exergame or virtual rehabilitation technologies as either an intervention or control; (4) examined outcomes not related to V̇O2peak, mobility or cardiovascular risk factors in individuals with stroke and (5) were preclinical animal studies.

Study selection procedure

Two independent reviewers (reviewer 1: KM (all studies), reviewer 2: one of LR, EW and KSN (divided equally) independently conducted title and abstract screening and full-text evaluation. All reviewers (KM, LR, EW and KSN) first piloted the screening template by independently screening a predetermined set of abstracts (n=50) and full-text articles (n=5). After piloting the screening template for abstracts and full-text screening, study reviewers deliberated, compared results and resolved any discrepancies to ensure the screening criteria and data extraction sheets were clear and consistently applied. A third reviewer (AT) was consulted as needed to resolve any conflicts.

Data extraction procedures

Two independent authors (reviewer 1: KM and reviewer 2: LR or EW) conducted data extraction. Extracted data included study and participant characteristics, AE intervention parameters and control group details, and outcome measurements. A comprehensive list of extracted data can be found in online supplemental appendix 1.1. In the event of disagreements, they were resolved by consensus between authors and a third reviewer (AT) in case of conflict. Pilot testing of the data extraction sheets was performed with a small number of articles (n=5) to ensure consistency between reviewers.

Exercise prescriptions of interest

AE was defined as the specific set of interventions of direct interest for the present NMA and is referred to as the decision set.22 Each intervention was categorised using the FITT principle: frequency, intensity, time and type of AE. Per the AEROBICS 2019 guidelines11 and the American College of Sports Medicine,23 AE intensity was based on the reported per cent heart rate max (HRmax) or HRR or ratings of perceived exertion (RPE) and defined as low (<54% HRmax, <40% HRR, <11/20 RPE), moderate (55%–69% HRmax, 40%–59% HRR, 12–15/20 RPE) or high (>70% HRmax, >60% HRR, >15/20 RPE) and included in the decision set. HIIT interventions were also included in the decision set and adapted using previous definitions.16 24 A decision-making algorithm was developed by KM and AT to assist with group allocation in the decision set. Briefly, if the exercise programme had a clearly described intensity level, it was grouped as defined (eg, moderate, 40%–59% HRR). If an exercise programme had overlapping intensity levels with a clear progression plan to achieve a given intensity (eg, 12 weeks of 40%–80% HRR, with 10% increase in %HRR every 4 weeks), it was grouped as the highest target intensity. An exercise programme with overlapping intensity levels without a clearly described progression plan was grouped conservatively as the lowest defined target intensity. Lastly, if an exercise programme had overlapping or unclear intensity levels but included a patient-reported intensity target value (eg, RPE), the participant-reported target value guided the exercise grouping for the present analysis. The decision-making algorithm was also used to assist with group allocation for AE control groups. The decision-making algorithm was implemented by KM and verified by consensus with LR and with consultation with AT in the event of disagreement.

Data analysis

Where applicable, descriptive data extracted was analysed descriptively, including mean±SD, mode and median (IQR). Data were organised and summarised in tables and figures and narratively described in the text.

Effect measures and data transformations

The number of participants, mean value and the SD of the outcome measurements in each group were used in the meta-analysis using mean differences. All extracted outcome data were converted into means and SDs. SDs were converted from the reported SE or CI.25 If SDs of change scores were not reported, they were imputed using a correlation coefficient derived from similar studies.26 Means and SDs were estimated following procedures outlined by Wan et al if medians, IQRs or ranges were reported.25 If METs were reported as an outcome of V̇O2peak, they were converted to stroke adjusted V̇O2peak since 1 MET=3.0 mL/kg/min in individuals poststroke.27

Pairwise meta-analysis

Pairwise meta-analyses were performed using a random-effects model, given that studies differed methodologically and clinically, resulting in between-study heterogeneity.28 Between-study heterogeneity of the treatment effects within each treatment comparison was assessed using the I2 statistic29 and its 95% CI and the magnitude of the between-study variance (τ2) and its 95% CI, as estimated using the restricted maximum likelihood estimator and the Q-profile approach.30 31

Network meta-analysis

A Bayesian random-effects NMA was conducted using Markov Chain Monte Carlo simulations. This analysis included outcomes from a network of at least 10 studies to ensure network connectedness and allow for meaningful comparisons. A common within-network between-study variance (τ2) was assumed since no clinically important differences in the heterogeneity were expected across treatment comparisons. Vague informative priors were assumed for all Bayesian NMA model parameters. The models were run for 100 000 iterations after discarding the first 10 000 iterations and thinning of 10. To evaluate the convergence of the model, 3 chains were used, and models were assessed using the trace plots of leverage vs Bayesian deviance residual, Gelman-Rubin and Geweke convergence diagnostics.32 Convergence was assumed using a combination of the Potential Scale Reduction Factor for each parameter (≤1.05), and if z-scores from the Geweke diagnostics are reported as larger than 1.96 or smaller than −1.96.33 For each NMA, the assumptions of consistency and transitivity were assessed a priori.34 35

To examine consistency at the global level, the deviance information criterion (DIC) was compared between an inconsistency model and a model in which consistency was assumed. The node-splitting method was used to examine local inconsistency between direct, indirect and network estimates, and statistical inconsistency was considered when p≤0.05.36 If significant inconsistency was detected, data were checked for potential abstraction errors, and in the absence of errors, significant inconsistency was explored by conducting meta-regression analysis using the potential effect modifiers. The plausibility of intransitivity was assessed via visual observation of scatterplots and box and whisker plots of the distribution of effect modifiers across studies. Effect modifiers included the proportion of females in each study arm, mean age and exercise intervention volume, which was calculated as the product of time (minutes per week), intensity (eg, 80%) and intervention duration (weeks).37

Network meta-regression was conducted for the primary outcome (V̇O2peak), considering potential treatment effect modifiers. The fit of the meta-regression model was carried out using a plot of the leverage values and the corresponding effective number of parameters (pD), total residual deviance (Dres) and the DIC. A larger DIC suggests a worse model fit, and a Dres value close to the total number of arms included in the analysis suggests a good fit. A combination of the pD, Dres and DIC values was used to help determine model choice when considering competing models. A difference in DIC≤3 suggests no significant difference between the competing models.38 This approach was used to compare the fixed versus random effects, consistency versus inconsistency and meta-regression versus random effect models.

Interventions of each outcome were ranked using the surface under the cumulative ranking curve (SUCRA).39 40 SUCRA values range from 0 to 100%, with higher SUCRA values indicating the higher likelihood that a treatment is in the top rank or one of the top ranks.41 Network diagrams were created to represent the available evidence in a figure. Each node represents the different interventions and comparators. The size of each node and the weight of the lines visually represent the proportional number of studies comparing two connected nodes within the network. Network estimates were reported in league tables, including posterior mean differences and 95% credible intervals (CrI). Forest plots were generated to present network estimates compared with usual care.

All statistical analyses were conducted within R statistical software (V.4.2.2). Random effects pairwise meta-analyses were conducted using the meta package,42 Bayesian NMAs were conducted using the BUGSnet package43 and the node-splitting method was conducted using the gemtc package.36

Sensitivity analyses

Comparison-adjusted funnel plots were examined for extreme and small study effects and potentially outlying studies for the primary outcome (V̇O2peak). When potential outliers were identified, they were excluded in a sensitivity analysis to assess the robustness of the results.

Risk of bias

Study quality and risk of bias of each study and outcome were independently assessed by two independent authors (reviewer 1: KM and reviewer 2: LR or EW) using the Cochrane Risk of Bias 2 assessment.44 Briefly, this assessment included rating the risk of bias (low, high or some concerns) of the following domains: random sequence generation; allocation concealment; blinding of participants and personnel; blinding of outcome assessment; incomplete outcome data; and selective reporting bias.44 Disagreement was resolved by consensus or with a third reviewer (AT). Risk of Bias 2 summary plots were generated using robvis.45

Grading the quality of evidence

One reviewer (KM) evaluated the quality of evidence by using the Grading of Recommendations, Assessment, Development and Evaluation (GRADE) for NMA using the direct, indirect and network estimates from the node splitting analysis.46–48 The certainty for each comparison and outcome was rated as high, moderate, low or very low, based on considerations of risk of bias, inconsistency, indirectness, publication bias, intransitivity,49 incoherence50 and imprecision.51 Judgements of imprecision were made using a minimally contextualised approach, with a null effect as the threshold of importance.52 For the primary outcome, treatments were then classified in groups from the most to the least effective using the minimally contextualised framework,52 which focuses on the treatment effect estimates and the certainty of the evidence. The GRADE assessment was then discussed with the authors to ensure consensus on the interpretation of findings.

Equity, diversity and inclusion statement

Our study included all identified RCTs of AE training after stroke, with no exclusion criteria on sex, gender, race, ethnicity and socioeconomic level. Our research team consisted of individuals working in three provinces across Canada and included four women and five men from different ethnicities, disciplines (rehabilitation, physiology, medical research, exercise science and population health) and stages across academic careers.

ResultsStudy selection

The flow of the systematic review is presented in figure 1. The database search yielded 7463 records. The examination of titles and abstracts resulted in the retrieval of 160 full-text records. Following full-text evaluation, 46 studies were included in the review (V̇O2peak NMA (n=28 studies, primary outcome), SBP NMA (n=11 studies, secondary outcome), 6MWT NMA (n=28 studies, tertiary outcome), 10 m Usual Gait Speed NMA (n=18 studies, tertiary outcome)). A complete list of studies, demographic information and post-intervention outcomes are reported in online supplemental appendix 2.0. NMAs of other outcomes including DBP (n=11 studies), resting heart rate (n=10 studies) and fast gait speed (n=18 studies) are reported in online supplemental appendix 3.0. Few studies reported on vascular risk factors which restricted meaningful NMAs, including body mass index (n=7 studies), LDL and HDL cholesterol (n=5 studies), fasting insulin (n=3 studies) and glucose (n=3 studies), glycosylated haemoglobin (n=3 studies), total cholesterol (n=5 studies) and waist circumference (n=2 studies).

Figure 1

Flow chart of studies identified and included in the systematic review and network meta-analysis. 6MWT, 6 min Walk Test ; FITT, frequency, intensity, time and type; SBP, systolic blood pressure.

Network geometry

Figure 2 displays the network plot detailing the treatment comparisons for V̇O2peak (figure 2A), SBP (figure 2B), 6MWT (figure 2C) and 10 m Usual Gait Speed (figure 2D). The V̇O2peak NMA (28 studies, n=1298 participants) included 6 interventions, 15 total pairwise comparisons and 10 direct pairwise comparisons (figure 2A). The SBP NMA (11 studies, n=648 participants, figure 2B) included 5 interventions, 10 total pairwise comparisons and 6 direct comparisons. The 6MWT NMA included 28 studies, n=1494 participants, 6 interventions, 15 total pairwise comparisons and 7 direct pairwise comparisons (figure 2C). The 10 m Usual Gait Speed NMA (18 studies, n=775 participants) included 6 interventions, 15 total pairwise comparisons and 7 direct pairwise comparisons (figure 2D).

Figure 2

Network geometry for (A) peak oxygen uptake (V̇O2peak), (B) systolic blood pressure (SBP), (C) 6 min walk test (6MWT) and (D) 10 m Usual Gait Speed. Each node indicates an aerobic exercise intervention and is weighted according to the number of participants who received the intervention. Each line connecting the nodes is weighted according to the number of studies which contain pairwise comparisons. HICT, high-intensity continuous training; HIIT, high-intensity interval training; Low, low-intensity continuous training; MICT, moderate-intensity continuous training.

Study characteristics

Characteristics of included studies are detailed in online supplemental appendix 2.0. A total of 2517 participants were included in this review. The common aerobic intervention groups included MICT (n=22 arms),17 53–73 high-intensity continuous training (HICT) (n=23 arms),61 68 74–94 HIIT (n=8 arms)17 54 60 65 67 95–97 and low (n=5 arms).63 69 74 80 88 The most common comparator was usual care (n=31 arms)53 55 57–59 62 64 65 70–73 75 78 79 81–87 89–93 95–98 which commonly included components of stroke rehabilitation such as gait retraining, active or passive range of motion, stretching, light resistance training, balance, functional training or education on physical activity, lifestyle and medication modifications and no exercise (n=4 arms).56 66 76 94 Briefly, sample size ranged from 5 to 105 participants (mean age range, 48–76 years; one study did not report the mean age of participants76), and study duration ranged from 3 to 24 weeks. The majority of studies were conducted in either North America (n=14 USA, n=3 Canada, n=1 Jamaica) and Europe (n=4 Denmark, n=3 Germany, n=2 Norway, n=2 Sweden, n=1 Belgium, n=1 Italy, n=1 Ireland, n=1 Hungary), followed by Asia (n=3 Taiwan, n=2 China, n=2 Israel), Oceania (n=4 in Australia, n=1 in New Zealand) and South America (n=1 Brazil). The proportion of females ranged from 0% to 65%, and three studies did not report sex distributions.64 83 85 Most studies were conducted with participants at least 6 months poststroke (n=41 studies, 87%), followed by the participants between 7 days and 3 months poststroke (6 studies, 13%).

Risk of bias

For the primary outcome, most studies were rated as having an unclear (n=15, 54%) or a high risk of bias (n=12, 43%). One study was rated as a low risk of bias (n=1, 3%). Missing outcome data due to participant dropouts and selection of the reported result due to unblinded outcome assessments were the most frequently identified sources of high risk of bias. A summary of the risk of bias assessment is presented in online supplemental appendix 2.1.

Network estimates of all AE prescriptions

The posterior mean difference and 95% CrI from all network comparisons are displayed in table 1 for V̇O2peak (table 1A), SBP (table 1B), 6MWT (table 1C) and 10 m Usual Gait Speed (table 1D). For the primary outcome, treatments were classified from the most to the least effective using the minimally contextualised framework52 in table 2. A GRADE table for each outcome and comparison is outlined in online supplemental appendix 2.2.

Table 1

League table for all mean differences and 95% credible interval (CrI) by outcome and aerobic exercise intervention

Table 2

Final classification of all aerobic exercise interventions, based on network meta-analysis of interventions for aerobic exercise and cardiorespiratory fitness in people poststroke

Cardiorespiratory fitness

The effect of each AE prescription compared with usual care on V̇O2peak is displayed in figure 3A. The posterior mean difference for all interventions ranged from 0.04 to 3.2 mL/kg/min. The largest network estimates were observed after HIIT (3.2 mL/kg/min, 95% CrI (0.7 to 5.6), moderate certainty), followed by HICT (2.4 mL/kg/min, 95% CrI (0.3 to 4.5), moderate certainty) and MICT (2.3 mL/kg/min, 95% CrI (0.3 to 4.4), moderate certainty). There were no significant differences between HIIT, HICT and MICT pairwise comparisons. Compared with usual care, the effect of low (0.08 mL/kg/min, 95% CrI (−2.0 to 1.9), low certainty) and no exercise (0.2 mL/kg/min, 95% CrI (−2.3 to to 2.8), low certainty) were the only non-significant comparisons.

Figure 3

Forest plots of the posterior mean difference and 95% credible intervals of each aerobic exercise intervention compared with usual care for outcomes (A) peak oxygen uptake (V̇O2peak), (B) systolic blood pressure, (C) 6 min Walk Test and (D) 10 m Usual Gait Speed. HICT, high-intensity continuous training; HIIT, high-intensity interval training; Low, low-intensity continuous training; MICT, moderate-intensity continuous training.

Systolic blood pressure

The effect of each AE intervention compared with usual care on SBP is displayed in figure 3B. The posterior mean difference for all prescriptions ranged from 0.2 to –7.1 mm Hg. The largest network estimates were observed after HIIT (−7.1 mm Hg, 95% CrI (−18.6 to 3.9), low certainty), followed by HICT (−6.3 mm Hg, 95% CrI (−19.2 to 3.0), low certainty), MICT (−2.8 mm Hg, 95% CrI (−10.5 to 5.9), low certainty) and no exercise (2.2 mm Hg, 95% CrI (−7.1 to 11.6), low certainty). There were no significant differences observed for any pairwise comparison.

6 min Walk Test distance

The effect of each AE intervention compared with usual care on 6MWT distance is displayed in figure 3C. The posterior mean difference for all prescriptions ranged from −2.3 to 41.9 m, with the largest significant network estimate following HICT (39.4 m, 95% CrI (16.1 to 66.5), low certainty). Compared with usual care, the effect of HIIT (41.9 m, 95% CrI (−12.1 to 94.9), low certainty), MICT (24.1 m, 95% CrI (−1.4 to 48.9) very low certainty) and low (12.4 m, 95% CrI (−32.9 to 54.2) very low certainty) were non-significant comparisons.

10 m Usual Gait Speed

The effect of each AE prescription compared with usual care on the 10 m Usual Gait Speed is displayed in figure 3D. The posterior mean difference for all prescriptions ranged from 0.05 to 0.34 m/s, with the largest significant effect from HIIT (0.34 m/s 95% CrI (0.13 to 0.55), moderate certainty), followed by MICT (0.20 m/s, 95% CrI (0.05 to 0.35), low certainty). Compared with usual care, the effect of no exercise (0.09 m/s, 95% CrI (−0.04 to 0.23), very low certainty), HICT (0.04 m/s, 95% CrI (−0.04 to 0.12), very low certainty) and low (−0.07 m/s, 95% CrI (−0.21 to 0.07), very low certainty) were non-significant comparisons.

Additional outcomes

Additional outcomes of DBP (n=11 studies), resting heart rate (n=10 studies) and fast gait speed (n=18 studies) are reported in online supplemental appendix 3.0. Briefly, HIIT interventions consistently ranked as the most superior intervention for improving DBP (SUCRA 79%), resting heart rate (49%) and fast gait speed (48%). There were no significant differences between any pairwise comparisons.

Superiority of ranking

Figure 4 displays the SUCRA curves, which indicate the percent probability that each intervention would rank in the top interventions for V̇O2peak (figure 4A), SBP (figure 4B), 6MWT (figure 4C), 10 m Usual Gait Speed (figure 4D). For V̇O2peak, the top three ranking interventions included HIIT (72%), followed by HICT (18%) and MICT (9%). For SBP, the top three interventions included HIIT (49%), HICT (44%) and MICT (4%). For 6MWT, the top three interventions included HIIT (48%), HICT (44%) and MICT (4%). For 10 m Usual Gait Speed, the top three interventions included HIIT (96%), MICT (2%) and no exercise (2%). SUCRA figures for additional outcomes are reported in online supplemental appendix 3.0

Figure 4

Probability of ranking for each intervention for outcomes (A) peak oxygen uptake (V̇O2peak), (B) systolic blood pressure, (C) 6 min Walk Test and (D) 10 m Usual Gait Speed. HICT, high-intensity continuous training; HIIT, high-intensity interval training; Low, low-intensity continuous training; MICT, moderate-intensity continuous training.

Model fits and network inconsistency

All model fit outputs and node-splitting analyses are provided in online supplemental appendix 2.3. For the V̇O2peak NMA, the random effects and consistency model (DIC=102, Dres=57) was superior fitting to the fixed effects model (DIC=128, Dres=95) and the inconsistency model (DIC=106, Dres=60). No meaningful differences were found between all model fit parameters for SBP, 6WMT and 10 m Usual Gait Speed. The node-splitting analyses for the primary outcome revealed no local inconsistency, as evidenced by no statistically significant differences between direct, indirect and NMA evidence (Bayesian p>0.05).

Sensitivity analyses

Sensitivity analyses of the primary outcome and their results are reported in online supplemental appendix 2.5. For the V̇O2peak NMA, eight studies17 60 63 65 67 68 76 88 were identified as potentially outlying through visual inspection of funnel plots. Overall, the effects for all AE interventions versus usual care remained statistically significant, such that posterior mean differences ranged from 0.14 to 3.04 mL/kg/min, with the largest from HIIT 3.0 mL/kg/min (95% CrI 1.0 to 5.0), followed by HICT 2.5 mL/kg/min (95% CrI 1.5 to 3.5) and MICT 2.1 mL/kg/min (95% CrI 1.1 to 2.9). The effectiveness and ranking of interventions remained consistent with the non-adjusted random effects model.

Network meta-regression

Network meta-regression model fits are for the primary outcome of V̇O2peak are in for online supplemental appendix 2.6. Model fit was not meaningfully different compared with the unadjusted random effects model for all potential effect modifiers. Age, sex and training volume did not yield any apparent modifying effect on the effect for each AE prescription.

Discussion

The present NMA represents the most extensive synthesis of AE data from randomised trials in stroke rehabilitation to establish the superiority of different AE interventions. In this systematic review and NMA, we analysed all relevant RCT data involving 47 trials and 2517 participants after stroke to establish the superiority and effectiveness of AE prescription for improving V̇O2peak, mobility and resting SBP. Compared with usual care, most AE interventions resulted in beneficial physiological and clinical adaptations relevant to stroke recovery, namely through improved V̇O2peak and mobility (6MWT distance and 10 m Usual Gait Speed). As hypothesised, we identified an intensity-dependent response such that higher-intensity AE (ie, HIIT and HICT) ranked as the most superior and effective across all outcomes of interest followed by MICT. Clinical decision-making by healthcare professionals and community exercise specialists should adopt a nuanced approach that balances exercise intensity and format, individualised needs and abilities, and specific clinical contexts. Nonetheless, our findings are relevant and may be used to inform updates to clinical practice guidelines.

HIIT, HICT and MICT were, in this order, the three most effective AE interventions for improving V̇O2peak, with effects ranging from 3.2 to 2.1 mL/kg/min. Using the minimally contextualised GRADE framework, we are moderately confident that our HIIT effect estimate is among the most effective intervention for improving V̇O2peak such that the true effect is likely to be close to the estimate of effect. However, due to some concerns with the risk of bias, there is a possibility that the effect estimate is substantially different. Nonetheless, when considering the mean differences for HIIT, HICT and MICT, V̇O2peak increased by at least 1.0 mL/kg/min compared with usual care and other typical interventions (ie, low and no exercise). Our findings are clinically significant given that V̇O2peak changes as low as 1 mL/kg/min are associated with a 10% reduction in cardiovascular mortality,99 and an increase of 1 MET (values in stroke=3.0 mL/kg/min)27 is associated with a 7% risk reduction in stroke hospitalisation100 and 9% lower risk for ischaemic stroke.4 Given the prognostic strength of V̇O2peak for stroke, cardiovascular disease, and all-cause mortality, and that it is often considered the fifth vital sign,101 the addition of V̇O2peak testing in stroke rehabilitation as routine care can be beneficial. However, it is important to acknowledge the clinical feasibility of cardiopulmonary exercise testing and potential for limited testing resources,19 particularly across various practice and low-resource settings. Nonetheless, exercise professionals are encouraged to follow an individualised approach for improved risk stratification, personalised exercise prescription and effective patient management to promote cardiovascular health and fitness after stroke.

The physiological mechanisms underpinning changes in V̇O2peak after HIIT among people poststroke have yet to be fully elucidated. Evidence from non-stroke populations suggests intensity-dependent peripheral and central cardiometabolic adaptations that may contribute to clinical improvements in V̇O2peak.102–107 From a peripheral perspective, an acute bout of HIIT activates signalling pathways involved with mitochondrial biogenesis and repeated activation of these pathways through HIIT may drive adaptations such as increased mitochondrial density, capacity and respiration in skeletal muscle,102–106 associated with increased V̇O2peak.107 While these mechanisms for the poststroke population are unclear, emerging evidence suggests that MICT is more effective than usual care for improving mitochondrial bioenergetics that has been positively correlated with V̇O2peak.59 Further research is needed to examine the intensity-dependent mechanisms underpinning mitochondrial adaptations after AE training among people poststroke.

The effects of AE on central cardiovascular adaptations and arterial function remain poorly understood after stroke. However, in adults with cardiovascular and cardiometabolic conditions, HIIT is superior to MICT for improving measures of cardiac contractility and systolic function,108 and endothelial function.109 In stroke, preliminary evidence suggests that higher intensity AE can improve diastolic function (eg, right atrial emptying fraction) compared with usual care,71 and that HIIT is more effective than MICT for improving systolic function (eg, cardiac output).60 HIIT can also elicit acute increases in novel arterial biomarkers, including insulin-like growth factor one and vascular endothelial growth factor, compared with MICT.110 111 While preliminary, this research highlights the potential for HIIT to elicit superior central cardiovascular adaptations compared with MICT after stroke and emphasises emerging areas of vascular research regarding the potential effects of exercise for improving novel cardiovascular and cerebrovascular biomarkers after stroke.112

Arterial hypertension is widely regarded as the most important modifiable cardiovascular risk factor for stroke; identifying effective non-pharmaceutical interventions, such as exercise, which can reduce SBP after stroke is particularly important.3 In the present NMA, we found that HIIT and HICT may reduce blood pressure by up to 7 mm Hg when compared with usual care and non-exercise interventions. When considering the lower CrI limit, HIIT and HICT may reduce blood pressure by up to 19 mm Hg. While this overall effect estimate is not statistically significant and of low certainty evidence, it may be clinically significant given that a 10–12 mm Hg reduction in SBP is associated with a 30% lower primary stroke risk.113 Our findings are novel, providing updated evidence from previous pairwise analyses (MD=−0.2, 95% CI −4.0 to 5.6) that did not demonstrate potential clinically important reductions in SBP.12 Our findings are the first to disaggregate the effects and superiority of several AE interventions for blood pressure management poststroke.

Importantly, our review determined the superiority of various AE interventions for improving clinical markers of stroke recovery, including the 6MWT distance and 10 m Usual Gait Speed. We identified a dose-response effect, whereby higher-intensity AE tended to be superior for improving 6MWT distance. HIIT demonstrated the greatest effect compared with usual care, although this finding was not significant (41.9 m (95% CI −12.1 to 94.9)). This was followed by significant improvements after HICT (39.4 m (95% CI 16.1 to 66.5)), suggesting that high-intensity exercise may be a key factor in improving walking distance. Given the low-certainty evidence of the included studies, these data are to be interpreted with caution. As such, the true effect may be substantially different from the effect estimate effect due to concerns of risk of bias, indirectness and imprecision.

We also found that HIIT was the superior intervention for improving Usual Gait Speed with effect estimates ranging 0.28–0.34 m/s compared with usual care. We are moderately confident in these effect estimates, and the effects observed are clinically significant, such that the minimal clinical important difference (MCID) for 10 m Usual Gait Speed is 0.13–0.19 m/s114 115 for people poststroke. For 6MWT, we found clinically significant improvements following HICT (39.4 m (95% CrI 16.1 to 66.5)), whereby MCIDs for 6MWT distance in stroke populations have been reported to range between 34 and 50 m.116–118 This finding is of low certainty and requires careful interpretation given concerns to risk of bias and imprecision of effect estimates. Nonetheless, our results extend the findings of a previous pairwise meta-analysis that found that all types of cardiorespiratory training can improve 6MWT distance by 32.1 m (95% CI 10.11 to 54.1) and Usual Gait Speed by 0.08 m/s (95% CI 0.03 to 0.11) compared with no exercise and usual care interventions.12 Given the observed dose–response relationship, where higher AE intensities proved to be both superior and clinically effective, stroke rehabilitation professionals should choose to prioritise HIIT or higher intensity AE using task-specific walking as the training modality54 to challenge the neuromotor and cardiorespiratory systems to promote walking recovery after stroke.

Limitations

Several limitations of this study should be acknowledged. Our risk of bias assessment revealed several methodological limitations that are consistent across all the exercise training literature in stroke, including poor exercise and control group adherence monitoring, lack of reporting of intention to treat analyses, lack of outcome assessment blinding and lack of allocation concealment. In addition, exercise intervention parameters and exercise compliance data were poorly reported, making it challenging to ascertain exercise grouping. To mitigate this challenge, we developed a decision-making algorithm that allows the review team to consistently determine exercise groups between reviewers. We also detected potential publication bias in our V̇O2peak NMA; however, our results remained consistent after excluding these studies in sensitivity analyses. In addition, there was a lack of consistent reporting regarding adverse events and safety outcomes within the literature, thus, we were to conduct meaningful safety analyses. Future research is required to explore the safety of high-intensity AE after stroke to help guide clinical decision-making. Lastly, the BUGSnet R package could only handle one covariate per NMA. Thus, we could not explore the effects of additive effect modifiers on our estimates. Future work using other complex modelling techniques could explore the effects of age, sex and other effect modifiers on V̇O2peak.