Two authors (E.R.A. and L.T.T.) systematically searched PubMed, CINAHL and Hinari databases from the date of inception of the databases until 15 September 2021. They restricted searches to articles involving human participants and written in English and French. The search strategy used the following keywords to query all the databases: exercise training, physical therapy, balance, postural balance, postural control, walking, QoL, stroke and cardiovascular accident. A third reviewer (O.K.) was consulted to resolve conflicts during the title, abstract screening and full-text evaluation. Additional manual searches included conference abstracts, bibliographies of candidate studies and recent systematic reviews for a comprehensive literature search.

Two authors (E.R.A. and L.T.T.) independently extracted data from eligible studies using a data extraction spreadsheet with predetermined content. Data included general information on the publications (first author’s name and year of publication), characteristics of the studies (sample size, randomisation and blinding), participants (age, gender and time since stroke), mode of interventions (AT, RT, or combination), the content of the interventions (modality, frequency, intensity, duration and sessions length), setting (hospital-based, home and/or community-based, or research laboratory-based) and outcomes (observation time points, measurement tools and follow-up). All data and other materials used in the review are available upon request.

The inclusion criteria for study selection were: (1) randomised controlled trials (RCTs) that involved adult patients (age ≥ 18 years); (2) diagnosed with a stroke; (3) executed structured exercise interventions based on AT, RT, or combination; (4) compared experimental groups with active control groups where participants received treatment such as conventional physiotherapy, or a home exercise programme; (5) reported outcome measures to evaluate balance and/or walking capacity and/or QoL.

Studies that included an additional intervention, such as conventional physiotherapy at the same period were excluded from examining the superiority of exercise over other physiotherapy interventions.

The authors used the Physiotherapy Evidence Database Scale (PEDro) (Blobaum 2006) to assess the methodological quality of included RCTs. The risk of bias in selected studies was analysed using the Cochrane risk of bias assessment tool. The risk of bias and quality of evidence were assessed by two authors independently (E.R.A. and L.T.T.). Disagreements between the two review authors regarding the methodological quality or risk of bias of some studies were resolved by discussion, with the participation of a third author (O.K.) if necessary.

Data synthesis was performed with Review Manager software (Version 5.3) under the random effects model in which a p < 0.05 (two-tailed) was considered significant. The treatment effect was measured by computing standard mean differences (SMDs) with 95% CIs. The authors performed subgroup analyses by stratifying results by exercise modes, volume and intensity of AT interventions and intervention provision settings. The sensitivity analysis was carried out by excluding studies with a high risk of bias and those with poor methodological quality to ensure the robustness of the data. A pre-set cut-off point of ≥ 50% was used to select trials for the sensitivity analysis (Nduwimana et al. 2020).

This systematic review and meta-analysis did not require formal ethical clearance because all data were obtained from publicly available sources and were analysed anonymously.

The authors identified 3245 records of possible interest in the electronic database searches. After removing duplicates, screening titles, abstracts, and reviewing full texts, 28 RCTs were eligible for qualitative synthesis and met the study’s inclusion criteria (Figure 1). Ultimately, 25 trials reported sufficient data to be included in the quantitative analysis yielding 1571 participants with sample sizes ranging from n = 12 to n = 128.

Descriptive study characteristics of the 28 included studies are shown in Table 1. It was found that 16 RCTs examined AT, 2 RCTs examined the effectiveness of RT, and 10 RCTs examined the effect of combining AT and RT. All studies included chronic stroke survivors except one (Mead et al. 2007), with the mean time post-stroke ranging from 4.9 to 99.2 months across all studies. Fourteen trials were conducted as hospital-based, six as home or community-based, and six trials as research laboratory-based. Twenty-five trials were considered for quantitative analyses (meta-analysis) because three trials did not report sufficient data (Elsner et al. 2020; Mead et al. 2007; Severinsen et al. 2014).

Of the 28 included trials for qualitative synthesis, 25 (89.3%) were of good methodological quality (Ada et al. 2003; Aguiar et al. 2020; Bonnyaud et al. 2014; Clark & Patten 2013; Combs-Miller et al. 2014; Danks, Pohlig & Reisman 2016; Drużbicki et al. 2016; Elsner et al. 2020; Gama et al. 2017; Globas et al. 2012; Gordon, Wilks & McCaw-Binns 2013; Ivey et al. 2015, 2017; Janssen et al. 2008; Jin et al. 2013; Lamberti et al. 2017; Lee et al. 2008, 2015; Lund et al. 2018; Marzolini et al. 2018; Mead et al. 2007; Olney et al. 2006; Ouellette et al. 2004; Quaney et al. 2009; Severinsen et al. 2014), and the remaining trials were of fair quality (Dean, Richards & Malouin 2000; Lo et al. 2012; Macko et al. 2005).

The majority of the included studies (78.6%) were carried out with a low risk (Cochrane risk of bias score > 3) of bias (Ada et al. 2003; Aguiar et al. 2020; Bonnyaud et al. 2014; Clark & Patten 2013; Combs-Miller et al. 2014; Danks et al. 2016; Drużbicki et al. 2016; Elsner et al. 2020; Gama et al. 2017; Globas et al. 2012; Gordon et al. 2013; Ivey et al. 2015, 2017; Jin et al. 2013; Lamberti et al. 2017; Lee et al. 2008, 2015; Marzolini et al. 2018; Mead et al. 2007; Olney et al. 2006; Quaney et al. 2009; Severinsen et al. 2014), and the remaining trials presented a risk of bias score less than three (Dean et al. 2000; Janssen et al. 2008; Lo et al. 2012; Lund et al. 2018; Macko et al. 2005; Ouellette et al. 2004).

Nine trials involving 405 participants reported the effects of exercise modes on balance (Appendix 1, Figure 1-A1). The analysis showed that neither AT alone nor AT combined with RT significantly improved balance.

Nineteen trials yielding 807 participants reported post-intervention effects of exercise modes on walking capacity (Figure 2). The overall analysis showed a significant effect in favour of experimental interventions (SMD = 0.28 [0.05, 0.51], p = 0.02). In the subgroup analysis, AT interventions were more effective than the control interventions on walking capacity, while RT or the combination of AT and RT were not.

Seven trials yielding 366 participants reported post-intervention effects of exercise modes on QoL (Figure 3). The overall analysis demonstrated a significant effect in favour of experimental interventions (SMD = 0.56 [0.12, 0.98], p = 0.01). In the subgroup analysis, AT combined with RT was more effective compared with the control interventions on QoL, while AT alone was not.

The authors compared the effect of AT interventions on walking capacity according to the dosage of interventions (Figure 4). With a duration of at least 120 min of exercise per week and an intensity of at least 60% of heart rate reserve (HRR) or rating of perceived exertion (RPE) above 14/60 per session (high dosage), AT interventions were more effective compared with the control interventions (SMD = 0.58 [0.12, 1.04], p = 0.01) while a lower to moderate dosage of AT interventions (< 120 min/week and < 60% HRR or RPE < 14/20) were not.

Eleven trials yielding 508 participants reported the effect of AT setting provision on walking capacity (Figure 5). Analyses showed an improvement in the walking capacity in favour of interventions executed in the hospital setting (SMD = 0.57 [0.06, 1.09], p = 0.03).

Treatment significance effects remained similar across different analyses involving only trials with PEDro scores ≥ 6 and Cochrane risk of bias scores > 3 (Figure 6).

Our study provides an updated review of the current evidence related to the use of exercise training protocols and optimal dosage to improve functioning in patients with chronic stroke. So far, meta-analysis of our study is the first to explore the effect of setting on different intervention types. Finally, our review strictly focused on structured exercise programmes for AT and RT interventions. This strategy prevents any parallel effects that may arise from the other means of rehabilitation or other programmes.

However, our results have some potential limitations. Firstly, only one study was maintained in which the impact of exercise on acute or sub-acute stroke was studied. As a result, the stratification of the studies according to the stages of stroke planned in our protocol could not be carried out. Secondly, the modality of exercise (e.g. treadmill, cycling, overground) was not assessed. It is plausible that the modality of exercise could impact the outcomes of interest, especially balance and walking capacity.