Physical fitness interventions for nonambulatory stroke survivors: A mixed‐methods systematic review and meta‐analysis

Abstract Introduction Physical fitness training after stroke is recommended in guidelines across the world, but evidence pertains mainly to ambulatory stroke survivors. Nonambulatory stroke survivors (FAC score ≤2) are at increased risk of recurrent stroke due to limited physical activity. This systematic review aimed to synthesize evidence regarding case fatality, effects, experiences, and feasibility of fitness training for nonambulatory stroke survivors. Methods Eight major databases were searched for any type of study design. Two independent reviewers selected studies, extracted data, and assessed study quality, using published tools. Random‐effects meta‐analysis was used. Following their separate analysis, qualitative and quantitative data were synthesized using a published framework. Results Of 13,614 records, 33 studies involving 910 nonambulatory participants met inclusion criteria. Most studies were of moderate quality. Interventions comprised assisted walking (25 studies), cycle ergometer training (5 studies), and other training (3 studies), mainly in acute settings. Case fatality did not differ between intervention (1.75%) and control (0.88%) groups (95% CI 0.13–3.78, p = 0.67). Compared with control interventions, assisted walking significantly improved: fat mass, peak heart rate, peak oxygen uptake and walking endurance, maximum walking speed, and mobility at intervention end, and walking endurance, balance, mobility, and independent walking at follow‐up. Cycle ergometry significantly improved peak heart rate, work load, peak ventilation, peak carbon dioxide production, HDL cholesterol, fasting insulin and fasting glucose, and independence at intervention end. Effectiveness of other training could not be established. There were insufficient qualitative data to draw conclusions about participants' experiences, but those reported were positive. There were few intervention‐related adverse events, and dropout rate ranged from 12 to 20%. Conclusions Findings suggest safety, effectiveness, and feasibility of adapted fitness training for screened nonambulatory stroke survivors. Further research needs to investigate the clinical and cost‐effectiveness as well as experiences of fitness training—especially for chronic stroke survivors in community settings.

One of the top research priorities, selected by stroke survivors, carers, and health professionals, is to investigate the potential of fitness training to reduce recurrent stroke risk and improve function and quality of life (Pollock, St George, Fenton, & Firkins, 2012).
What is known already is that fitness training facilitates secondary prevention of cardiovascular morbidity (Garber et al., 2011), reduces disability, and improves walking (Saunders et al., 2016), quality of life (Carin-Levy, Kendall, Young, & Mead, 2009), psychosocial functioning (Carin-Levy et al., 2009), and adaptation to life after stroke (Reed, Harrington, Duggan, & Wood, 2010). This evidence underpins guidelines for community-based exercise after stroke services in the UK (Best et al., 2010;Poltawski et al., 2013) and clinical guidelines across the world (Billinger et al., 2014;MacKay-Lyons et al., 2013;Royal College of Physicians Intercollegiate Stroke Working Party, 2016;Scottish Intercollegiate Guidelines Network, 2008Stroke Foundation, 2017). These guidelines mainly pertain to ambulatory stroke survivors, however. There appears to be comparatively little research on fitness training for nonambulatory stroke survivors (Billinger et al., 2014;Saunders et al., 2016; i.e., those unable to walk at all or without physical assistance from at least one other person), who make up approximately 20% of the stroke population (Kwah, Harvey, Diong, & Herbert, 2013;Veerbeek, Van Wegen, Harmeling-Van der Wel, & Kwakkel, 2011); 53 of the 58 studies in the Cochrane systematic review on fitness training after stroke (Saunders et al., 2016) involved ambulatory stroke survivors. Fitness training after stroke often involves walking (Saunders et al., 2016) and is therefore not suitable for most nonambulatory stroke survivors, who are thus disadvantaged by the lack of evidence-based physical fitness training that is adapted to their mobility restrictions. As nonambulatory stroke survivors are inevitably more sedentary than their ambulatory counterparts, their risks associated with prolonged sitting (Rezende, Rodrigues Lopes, Rey-López, Matsudo, & Luiz, 2014) are increased.
In summary, improving fitness in nonambulatory stroke survivors is a top priority, but there is a dearth of evidence-based guidance to inform practice. To the knowledge of the authors, there is no published systematic review on this topic. The aim of this mixed-methods systematic review and meta-analysis was to synthesize published literature on physical fitness interventions for nonambulatory stroke survivors and evaluate the evidence for their effects on fitness, function, activity and participation, quality of life, acceptability, and feasibility.

| Design
This review was designed as a mixed-methods systematic review and meta-analysis. The framework by Thomas, Ciliska, Dobbins and Micucci (2004), designed for synthesizing quantitative and qualitative evidence, was used to comprehensively integrate evidence on case fatality, effects, feasibility, and acceptability. The following sections describe the study eligibility criteria for this review.

| Types of studies
Any type of quantitative, qualitative, or mixed-methods (i.e., comprising a quantitative and qualitative element) study was included (e.g., randomized and nonrandomized, crossover, cohort, and case studies). For the analysis of case fatality and feasibility, data from all included studies were used. For the analysis of effects, only data from randomized controlled trials (RCTs) were used, given the increased risk of bias in non-RCTs; for the analysis of acceptability, data from mixed-methods and qualitative studies were used. Systematic reviews were excluded; however, their reference lists were searched to ensure all relevant studies were included. In order to have access to all relevant data, articles had to be full reports, published in English.
investigate the clinical and cost-effectiveness as well as experiences of fitness training-especially for chronic stroke survivors in community settings.

K E Y W O R D S
exercise, fitness, nonambulatory, rehabilitation, stroke, systematic review

| Types of participants
Only data pertaining to nonambulatory stroke survivors were included, as generalizing from ambulatory participants was considered inappropriate. Nonambulatory adult stroke survivors (age ≥18 years) were included, regardless of type and time since stroke, or any comorbidities. In studies where information about ambulatory status was absent or unclear, authors were contacted. Where it was not possible to obtain data relating to nonambulatory stroke survivors, studies were excluded. To the authors' knowledge, there is no standard definition for "nonambulatory." The Functional Ambulation Category (FAC; Holden, Gill, Magliozzi, Nathan, & Piehl-Baker, 1984) is a validated and widely used tool to describe walking ability after stroke. In this review, "nonambulatory" was defined as an FAC score ≤2, ranging from being completely unable to walk to being dependent on continuous/intermittent physical assistance of at least one person during walking, to help with balance or coordination (Holden et al., 1984).

| Types of interventions
Improving cardiorespiratory fitness is crucial for secondary stroke prevention (O'Donnell et al., 2016) and therefore a key element in many fitness interventions after stroke (Saunders et al., 2016).
Studies were therefore included if published intervention descriptions comprised structured activities aimed at improving healthrelated fitness (Garber et al., 2011). The importance of skill-related fitness was acknowledged; however, studies that focused exclusively on the latter (e.g., mirror-box training to improve dexterity) were excluded. Similarly, voluntary muscle contraction was considered a key intervention ingredient. Therefore, studies were excluded if voluntary muscle contraction was not an essential component of the intervention (e.g., passive movement, electrical stimulation, or diet). Studies comprising only unstructured recreational or occupational physical activity were also excluded, as extracting information about dose would not be possible.

| Types of setting
Interventions delivered in any type of setting (e.g., hospital, laboratory, community) were included, but they had to be land-based.

| Types of comparisons
Studies were not required to have a comparison, but those that did were only included if this provided information about the effects of the fitness intervention, that is, fitness training versus placebo, no intervention, usual care, or another intervention. Studies where a health-related fitness intervention was compared to the same intervention plus an intervention not requiring active voluntary muscle contraction (e.g., a diet) were excluded. Data were compared between baseline and end of intervention, and between baseline and follow-up (where provided).

| Types of outcome measures
Quantitative studies were included if outcomes comprised at least one health-related fitness component, as defined by the ACSM (American College of Sports Medicine, 2013), specified below). Studies were excluded if they only reported skill-related fitness outcomes. Outcomes were categorized into International Classification of Disability and Functioning (ICF;World Health Organization, 2001) domains where possible, to enable comparison to recommended stroke datasets (Geyh et al., 2004;Silva et al., 2015).

3.
Feasibility, operationalized as the number of patients assessed for eligibility and those randomized (or allocated otherwise to an intervention), attendance, number of dropouts and adverse events, and acceptability of the intervention, reported by study participants. Review authors extracted data on dropouts in the period between intervention start and end of study and then categorized these as: possibly intervention-related, general health-related, logistics-related, and refusal to participate-if this could be deduced from the text. Otherwise, dropouts were categorized as unknown or not reported. These data were extracted from all studies included in this review.

| Search terms and databases
A combination of controlled Medical Subject Headings (MeSH) and free-text terms relating to the key search terms of "stroke," "physical activity," and "non-ambulatory" were used to search the
For the analysis of intervention effects, only data from RCTs were used, as this type of design yields the highest quality evidence.
Randomized crossover studies were also included-but only up to and including the point of crossover. Data from non-RCTs were analyzed descriptively only. For a comprehensive overview, data from all included studies are reported in the data tables (Tables 4-7). For the meta-analysis, only outcomes used in two or more RCTs were entered; outcomes used in one study only are described in the text and presented in the tables. To synthesize quantitative data from RCTs, RevMan 5.3 software (RevMan 2014) was used for meta-analysis purposes (Cochrane Collaboration, 2014). Where studies used varying subscales of the same outcome measure (e.g., the full Fugl-Meyer or its lower limb subscale only), the standardized mean difference (SMD) was used instead of the mean difference (MD). Only data reported as standard deviation were entered in the meta-analysis; data presented as standard error were converted to standard deviation before being entered. Data reported as medians and (interquartile) ranges, which did not allow their distribution to be examined for skewness, were not included in meta-analysis (Higgins & Green, 2011). In cases where multiple baseline assessments were reported that were not significantly different, the last baseline measure was used. Final values at the end of intervention and at follow-up (where included) were used. To establish the odds of regaining independent walking, an odds ratio (OR) was computed. Variability was assessed with the Chi-square test for statistical heterogeneity and the I 2 statistic for inconsistency across studies, which are both included in the RevMan forest plots. However, as the Chi-square test has low power in meta-analyses when the sample size is small or when the number of events is small, the significance level was set at 0.10 rather than at 0.05, and a random-effects model was used (Higgins & Green, 2011).
These processes also ensured comparability with the Cochrane systematic review on physical fitness training after stroke by Saunders et al. (2016).
For the analysis of feasibility, relevant data on adverse events and dropouts from all studies were included. For case fatality, the number of deaths in each group and the total number of participants in each group were entered into the meta-analysis as dichotomous outcomes and the odds ratios (OR) were computed.
For the analysis of acceptability of interventions, the plan was to use a thematic synthesis of qualitative data. However, no qualitative studies and only two mixed-methods studies could be included, which had very little qualitative information pertaining to nonambulatory participants, and this is presented narratively.
Following the separate analysis of quantitative and qualitative data, the framework proposed by Thomas et al. (2004) was used to synthesize these data.
The quality of the two mixed-methods studies White et al., 2013) was rated as low, as the overall score is the lowest score of the study components (Pluye et al., 2011; Table 3).
All studies reported the profession of staff delivering the intervention, with the exception of Cho et al. (2015) and Potempa et al. (1995), but exercises were supervised in all studies. Only one study mentioned a home program , but no further details were reported. Seventeen of 33 studies (52%) indicated that participants were given information to aid motivation, but none appeared to include a theory-based strategy.
Intensity of strength training as part of the high-intensity functional exercise program was "high" (i.e., [8][9][10][11][12]) in one study (Rosendahl et al., 2006), "somewhat strong" to "strong" in another study (Leroux, 2005), and not clearly reported in two studies Tsaih et al., 2012). Intensity was monitored in one study only (Leroux, 2005). Session duration ranged from 30 min  to 1.74 ± 0.15 hr . Program duration ranged from 4 weeks  to 3 months (Rosendahl et al., 2006). The number of sessions varied between 12  and 50 . Progression was described in three studies (Leroux, 2005;Rosendahl et al., 2006;Tsaih et al., 2012), but not in Richards et al. (1993). Brisk walking was used in one cohort study (Batcho et al., 2013), but how this was adapted for nonambulatory participants was not explained. Modified jump training was used in one case series . Intensity was set by the patient and therapist, but was not described. Two studies monitored cardiovascular responses (Batcho et al., 2013;Mehrholz et al., 2006).
BWSTT was used in four RCTs Franceschini et al., 2009;Teixeira da Cunha Filho et al., 2001 and four other studies Plummer et al., 2007;Vidoni et al., 2008; Table 6). Session frequency ranged from 3 ×  to 5 × per week Franceschini et al., 2009;Hesse et al., 1994Hesse et al., , 1995Plummer et al., 2007;Teixeira da Cunha Filho et al., 2001. Intensity was not described in any study; Plummer et al. (2007) was the only study to monitor heart rate, while Vidoni et al. (2008) assessed heart rate and blood pressure prior to each session. Session duration ranged from 15  to 30 min Hesse et al., 1995;Plummer et al., 2007). Average program duration ranged from 5 Hesse et al., 1994) to 16 weeks . In other studies, the intervention ended when participants achieved independent walking  or were discharged Teixeira da Cunha Filho et al., 2001. The number of sessions, where stated, ranged from 18  to 45 . Walking was assisted by one or more therapists, while BWS did not exceed 50% in any study. Progression was described in all studies, which was achieved by reducing BWS and/or increasing speed. TA B L E 2 (Continued)

References
Robot-assisted walking training, using a total of four different devices across studies, featured in 11 studies Cho et al., 2015;Hesse et al., 2010Hesse et al., , 2012Husemann et al., 2007;Mayr et al., 2007;Morone et al., 2011;Ochi et al., 2015;Stoller et al., 2015;Tong et al., 2006; Table 5). The Lokomat was used in five studies Cho et al., 2015;Husemann et al., 2007;Mayr et al., 2007;Stoller et al., 2015). The G-EO Systems Robot was used in two studies  and the Gait Trainer (GTII) was used in three studies Tong et al., 2006), while the Gait Assistance Robot (GAR) was used in one study . Training frequency ranged from 1× per week  to 2× per day . Intensity was not specified as such in any of the studies, but some monitored cardiovascular responses Stoller et al., 2015;Tong et al., 2006). Session duration ranged from 15  to 30 min net training time , although the total session duration in Husemann et al. (2007) was 60 min.
Program duration ranged from 2  to 9  weeks, but in Mayr's study  this comprised only two, three-week blocks of Lokomat training. The number of sessions ranged from 4  to 45  and was 20 in most studies Hesse et al., 2012;Husemann et al., 2007;Morone et al., 2011;Ochi et al., 2015). In studies using BWS, this was set at a maximum of 50% and reduced as soon as possible and speed was increased while preserving an optimal gait pattern. Progression in the study by Ochi et al. (2015), who did not use BWS, was not clearly described.

| Cycle ergometer training
Four RCTs Potempa et al., 1995;Wang et al., 2014aWang et al., , 2014b) and one randomized crossover study  used cycle ergometer training, including lower limb cycling ; Yang et al., 2014) or upper/lower limb cycling , while Potempa et al. (1995) did not specify the type of cycling. The study by Lennon et al. (2008) included two "life skills" classes. Frequency ranged from 2×  to 5× per week .

| Other training
Shea et al. (Shea & Moriello, 2014) delivered an adapted, classical Pilates program comprising of exercises in a lying/seated position for 9 months-the longest intervention period reported. Exercises were progressed, but intensity was not reported.  adapted dance techniques, so they could be performed in sitting. Improvisation was used to encourage participants, but otherwise progression was not clear. Intensity, which was moderate, was TA B L E 3 Quality assessment of mixed-methods studies according to the Mixed Methods Appraisal Tool (Pluye et al., 2011)
Only the intervention and usual care groups were included in this meta-analysis. Ng et al. (2008)   All data were extracted from publications, except in cases indicated by: a Data supplied by author, analysed by review authors (ML, FvW). b Median (range). c Analysed data supplied by the author. d Data from all study participants including those who were not non-ambulatory after stroke, where data from the latter were not available. NR data not reported by study authors.

TA B L E 5 Overview of intervention parameters in intervention groups (and control groups where included)
Author ( Data at point of cross-over not presented; data from RAGT phases combined for both groups and compared with data from non-RAGT phase combined for both groups: 1-6: No significant between group differences 7. No between group differences in total MBI but significant difference in "transfer" item in favour of the RAGT group (p < 0.05).

Author (year) Study design Assessment time points and outcome measures Results
Husemann (2007) 1-3, 5-7: no significant between-group differences 4. Significant difference in favour of intervention group in reduction of fat mass (p = 0.012), no significant betweengroup differences in body weight or body cell mass.

Richards (1993) RCT with 3 arms
Baseline and end of intervention at 6 weeks and 3, 6 months follow-up 1.

| Outcome measures
A total of 105 different outcome measures were reported across the 33 studies, including 74 used in single studies only. A total of 44 (42%) were health-related fitness outcomes (Table 6).

White (2013) b
Mixed methods cohort study +

| Assessment times
Baseline measures were reported in all but one study , which only measured outcomes at 6 months post-study entry.

| Setting
Twenty

| Effects of interventions
Outcomes from all studies are reported in Table 6. Five RCTs could not be included in some meta-analyses: Some or all data were presented as medians Husemann et al., 2007;Ochi et al., 2015), end-of-study data were only presented in graphical form , and only one nonambulatory stroke survivor was included in each group , while one randomized crossover study did not report data at crossover point .

| Effects on primary outcomes
Alpha was set at 0.10 instead of the conventional 0.05, for reasons explained in the Section 2.

Case fatality
Out of 33 studies involving 910 participants, 29 studies including 739 participants reported case fatality. Within these, 10/739 deaths (1.35%) were reported over the entire study period: 7/400 (1.75%) in all intervention groups and 3/339 (0.88%) in all control groups (Table 7). There were no deaths in any of the cycling or other intervention-type studies-although two studies (Potempa et al., 1995;White et al., 2013) Table 7). Both deaths occurred in one study , but it was unclear whether this occurred during the intervention itself or just within the intervention period. The difference in case fatality between groups was not statistically significant (OR 0.69, 95% CI 0.13 to 3.78, p = 0.67, I 2 = 0%; Figure 2). There were no deaths in any of the 10 other walking studies (Batcho et al., 2013;Hesse et al., 1994Hesse et al., , 1995Hesse et al., , 2010Hesse et al., , 2012Leroux, 2005;Mayr et al., 2007;Mehrholz et al., 2006;Plummer et al., 2007;Vidoni et al., 2008), while two did not report case fatality Richards et al., 1993).

Cardiovascular and respiratory functions [ICF domain b4]
Cardiac risk score None of the RCTs on assisted walking measured cardiac risk score. One cycle ergometer study measured cardiac risk score; Lennon et al. (2008) reported changes, but due to the small number of participants, only descriptive data are presented (Table 6).
Heart rate One walking study measured resting heart rate ; however, there was no effect compared with the control group. At the end of walking training, there was a significant increase in peak heart rate in the intervention compared to the control group (MD 9.3, 95% CI −0.7 to 19.2, p = 0.07, I 2 = 32%; Figure 4) in three studies Stoller et al., 2015;Teixeira da Cunha Filho et al., 2001). Stoller et al. (2015) found a significant difference in favor of the intervention group in terms of training intensity, heart rate, and heart rate reserve (p < 0.002).
TA B L E 7 Overview of dropouts involving non-ambulatory participants only (intervention period, follow up period-where included) and adverse events
There were no follow-up data.
Oxygen (VO 2 ) uptake At the end of walking training, peak oxygen uptake was significantly increased compared to control interventions F I G U R E 2 Comparison assisted walking training versus control-end of intervention. Outcome: case fatality

Drop out a (number of non-ambulatory stroke participants) during intervention period and follow up period (where included) Adverse events b (number of non-ambulatory stroke participants experiencing event, and event description as stated by authors)
Possibly
Respiratory exchange ratio (RER) At the end of walking training, there was no significant difference between intervention and control groups in peak RER (MD 0.01, 95% CI −0.01 to 0.03, p = 0.34, Chang et al., 2012;Stoller et al., 2015).
At the end of one cycling training RCT (Potempa et al., 1995), no significant difference was found in peak RER; however, there was a significant improvement in peak CO 2 production in the intervention compared to the control group (p < 0.01). There were no follow-up data.
Peak ventilation (VE peak) At the end of walking training, there was no significant difference in peak VE between intervention and control groups (MD 0.87 L/min, 95% CI −4.75 to 6.49, p = 0.76, Chang et al., 2012;Stoller et al., 2015).
At the end of one cycle ergometer training RCT, Potempa et al. (1995) found a significant improvement in peak ventilation in the intervention compared with the control group (p < 0.01). There were no follow-up data.
Other cardiorespiratory functions After walking training, Stoller et al. (2015) found no significant difference in any of their cardiorespiratory performance measures (Table 6) compared with the control intervention.
These findings were echoed in the RCT by Chang et al. (2012).
After cycling training, one RCT  reported changes in forced expiratory volume; however, only descriptive data could be presented (Table 6).
Workload One walking training RCT (Teixeira da Cunha  found no significant difference in workload during exercise testing between walking and control groups at intervention end. At the end of one cycle training RCT (Potempa et al., 1995), a significant improvement in workload was found during maximal exercise in the intervention compared to the control group (p < 0.0001). Lennon et al. (2008) reported changes in peak wattage following their cycling intervention, but due to the small number of participants, only descriptive data are presented (Table 6). There were no follow-up data.
F I G U R E 3 Comparison assisted walking training versus control-follow-up. Outcome: case fatality F I G U R E 4 Comparison assisted walking training versus control-end of intervention. Outcome: peak heart rate (bpm) F I G U R E 5 Comparison cycle ergometer training versus control-end of intervention. Outcome: peak heart rate (bpm) Rate of perceived exertion Rate of perceived exertion (RPE) was assessed in two walking training RCTs: No significant differences between intervention and control groups were found at the intervention end Franceschini et al., 2009) or at follow-up ).
One cycle training RCT  assessed RPE, but due to the small numbers of nonambulatory participants, only descriptive data are presented (Table 6). There were no follow-up data.
Exercise tolerance One walking training RCT measured the total time pedaling during the testing protocol (Teixeira da Cunha , but found no significant difference between intervention and control groups at intervention end.

Metabolic functions [ICF domain b5]
Body weight At the end of robot-assisted walking, one RCT  found a significant reduction of fat mass compared with conventional walking rehabilitation (p = 0.012); however, there were no significant between-group differences in body weight or body cell mass. There was no follow-up. Morone et al. (2011) was the only study on walking to measure BMI at baseline and discharge (but not end of intervention); however, no significant between-group difference was found.
At the end of cycle ergometer interventions, there was no significant difference in body weight between groups (MD −0.58 kg, 95% CI −8.12 to 6.97, p = 0.88, I 2 = 48%; Potempa et al., 1995;. Lennon et al. (2008) measured waist girth and BMI, but as there were only four ambulatory participants in each group, only descriptive data are presented (Table 6).
There were no follow-up data.
Serum lipid profiles None of the walking training RCTs measured any serum lipid levels.
One cycle training RCT measured total cholesterol ; however, due to the small number of nonambulatory participants, only descriptive data are presented (Table 6). Two cycle training RCTs ) measured total triglycerides: Following the end of the intervention, there was no significant difference between intervention and control groups (MD −0.18 mmol/L, 95% CI −0.59 to 0.23, p = 0.39, I 2 = 98%).
Two cycle training RCTs ) measured high-density lipoprotein (HDL) and low-density lipoprotein (LDL): Following the end of the intervention, HDL levels had improved significantly in the intervention compared to the control group (MD 0.06 mmol/L, 95% CI 0.00 to 0.13, p = 0.07, I 2 = 0%; Figure 7).
In contrast, the same two studies   to control groups . Furthermore, by combining data on fasting glucose and 2-hr plasma glucose, Wang et al. (2014a) found that significantly more participants in the intervention (48%) compared to the control group (18%) improved their glucose tolerance (p < 0.05).

Movement-related functions [ICF domain b7]
Walking endurance A mix of 5MWT and 6MWT was used across studies; therefore, the average distance per minute walking during Only one cycling study measured walking endurance ; however, there was only one nonambulatory stroke survivor in each group (Table 6).
At the end of walking training, there was no change in the MIlower limb subscale between groups (MD 1.8, 95% CI −5.9 to 9.5, p = 0.65, I 2 = 20%; Chang et al., 2012;Mayr et al., 2007;. Three further studies found no significant differences in the MI at intervention end Husemann et al., 2007;Tong et al., 2006). Mayr et al. (2007) used the MRC scale, but due to the small numbers involved, only descriptive data are presented (Table 6). Rosendahl et al. (2006) used the 1RM to measure leg strength; however, due to the small number of nonambulatory participants, only descriptive data are presented (Table 6) Three studies conducted a follow-up Morone et al., 2011;; there was no significant effect of walking compared to control interventions on the MI. Findings from the meta-analysis (MD 6.5, 95% CI −1.9 to 14.9, p = 0.13, I 2 = 0%) agreed with those by Franceschini et al. (2009).
None of the cycle training studies included any measures of muscle strength or power.

Mobility [ICF domain d4]
Measuring walking outcomes in a nonambulatory population was challenging, and different studies used different protocol adaptations (although they were not always described); for example, in some studies participants were allowed to use devices (including parallel bars) and assistance from staff, while in others this was not permitted. In some studies, walking was only evaluated in those able to walk, while in other studies outcomes were scored as "zero" if participants were unable to walk independently or without aids, walk continuously, or complete the required time or distance. In other studies again, if participants were unable to Two further RCTs showed significant improvements in the FAC compared to control interventions Tong et al., 2006); however, two other RCTs Husemann et al., 2007) found no significant between-group differences at the end of the intervention. Three walking RCTs conducted a follow-up using the FAC; Ng et al. (2008) found a significant improvement in the FAC in favor of the intervention group (p = 0.018).
FAC data in the study by Morone et al. (2011) were not presented in a format that could be used for this meta-analysis. In that study, four groups were compared (Table 5) and the only significant improvement found was in the walking compared to the control subgroups that included participants with more severe paresis (p = 0.001). This showed that the odds of becoming an independent walker at the end of a walking intervention increased 2.73-fold compared with the control group (OR 2.73, 95% CI 0.97-7.71, p = 0.06, I 2 = 51%; Figure 12).
None of the cycling interventions reported the odds of regaining independent walking.
Walking speed After assisted walking interventions Mayr et al., 2007;Richards et al., 1993;Rosendahl et al., 2006;Teixeira da Cunha Filho et al., 2002;Tong et al., 2006;Tsaih et al., 2012), there was a significant improvement in maximum walking speed in the intervention compared with the control group (MD 0.10 m/s, 95% CI 0.01 to 0.18, p = 0.02, I 2 = 67%). Rosendahl et al. (2006) also measured self-paced walking speed, but there was virtually no change in either intervention or control group, both at the end of intervention and follow-up.
Of the remaining walking RCTs, Franceschini et al. (2009) and  found no significant between-group differences in speed during the intervention period, while Ochi et al. (2015) found a trend toward improvement in the intervention compared with the control group (p = 0.07). Six RCTs on walking training F I G U R E 1 0 Comparison assisted walking training versus control-end of intervention. Outcome: walking endurance (m/min) F I G U R E 11 Comparison assisted walking training versus control-end of intervention. Outcome: FAC included a follow-up; however, Richards et al. (1993) did not report data. Meta-analysis including four RCTs Morone et al., 2011;Rosendahl et al., 2006) showed no significant difference between intervention and control groups at 6month follow-up (MD 0.11, 95% CI −0.05 to 0.27, p = 0.19, I 2 = 71%), and neither did Franceschini et al. (2009).
Only one cycling study measured walking speed ; however, there was only one nonambulatory stroke survivor in each group, whose outcomes did not change (Table 6).
Gait kinematics At the end of the walking intervention, Husemann et al. (2007) found no significant between-group differences in cadence, stride duration, stance duration, or single support time.
This study did not include a follow-up. At follow-up, Dean et al. (2010) found no significant differences in stride length between intervention and control groups, measured in participants who had become able to walk. The study by Richards et al. (1993) included gait kinematics, but data were not reported, while Yagura et al.
None of the cycle interventions measured gait kinematics.
Self-rated walking Using the modified EU Walking Scale, Mayr et al. (2007) found that average scores in both groups had improved at the end of the walking-based intervention, but due to the small number of nonambulatory participants, only descriptive data are presented (Table 6). There was no follow-up. At the end of the walking intervention, nor at follow-up, did Franceschini et al. (2009) find any between-group difference in the Walking Handicap Scale.
In contrast, Dean et al. (2010) found a significant improvement on a self-rated walking questionnaire in the walking compared with the control group at 6-month follow-up (MD 1.0, 95% CI 0.1 to 1.9).
None of the cycling interventions assessed self-reported walking ability.
Mobility At the end of walking training, Elderly Mobility Scale scores significantly improved in the walking compared to the control group in two RCTs Tong et al., 2006), as well as at follow-up (Ng et al., 2008).
The average time for the Timed Up and Go improved in one RCT  in both intervention and control groups following walking training; however, due to the small sample, no further analysis was undertaken (Table 6). At the end of walking training, average Rivermead Mobility Index (Gross function) scores improved in the RCT by Mayr et al. (2007), but due to small numbers, no further analysis was undertaken. Morone et al. (2011) did not report data at the end of their intervention, but at follow-up, they noted a significant improvement in the walking compared to the control subgroups that included participants with more severe paresis (p = 0.001). There were no significant between-subgroup differences between those with less severe paresis.
None of the cycling studies included any mobility measures.

Movement-related functions [ICF domain b7]
Voluntary movement control At the end of walking training, a significant improvement was seen in the Fugl-Meyer (lower limb) scores compared with control interventions Richards et al., 1993;MD 3.19, 95% CI −0.17 to 6.55, p = 0.06, I 2 = 0%; Figure 13). However, two further walking RCTs found no significant between-group differences in Fugl-Meyer scores . At follow-up, Richards et al. (1993)) found no significant difference between intervention and control groups in the Fugl-Meyer (lower limb and balance) scores.

Across cycle ergometer interventions, different sections of the
Fugl-Meyer were used; therefore, the SMD instead of the MD was computed. Following training, no significant differences were seen in three studies (Potempa et al., 1995;Wang et al., 2014aWang et al., , 2014b; SMD 0.59, 95% CI −0.26 to 1.43, p = 0.17, I 2 = 82%), while in the study by Yang et al. (2014), only one nonambulatory stroke survivor took part in each group, both of whom showed minimal improvement (Table 6).

Trunk control Two walking training RCTs used the Trunk Control
Test Morone et al., 2011 found no significant difference between the intervention and control groups, either at the end of intervention or at follow-up. Morone et al. (2011) did not report end-of-intervention results, but at discharge, there was a significant improvement only within the subgroup of participants with severe paresis who had undertaken walking training, compared with the control group (p = 0.001).
Balance At the end of walking training, there was no significant difference between intervention and control groups in the Berg Balance Scale (BBS; MD 3.97, 95% CI −1.28 to 9.21, p = 0.14, I 2 = 0%; Richards et al., 1993;Rosendahl et al., 2006;Tsaih et al., 2012). One further RCT  also found no significant between-group difference in balance.  Richards et al. (1993) found no significant difference between intervention and control groups in the Fugl-Meyer (balance) score at follow-up.
None of the cycling RCTs included any balance outcomes.
Falls Only one study assessed the number of falls and the percentage of fallers; although no data were available for the intervention end, Dean et al. (2010) reported no significant differences between walking training and control groups at 6-month follow-up.
Resistance to passive movement Resistance to passive movement was assessed with the Ashworth Mayr et al., 2007;Morone et al., 2011) or modified Ashworth Husemann et al., 2007) scales in five walking training RCTs.
At the end of walking training, two RCTs Husemann et al., 2007) found no significant between-group difference in resistance to passive movement. Morone et al. (2011) did not report data at intervention end, and the number of participants in the study by Mayr et al. (2007) was too small for further analysis (Table 6). At follow-up, Franceschini et al. (2009) and Morone et al. (2011) found no significant difference between groups in this outcome.
One cycling study indicated that the modified Ashworth scale had been used, but data were not reported .

Body functions [ICF domain b]
Morone et al. (2011) was the only study to use the Canadian Neurological Scale at baseline and at discharge, but not at intervention end. All groups improved, but between-group differences were not specified.

Sensory functions [ICF domain b2]
Proprioceptive sensibility of the lower limb was assessed in one walking training RCT ; no significant differences were found between the intervention and control groups at the end of intervention or follow-up.
One study used the Albert's Test for perceptual neglect , but no significant between-group differences were found at the end of the walking training intervention or at follow-up.

Mental functions [ICF domain b1]
Anxiety and depression None of the walking RCTs assessed effects of training on psychological function, including cognition or mood.
Only one cycle training RCT  used the Hospital Anxiety and Depression Scale (HADS). However, as only four nonambulatory participants were included in each group, only descriptive data are presented (Table 6).

Activities and Participation [ICF domain d]
The Barthel Index (BI) or modified BI was used in eight walking RCTs including a crossover study Franceschini et al., 2009;Husemann et al., 2007;Morone et al., 2011;Richards et al., 1993;Tong et al., 2006;Tsaih et al., 2012); however, only data from Ng et al. (2008), Richards et al. (1993),  could be entered into the meta-analysis, as Morone et al. (2011) only reported a p value (<0.029), and reasons for not including other studies were stated above. No significant difference between intervention and control groups was found at the end of intervention (SMD 0.20, 95% CI −0.28 to 0.67, p = 0.42, I 2 = 0%). The remaining RCTs also found no significant difference in BI between intervention and control groups at intervention end Husemann et al., 2007;Tong et al., 2006).
The Functional Independence Measure (FIM) was used in five walking training RCTs (Ng et al., 2008;Ochi et al., 2015;Teixeira da Cunha Filho et al., 2001;Tong et al., 2006;, although different sections were used: Teixeira da Cunha  used the locomotor subscale and Ochi et al. (2015) used the mobility subscale, while  and Tong et al. (2006) used the full FIM instrument and the paper by  included graphs of the FIM total, motor, and gait subscales. There was no significant difference between intervention and control groups, both at the end of the intervention Tong et al., 2006;at follow-up (Ng et al., 2008), in any of these outcomes. One walking training RCT used the Adelaide Activities Profile . Baseline data were not reported, and outcomes were only measured at 6 months after study entry. At that point, no significant differences between the intervention and control groups were found. At follow-up, Franceschini et al. (2009) and found no significant between-group differences in the BI. This was in contrast to Morone et al. (2011), who did find a significant difference-but only in favor of the subgroup of participants with the low F I G U R E 1 3 Comparison assisted walking training versus control-end of intervention. Outcome: Fugl-Meyer (lower limb) motricity intervention group compared to those in the control group (p = 0.006). Richards et al. (1993) found no significant difference between intervention and control groups in the Barthel Ambulation score, both at the end of the intervention and at follow-up.
Two walking training RCTs used the Rankin  or modified Rankin Scale . Franceschini et al. (2009) found no significant difference between the intervention and control groups either at the end of the intervention or at follow-up. At discharge, Morone et al. (2011) only found a significant improvement in favor of the subgroup of participants with low motricity partaking in the intervention compared to the control group (p < 0.029).
At the end of the intervention, cycle ergometer training resulted in significant improvements in favor of the intervention groups in the BI in two studies (MD 19.5, 95% CI 13.8 to 25.2, p < 0.00001, Wang et al., 2014aWang et al., , 2014b by the same author. There were no follow-up data. One cycle ergometer study  used the Frenchay Activities Index. However, due to the small number of nonambulatory participants, only descriptive data are provided (Table 6).
tervention groups across all walking interventions, compared with 47/299 (16%) in the control groups, with another six nonallocated adverse events reported by Stoller et al. (2015). Reasons for dropout, considered by the review authors to be exercise interventionrelated, included anxiety associated with treadmill training  and discomfort from wearing the harness . Cho et al. (2015) did not report any specific figures but attributed a "high dropout rate" to deteriorating health status and "adverse dermatological effects." Across all cycling interventions, there were 9/49 (18%) dropouts in the intervention and 10/49 (20%) in the control groups. Reasons for dropout, considered to be exercise intervention-related by the review authors, included discomfort in the affected leg . In the other intervention category, White et al. (2013) did not specify the ambulatory status of their only dropout. In the remaining two studies Shea & Moriello, 2014), one of six participants (17%) had an adverse event in the intervention groups (there were no control groups in this category) and there were no dropouts from adverse events considered to be intervention-related.

| Acceptability of the interventions
There were no qualitative studies, and only two cohort studies White et al., 2013) incorporated a qualitative element, exploring participants' views on the intervention provided. During their dance intervention, the instructor kept a journal containing participant feedback , but there was no feedback from any of the nonambulatory stroke survivors. Following Masterstroke, a mixed exercise and education program (White et al., 2013), semistructured interviews were conducted, in which three of four nonambulatory participants took part.
The themes and quotes described below were all linked to nonambulatory participants by the study authors.

| Perceived benefits
All participants in the Masterstroke program (White et al., 2013) valued the exercise component. One of the nonambulatory participants highlighted how perceived improvements in strength and stamina helped with getting up and down off a chair, while another expressed how they benefited from encouragement by health professionals. Participants also reported improved balance and mobility following the dance intervention . The benefits of group exercise were expressed in both cohort studies White et al., 2013), as expressed by participants feeling less isolated and reassured by peer support. Participants reported feeling more positive following a group-based dance intervention . Music was also expressed as an important social factor for reminiscing and enjoyment of the intervention. In addition to health benefits, psychosocial benefits from being in a group included vicarious learning and sharing empathy with other stroke survivors (White et al., 2013). In the dance intervention , all participants derived a sense of pride from performing in front of a small audience, which they indicated as their favorite component.

| Goal attainment
Goal setting was a central component of the Masterstroke program (White et al., 2013), and although not everyone achieved theirs, participants appreciated that the exercises were aimed at their personal goals.

| Lifestyle modification
One nonambulatory participant expressed that knowing staff at the gym was a key element in maintaining motivation to exercise after completing the Masterstroke program (White et al., 2013). The same participant also reported that information on diet was important to maintain body weight following study end. on the low number of intervention-related adverse events, a low dropout rate, and similarity in case fatality between intervention and control groups over the intervention period, most intervention procedures included in this review could reasonably be considered to be feasible.

| D ISCUSS I ON
Other key findings related to study quality, participants, interventions and comparisons, outcome measures, settings, and effects, feasibility, and acceptability will be discussed below.

| Study quality
Study quality varied; most studies were rated as "moderate." Selection bias affected all studies, with few reporting the proportion of participants agreeing to participate, or sufficient information to judge the representativeness of the study population. These aspects could be better reported in future.

| Participants
The lack of clear and standardized descriptors of ambulatory ability levels made it difficult to select and compare relevant studies.
Despite utilizing the criterion of FAC score ≤2, a clinically diverse group was included in this review, which might have led to heterogeneity in intervention effects (Higgins & Green, 2011). Future studies should attempt to specify participants' walking ability using a standardized scale (e.g., the FAC), to enable better comparison of studies.
Only a few studies included participants more than 6 months poststroke. In this population, it is particularly important to prevent recurrent stroke, which accounts for approximately 30% of all stroke (Hankey, 2014)

| Interventions and settings
Most studies used walking interventions, assisted by therapists, BWST, and/or robotic equipment. As most participants were within 3 months poststroke, the emphasis on walking seemed appropriate, as this is an important rehabilitation goal at this stage (Jørgensen et al., 1995). The use of electromechanical devices may be feasible within a rehabilitation setting (although none of the studies reported costs); however, within community settings, cost, space, and staff training requirements may pose barriers. Importantly, this type of training precludes the opportunity for social interaction with peers, which is an important motivator for stroke survivors (Nicholson et al., 2013). Only six studies (Batcho et al., 2013;Lennon et al., 2008;Leroux, 2005;Rosendahl et al., 2006;White et al., 2013) used group training, and only five were undertaken in the community Batcho et al., 2013;Leroux, 2005;Shea & Moriello, 2014) including care homes (Rosendahl et al., 2006;Tsaih et al., 2012). This highlights an important gap, as guidelines recommend the continuation of fitness training-preferably in group format-after hospital discharge (Best et al., 2010;Billinger et al., 2014 Most interventions were of a short duration, except for one walking (Rosendahl et al., 2006) and one Pilates intervention (Shea & Moriello, 2014). Therefore, the limited effects found in this review may partially be due to the short training duration.
All interventions were tailored to individuals, but methods were not always described sufficiently to enable replication-with the exception of the study by Shea & Moriello (2014).

| Comparisons
Most studies that included a comparison group comprised usual care, but without sufficient detail to enable replication (Table 5). Some variation is unavoidable due to the individualized nature of stroke care; however, more detailed reporting (Hoffmann et al., 2014;Slade et al., 2016) will increase reproducibility and comparability of usual care. Most studies with usual care as the comparator were dosematched; however, some of the electromechanical gait studies were confounded by preparation time. and participation-a division also reflected in the ICF core set for stroke (Geyh et al., 2004). The predominantly biomedical approach to research on fitness training after stroke, which emerges from this review, is also demonstrated by the lack of psychosocial outcomes, with only one study (White et al., 2013) evaluating quality of life.

| Outcome assessment
Given the high prevalence of anxiety and depression after stroke (Hackett & Pickles, 2014), further research on the effects of fitness training on mood is warranted (Sims et al., 2009).
Importantly, none of the studies included any measure of costs.
A recent study demonstrated the cost-effectiveness (Collins, Clifton, van Wijck, & Mead, 2018) of a clinically effective community-based fitness training program for ambulatory stroke survivors (Mead et al., 2007), but more health-economic evidence is required for service development.
Taken together, this review indicates that studies using assisted walking interventions primarily assessed skill-related and only few health-related fitness outcomes, whereas the reverse seems to be the case in studies evaluating cycling interventions. This pattern offers limited opportunity for comparing assisted walking and cycling intervention categories. Therefore, in order to strengthen this body of evidence, an agreed standardized toolkit of outcome measures is required that are valid and meaningful to service users and providers, reflect a biopsychosocial paradigm, and include health economics measures.

| Effects on primary outcomes
The majority of RCTs used an ITT analysis, but in those that did not, treatment effects may have been subject to bias (Higgins & Green, 2011).

Case fatality
Fatalities were rare; deaths only occurred in walking intervention groups, but these comprised the majority of participants. There was no suggestion that fatalities occurred during the intervention itself. Between intervention end and follow-up, risk of death was increased 4.75-fold for participants in walking-based interventions, but this was only borderline significant. Case fatality in the review by Saunders et al. (2016) was even lower; 0.46% of all participants died before intervention end and 0.72% before follow-up. The low number of deaths may relate to stringent criteria, whereby participants with contraindications to exercise were excluded. It is also likely that participants were self-selected, with only those feeling able agreeing to take part. Together, these points question the external validity of the findings, but underline the importance of thorough screening as one of the factors underpinning low case fatality.

Cardiovascular and respiratory functions
Assisted walking training improved peak heart rate, peak oxygen uptake capacity, and oxygen consumed during walking, suggesting better aerobic fitness. However, this evidence was based on three RCTs of moderate-to-strong methodological quality only Stoller et al., 2015;Teixeira da Cunha Filho et al., 2001. Medication and age may influence heart rate within this population, and therefore, results may not represent the actual cardiac training effect. The improvement in peak oxygen uptake was below the minimal clinically important difference (MCID) of 10 ml/kg/ min (Puente-Maestu et al., 2016). As there were no follow-up data, longer-term benefits of assisted walking training remain unknown.
Measures of peak cardiopulmonary performance were collected by two high-quality walking training RCTs only Stoller et al., 2015). Stoller et al. (2015) noted that despite their intervention group reaching a significantly higher training intensity than the control group, they did not manage to maintain their target because of fatigue. Chang et al. (2012) attributed the limited effect of training to the short intervention period, which was only 2 weeks.
These observations suggest that the training dose may not always have been sufficient to reach an effect.
Cycle ergometer training improved peak heart rate, work load, peak ventilation, and maximum carbon dioxide production compared with controls at intervention end, but the evidence was more limited than in the walking-based studies. Evidence for benefits on peak heart rate was based on three RCTs including one low-quality (Potempa et al., 1995) and two high-quality RCTs , but evidence for the remaining outcomes was based on one low-quality RCT only (Potempa et al., 1995). As there were no follow-ups, any carryover effects remain unknown. In contrast, in mostly ambulatory stroke survivors, cardiorespiratory training did improve peak oxygen uptake and exercise tolerance (Saunders et al., 2016), suggesting that these effects cannot be generalized to nonambulatory stroke survivors.

Metabolic functions
There was a paucity of data on the effects of assisted walking on risk factors for stroke.  , so would need to be replicated before any conclusions can be drawn. Impaired glucose tolerance, a measure recognized by the World Health Organization (World Health Organization, 2006), may be more clinically relevant than fasting glucose per se in future studies, as it is a known risk factor for atherosclerosis and stroke.
As these findings show potential for fitness training to contribute to secondary stroke prevention-a recognized research priority (Pollock et al., 2012)-future studies should include measures of serum lipids, insulin sensitivity, or glucose tolerance.

Movement-related functions: walking endurance and strength
Assisted walking resulted in a borderline significant improvement in walking endurance at intervention end and a significant improvement at follow-up, compared to control interventions. When converted to the distance walked in 6 minutes, the effect might also be clinically significant, exceeding the MCID of 34.4 m (Tang, Eng, & Rand, 2012)-however, challenges in undertaking walking-based outcomes in a nonambulatory population complicate interpretation.
This evidence was based on five RCTs, comprising one low-quality  and four moderate-quality Mayr et al., 2007;Morone et al., 2011;Teixeira da Cunha Filho et al., 2001 studies. However, one high-quality RCT ) that could not be included in the meta-analysis found no significant effect at the end of intervention or follow-up. These findings align with the review including mostly ambulatory stroke survivors (Saunders et al., 2016).
Mixed training in the cohort study by White et al. (2013) resulted in patient-reported improvements in strength and stamina.
However, it was difficult to corroborate these perceptions in other studies using more objective measures Franceschini et al., 2009;Husemann et al., 2007;Mayr et al., 2007;Tong et al., 2006). These findings align with those from Saunders et al. (2016), where effects of fitness training on strength were inconsistent.

Mobility
The effect of assisted walking on walking independence, assessed with the FAC, was uncertain, both at the end of the intervention and at follow-up. This evidence is based on eight RCTs, including four high-quality Franceschini et al., 2009;Ochi et al., 2015), three moderate-quality Morone et al., 2011;Teixeira da Cunha Filho et al., 2001, and one low-quality RCT . There was no significant benefit from walking compared with control interventions in terms of the percentage of independent walkers at the end of the study. However, at follow-up, two medium-quality RCTs Morone et al., 2011) showed a significant 2.73-fold increase in the odds of achieving independent walking in the intervention compared to the control group. This effect may be due to an increase in habitual walking following discharge from hospital, and this would be useful to examine with activity monitors in future.
These findings concur to some extent with the Cochrane systematic review (Mehrholz, Thomas, Werner, et al., 2017) on electromechanical-assisted gait training, which found that this technology increased the chance of independent walking in dependent walkers. This comparison needs to be interpreted with caution, however, as "dependent walkers" were defined as those with an FAC <4 (which includes those requiring supervision but able to walk without mechanical assistance), while data were analyzed at intervention end only. A comparison with the Cochrane systematic review on treadmill training and body weight support  could not be undertaken, however, as this did not differentiate between outcomes in ambulatory and nonambulatory participants.
The effects of walking training on self-reported walking ability compared with control interventions were based on two medium-quality studies Dean et al., 2010) and one high-quality RCT .
It was challenging to obtain reliable measures of gait kinematics in this population, and any changes need to be interpreted with caution. For example, an increase in speed may have been the result of fewer participants scoring "zero" in some studies. Walking training significantly improved maximum walking speed in intervention compared to control groups, but effects were lost after the intervention end. This evidence is based on eight RCTs, including one high-quality , four moderate-quality Mayr et al., 2007;Rosendahl et al., 2006;Teixeira da Cunha Filho et al., 2002), and three low-quality studies Tong et al., 2006;Tsaih et al., 2012). In the systematic review by Saunders et al. (2016), effects of fitness training on walking endurance and speed did carry over after the intervention, which suggests that training for nonambulatory stroke survivors might need to continue, possibly because it may be more difficult for this population to practice safely and independently. Walking training did not improve any gait kinematics at the end of the intervention or at follow-up, but only three RCTs (two medium-quality Husemann et al., 2007) and one low-quality RCT ) were able to measure a selection of these. Effects of walking training on mobility were mixed, with significant improvements in the Elderly Mobility Scale shown in one low-quality  and one high-quality RCT (Ng et al., 2008) at the end of the intervention and in one RCT at follow-up , but inconclusive findings in the Rivermead Mobility Index and TUG due to a paucity of data Morone et al., 2011;Tsaih et al., 2012).

Movement-related functions
Evidence for the effects of fitness training on voluntary movement control, trunk control, balance, falls, and resistance to passive movement was limited. The effect of assisted walking training on voluntary motor control, assessed with the Fugl-Meyer, was uncertain. This evidence is based on two high-quality Ochi et al., 2015) and two low-quality RCTs . Walking training did not improve trunk control compared with controls at intervention end, while data at followup were inconclusive. Evidence for trunk control was based on one high-quality ) and one moderate-quality  RCT. Walking training, compared to control interventions, had no significant impact on balance at the end of the intervention, but between end of intervention and follow-up, there was an indication of improvement. This is perhaps to be expected, as during the intervention, participants would have been supported by therapists and/or equipment, but afterward, without such support, participants' balance would have been challenged more often during habitual daily activities. This evidence is based on five RCTs, including one high-quality , one moderate-quality (Rosendahl et al., 2006), and three low-quality RCTs Tong et al., 2006;Tsaih et al., 2012). The effect of walking training on falls could not be established, due to a paucity of data.
As falls prevention is an important clinical consideration in nonambulatory stroke survivors (Bernhardt, Ellis, Denisenko, & Hill, 1998), future studies should include valid measures of balance and falls.
Walking training did not seem to have any differential impact on resistance to passive movement. This evidence is based on one highquality ) and four moderate-quality Husemann et al., 2007;Mayr et al., 2007;Morone et al., 2011) studies, suggesting that fitness training does not exacerbate hypertonia.
Cycling resulted in no significant benefit in voluntary motor control, assessed with the Fugl-Meyer, compared with control interventions. This evidence came from two high-quality  and one low-quality (Potempa et al., 1995) RCT. This is perhaps not surprising, as the Fugl-Meyer does not comprise any cyclical actions. The effects of cycling training on balance, trunk control, and resistance to passive movement are not known, as these measures were not included or reported.

Body and Sensory functions
Effects of walking training on neurological function (CNS), lower limb proprioception, and perceptual neglect were inconclusive due to a paucity of data.

Mental functions
The effects of walking on mood are not known, as none of the walking RCTs included an outcome to this effect. One cycle training RCT assessed mood, but findings were inconclusive due to a paucity of data. The systematic review on fitness training by Saunders et al. (2016), which included mostly nonambulatory stroke survivors, found inconsistent effects on mood. The impact of fitness training on mood is an important gap in the evidence, as many stroke survivors experience depression and/or anxiety (Kim, 2017). Participants in a mixed training/education program (White et al., 2013) expressed psychosocial benefits from group-based training, including enhanced motivation to exercise and benefits from seeing how others had learned to cope with a similar condition. These findings are worthy of further investigation.
None of the studies assessed the effects of fitness training on cognition (the top research priority selected by stroke survivors, carers, and health professionals (Pollock et al., 2012), which should be explored in future studies, especially as other reviews have shown benefits of exercise after stroke on cognition (Cumming, Tyedin, Churilov, Morris, & Bernhardt, 2012;Garcia-Soto, Lopez de Munain, & Santibanez, 2013)).

Activities and participation
Most of the walking training RCTs showed no significant benefits for activity and participation compared to control interventions, as assessed with the FIM, BI, or Adelaide Activities Profile. This evidence is based on 12 RCTs, including three high-quality Ochi et al., 2015), five moderate-quality Dean et al., 2010;Husemann et al., 2007;Morone et al., 2011;Teixeira da Cunha Filho et al., 2001), and four low-quality studies Tong et al., 2006;Tsaih et al., 2012;. Two walking RCTs (one high quality ) and one moderate quality ) examined the effects of training on stroke-related disability, assessed with the (modified) Rankin Scale, but found no difference compared with controls. It is plausible that walking training, which comprises repetitive practice of a specific cyclical task, does not carry over to tasks that are discrete and complex. The lack of effect of fitness training on disability (other than walking-related) was echoed in the systematic review by Saunders et al. (2016).
Cycling resulted in a significant improvement in the Barthel Index (BI) at the end of training, based on two high-quality RCTs by the same author . Changes in the BI, following cycle ergometer, were clinically important, as the detected mean difference was 19.4 points, much higher than the MCID of 1.85 points (Hsieh et al., 2007). These promising findings need to be replicated in other studies, however, before any conclusions can be drawn.

| Feasibility
Reporting of recruitment rates, dropouts, adverse events, and attendance varied; only just under 50% of studies included in this review fully reported this information. However, it must be acknowledged that many studies were published before the CONSORT guidelines (Schulz, Altman, & Moher, 2010). Across studies reporting this information, on average 22% of all patients screened were eligible, but for planning future studies, more consistent reporting of this number is required.
Attendance, although only reported in just over 50% of studies, was generally high, which supports feasibility. However, better reporting of attendance, which is also poorly reported in exercise stud- there were very few intervention-related adverse events, which included anxiety associated with treadmill training , discomfort from wearing the harness  and "adverse dermatological effects"  in walking interventions, and discomfort in the affected leg during cycling . Fatigue was commonly reported, but did not necessarily lead to dropout. In this review, only dropouts in the period between intervention start and end of study were noted, but between randomization and intervention start, 29 additional dropouts occurred, in many cases because participants were not able to tolerate the study's exercise testing protocol.
Experiences from only three nonambulatory stroke survivors could be included in this systematic review, which were generally positive: Participants reported benefits from the exercise component that was tailored to their goals, helped to increase strength and stamina, and provided a supportive group atmosphere providing mutual support from peers and professionals (White et al., 2013).
However, it is clear that more research is required to gain a deeper understanding of participants' perceptions of fitness interventions in order to optimize their uptake and maintenance.

| Strengths and limitations
In terms of the evidence included in this review, there was a paucity of high-quality quantitative-and particularly qualitative-evidence, as discussed above. These limitations impact on the conclusions that can be drawn in this review, and recommendations for strengthening the evidence base will be discussed below.
In terms of review methodology, a systematic and comprehensive literature search was conducted. However, despite best efforts, other relevant studies may have been overlooked. Reporting of ambulatory status was generally poor, and although authors were contacted where required, data were not always available, and therefore, some studies had to be excluded. Studies in languages other than English also had to be excluded, due to resource limitations. Taken together, these limitations mean that not all potentially relevant literature could be included in this review.

| Implications for practice
This review provides evidence that assisted walking and cycle ergometer training may improve health-and skill-related fitness, as well as functional outcomes in carefully selected nonambulatory stroke survivors, but no firm conclusions could be drawn. Training did not carry over into activity and participation, however; therefore, if these domains were to be among the participant's personal goals, they would require more targeted interventions.
Adverse event reporting was patchy; however, the low incidence of intervention-related adverse events and similarity in case fatality over the intervention period suggest that the adapted interventions, delivered by qualified staff, were safe for those who had been selected. Although the evidence requires strengthening, postponing implementation until such time would mean that this population remains sedentary and at high risk of further cardiovascular disease. Therefore, health and exercise professionals, as well as policymakers, should be encouraged to create more opportunities where this emerging body of evidence can be implemented judiciously by suitably trained professionals, to enable nonambulatory stroke survivors to become less sedentary and more physically active (Ezeugwu & Manns, 2017;Kerr, Dawson, Robertson, Rowe, & Quinn, 2017).

| Implications for future research
Descriptions of different levels of walking ability after stroke need to be agreed and standardized to enable better comparison between studies. One of the strengths of this review is the attempt to use a standardized tool to describe the term "nonambulatory," that is, the FAC (Holden et al., 1984). This may facilitate comparison across studies in future and enable further research to build upon this review.
To strengthen the evidence and facilitate trial planning, future studies should improve their reporting of a number of aspects, especially the number of participants initially approached, as per CONSORT guidelines (Schulz et al., 2010). Reporting of intervention-related adverse events should be improved to provide a more accurate estimate of safety. Future studies should also report all components of fitness interventions and comparisons, as per TIDieR (Hoffmann et al., 2014) and CERT (Slade et al., 2016)  More qualitative or mixed-methods studies are required to gain deeper insight into participants' experiences of interventions, to ensure these are acceptable, aimed at what matters most to them, and encourage maintenance of physical activity.

| CON CLUS ION
This mixed-methods systematic review and meta-analysis on the case fatality, effects, experiences, and feasibility of physical fitness interventions for nonambulatory stroke survivors showed emerging evidence that assisted walking and cycle ergometer training, com-

ACK N OWLED G M ENTS
The first author was funded by a Glasgow Caledonian University PhD bursary. The author team would like to thank all authors who made their data available for this review.

ML declares no conflict of interests. DS is a Director of Later Life
Training, a not-for-profit training organization, which delivers