Exercises

For rate of falls, based on pooled results from eight studies (765 participants) investigating exercises to prevent falling in people after stroke as well as our sensitivity analysis for single interventions (626 participants), we have low confidence that exercise may reduce the rate of falls in favour of the experimental group. Based on our pooled results, exercises resulted in a 28% reduction of the fall rate; and when considering exercises as a single intervention, a 34% reduction of fall rate was observed. Sensitivity analysis for studies conducted in the chronic phase after stroke (205 participants), however, showed little or no difference in the rate of falls. There was also little or no difference in the rate of falls when only studies at low risk of bias were included (462 participants).

For the number of fallers, based on pooled results from 10 studies (969 participants) and sensitivity analyses for single interventions (796 participants), chronic phase after stroke (375 participants), and low risk of bias (462 participants), we have very low confidence in the finding that there may be little or no difference in the number of fallers within the exercises group compared to controls. Overall, our findings suggest that exercises may result in a beneficial reduction in the number of times that fallers fall, but to date there is no clear evidence that exercises change fallers into non‐fallers. The certainty in the evidence (GRADE) for this analysis ranges from very low (number of fallers) to low (rate of falls).

Results for secondary outcome measures were sparse, with the exception of quality of life, but because of the heterogeneity of outcome measures used we decided not to pool these results. Studies assessing the effect of exercises on preventing falls included a total of 14 measures of quality of life. In four of these measures (Adelaide Activities Profile 'service to others' subscale, FES‐I, SF‐36 mental dimension and mental health subscales, and SIPSO), an improvement was reported in favour of the experimental group.

Environmental adaptations

Rate of falls or number of fallers did not appear to reduce when comparing the intervention to the control condition.

• Social environment: predischarge home visits compared to predischarge interviews in the hospital (very low GRADE evidence, 85 participants). Mood was found to be increased more in the home visits group.

• Vision improvement: provision of single lens distance vision glasses instead of multifocal glasses (very low GRADE evidence, 43 participants).

• Aids for personal mobility: use of a servo‐assistive rollator compared to a conventional walking‐oriented therapy (very low GRADE evidence, 42 participants).

All analyses within the 'Environmental adaptations' section relied solely on one study.

Interpretation of results

In contrast to the findings of the original version of this review, exercises may reduce falling post stroke, based on a reduced rate of falls. Our results showed different results in rate of falls and number of fallers. Absence of data on rate of falls from Green 2002 and Holmgren 2010 may have resulted in different findings for the pooled analysis of rate of falls. Of interest are the results of Marigold 2005 and Ada 2013, which are the only trials individually demonstrating a reduction of rate of falls resulting from an intervention programme consisting of agility exercises and treadmill training, respectively. It should be noted that there is a difference for the analysis and subsequent result between Marigold 2005 and our review. In Marigold 2005, the number of falls and the number of fallers were analysed with a Mann‐Whitney U test and a Chi² test respectively, and showed no statistically significant between‐group difference. Based on the information from Marigold 2005, we were able to calculate the parameters of interest for inclusion into our meta‐analysis (see Analysis 1.1 and Analysis 1.2). Surprisingly, for the rate of falls this resulted in a statistically significant between‐group difference in favour of the agility programme (Analysis 1.1). For planning future trials and implementation in clinical practice, this trial seems to give an important message about the content of an intervention to prevent falls. Ada 2013 used two intervention conditions related to exercises, forcing us to pool the two intervention groups into one group to circumvent the problem of study multiplicity in the meta‐analysis. This merge might have caused an under‐ or overestimation of the true effect of the individual exercise programmes. However, in meta‐analyses a general pooling of related control interventions is suggested, therefore this potential bias is intrinsic to the analysis.

Andrade 2017 found tDCS to reduce falling post stroke. However, these results need to be interpreted with caution since the results were obtained from a single study with limited sample size, of which the power analysis was not conducted for the number of fallers. Interestingly, a recent meta‐analysis investigating the effect of tDCS on balance found little or no effects (Li 2018), which seems contrary to the results of Andrade 2017, since balance is an important predictor of falling post stroke (Ashburn 2008; Maeda 2015). Falls remain complex, however, for they are predicted by multiple factors (Ashburn 2008).

Results in Haran 2010 are only to be interpreted for post‐stroke patients who are regular wearers of multifocal glasses.

Regarding environmental adaptations to reduce falling post stroke, Drummond 2012 used in parallel to the randomised controlled trial a cohort study which included patients to whom the clinicians believed a home visit was essential. This has a clinical advantage, but might reduce statistical quality. Furthermore, they described procedural problems in administering the home visits, which were reported to be altered in a definitive trial. Finally, they also found the mean (SD) cost of a home visit to be GBP 183 (GBP 81), which might be an extra limitation for implementation in low‐income countries.

Our hypothesis that fall‐prevention strategies are more effective in earlier phases post stroke remains unconfirmed since we found too few studies with a population in that early rehabilitation post‐stroke phase. Future updates might be potent to answer this query.

In this update, we have used a broad categorisation of interventions to synthesise data, and have not further documented details of interventions. We note that variability is present regarding intervention type within our exercises' meta‐analyses. In our opinion the heterogeneity is substantial, and forms — in combination with the limited number of included studies — the main reason why we did not perform sensitivity analyses on the effect of type of exercise intervention. Exercise, furthermore, is an umbrella term covering a wide range of interventions, which should be described in detail in order to understand key elements such as dose, content and intensity. In a future update of this review, specific analyses of type of exercise intervention could be possible if both future exercise trials are conducted and trials provide adequate detail of exercises investigated. Future sensitivity analyses could investigate whether multiple or multifactorial exercise interventions are more beneficial than single‐type or focused exercise interventions, in order to understand the differential effect of type of exercises provided on falls.

Other remarks

It should be further noted that some trials reported interesting post hoc analyses. Marigold 2005 showed that for their participants with a history of falls, eight out of 15 continued to fall in the intervention group compared with 13 out of 15 in the control group (P = 0.05). Dean 2012 indicated fewer falls in the intervention group for their fast walkers but more falls in the intervention group for their slower walkers. Both studies contribute to the current belief that interventions should be developed for specific subgroups of people with stroke. Again, this information can contribute to the future development of interventions for preventing falls in people after stroke. Andrade 2017 found that the group receiving sham tDCS experienced a higher fall risk compared to the other participants (P < 0.01). This result — combined with their discovery of a bilateral tDCS montage to be associated with significantly higher Berg Balance Scale (BBS) and Falls Efficacy Scale ‒ International (FES‐I) scores compared to anodal and cathodal currents — might be of use when considering a broad scope on fall prevention, including measures for fall risk. It should be noted, however, that only people with high risk of falling were included in the trial. Additionally, no differences between montages were found on the Four Square Step Test (FSST) and Overall Stability Index (OSI), which were both used as outcome measures to quantify the risk of falling.

Holmgren 2010 used an intervention that was classified as multiple, since it comprised exercises, implementing exercises into real‐life situations and educational sessions. The trial found no significant reduction of number of fallers, but a significant increment in quality of life. In addition, the complex aetiology of falls might point towards a more holistic interventional approach, outlining the importance of this idea for planning future trials in this field.

We could neither perform meta‐analyses nor narrative description regarding the benefits of interventions targeted at reducing falling post stroke on our secondary outcomes, due to lack of reporting on these outcomes in the included studies. Future studies are advised to include these outcomes as part of the trial design.

Fall assessment

A major concern arising from the trials included in this review is the definition of falls. Of the 16 trials, only nine provided a definition of a fall, and among these seven different definitions were used, with two trials presenting a definition not referenced to previous literature (Dean 2010; Taylor‐Piliae 2014). Although the content of these different definitions might not be significantly different, uniformity should be sought in future trials evaluating interventions for preventing falls in people after stroke, with a consensual definition of a fall, such as the one developed by the Preventions of Falls Network Europe (Lamb 2005).

Also relating to fall assessment, four of the included studies only collected data on falls during the intervention period. The disadvantage of this approach is that the presence of a latency of physiological changes might cause an overestimation in fall occurrence since benefits from exercises are not yet physically established (Lamb 2005). On the other hand, there may also be a decay of beneficial effects during follow‐up. These limitations strengthen our recommendation to follow the methodological gold standards of fall research presented by Lamb 2005.

Power calculation

For nine out of 14 trials, the primary aim was to prevent falls. For seven of the nine, a power calculation was performed based on establishing a reduction in falls/fallers. Andrade 2017 performed a power calculation to establish this on another primary outcome measure (Four Square Step Test (FSST)) and Taylor‐Piliae 2014 reported about a power calculation to determine their sample size but neither specified time point (a priori or post hoc) nor outcome measure used. The single aim of Holmgren 2010 was to prevent falls, but the power calculation was based on finding an increase in the Berg Balance Scale score, leading to a total of 34 participants recruited. The non‐significant finding for falls and fallers may have resulted from an inadequate sample size for this outcome. Unpublished stroke subgroup results from Haran 2010 should also be interpreted with caution; this analysis is underpowered, since a limited group of the total study sample was selected for our analysis. Future trials need to be of adequate size, with a power calculation based on reasonable estimates of effect size, resulting in trials with many more participants. As an example, Marigold 2005 calculated that, based on their fall data, a sample size of 292 participants per exercise group would be required to detect differences for the number of fallers in a definitive trial. Future trials will probably need to be multicentred and perhaps international.

Risk of bias

The causes of falls in people after stroke are complex, and a trial aimed at preventing falls requires a complex intervention. We assessed the majority of trials included in this review as having a low risk of selection, attrition and reporting biases, as well as a reliable ascertainment of falls/fallers outcome (see Figure 3). Of concern was the level of detection bias (blinding of outcome assessment for falls). In the majority of trials (93.75%), we assessed this as being at high risk of bias, as studies used a self‐reported questionnaire or a falls calendar or diary. These methods rely on active registration by the participant, with telephone calls to the participants if (monthly) fall calendars are not returned. Nevertheless, we assessed this as being at high risk of bias, since the assessors, who were in this case the participants themselves, were probably not blinded to group allocation, and because the accuracy of prospective reporting methods may lead to over‐ or under‐reporting of falls (Lamb 2005). In an attempt to mitigate this risk, Mansfield 2018 used one blinded assessor to interview the participant about the circumstances of the fall, and two blinded assessors to determine if any falls should be excluded from the analysis. This approach could serve as example for future studies and warrants further consideration. Kunkel 2011, comparing retrospective interviews and prospective falls diaries over a 12‐month period in a cohort of 122 people with stroke, found an 83% agreement between the methods in the classification of fallers. Yet frequent repeat fallers reported falls during the retrospective interview but did not record all falls in the diary. Excluding these outliers, a similar number of falls were reported using either method. Our results for detection bias and the findings from Kunkel 2011 indicate that monitoring falls accurately in a chronic, community‐dwelling population remains difficult. Although prospective methods are considered preferable (Hauer 2006), future trials could include both retrospective and prospective methods. Nevertheless, preliminary studies investigating novel assessments of falls, such as portable activity monitors, seem warranted for future research.

To correct for included studies at high/unclear risk of bias (scoring high/unclear risk on one item in addition to the 'blinding of outcome assessment' item), we performed an additional sensitivity analysis in the exercise comparison. Surprisingly, this produced results conflicting with the original significant finding in rate of falls. Subsequent stepwise addition of studies at high/unclear risk of bias yielded results that strongly depended on study selection. Of interest is our observation that by reincorporating Ada 2013 in the analysis, the significant finding returned regardless of the amount of remaining studies at high/unclear risk of bias that were subsequently reincluded. Ada 2013 scored unclear risk of bias at concealment methodology, since details were lacking regarding the adopted concealment method. A conclusive statement with respect to study quality is that more well‐designed, low bias, adequately powered trials are needed to increase quality of the evidence, which would be potent to more robustly outline the true significance of fall‐oriented interventions.

GRADE

This updated version included the GRADE approach according to Atkins 2004. The quality of the evidence ranged from very low in the majority of the comparisons to low in the two comparisons that were found to be significant. The most frequent reasons for downgrading were the lack of blinding, which is difficult to perform on outcome measures related to fall assessment, and the fact that most comparisons only consisted of one study. Future updates of the review might show increased quality of the evidence due to inclusion of more studies per comparison.

Stroke research

Our search strategy identified two systematic reviews on the topic of falls prevention after stroke. First, Batchelor 2010 included 13 studies and found exercises not to be beneficial in reducing either rate of falls or number of fallers post stroke. There is a discrepancy between the studies included in Batchelor 2010 and our review. In Batchelor 2010, the type of interventions included were all those that may affect falls outcome. On that basis, they also included trials in which falls were classified as an adverse event. Their review therefore contains six trials that we have excluded from our review, because we applied the stricter inclusion criterion that interventions had to be aimed at preventing falls. This is an important distinction. Trials evaluating interventions such as very early mobilisation after stroke or early supported discharge include falls as an outcome. The aim of these interventions, however, is not to prevent falls but to improve functional outcome. Because (very) early mobilisation or early supported discharge might be associated with an increase in falls, they are included as an outcome measure. We believe that a stricter approach — i.e. including only trials where the aim was to prevent falls — is justified, as otherwise study hypotheses are mixed and results become difficult to interpret.

Second, Winser 2018 conducted a meta‐analysis regarding the effect of Tai Chi in neurological disorders, including the effect on falling post stroke. They found Tai Chi to reduce falling post stroke, based on one study that was also included in our meta‐analysis (Taylor‐Piliae 2014). Their results agree with our findings of Taylor‐Piliae 2014, although we merged their two intervention conditions and we used a random‐effects model to evaluate the rate ratio, whereas Winser 2018 used a fixed‐effect model to evaluate an odds ratio. Despite the fact that merging both intervention conditions in our review reduced the effect size, our findings support that Tai Chi might be beneficial in preventing falls post stroke. The superiority of one exercise intervention over the other remains unknown at present, since there is a lack of studies investigating the fall‐preventing capacity of different exercise interventions post stroke, and this is probably also an important avenue for future research.

Elderly research

Finally, we feel that a more in‐depth appraisal of fall‐prevention reviews, focusing on elderly people living in the community, is important to the readership of this article. Besides the larger number of studies included in these reviews, there is a first notable contrast with our results in Gillespie 2012: multiple‐component group exercises and individually prescribed multiple‐component home‐based exercises did show a beneficial effect for reducing the rate of falls and the risk of falling in this population. We should be cautious, however, when considering whether these interventions might be suitable for preventing falls in people after stroke. Three trials included in this review (Batchelor 2012, Dean 2012 and Taylor‐Piliae 2014) examined the effect of an intervention containing an exercise programme developed for older people: the Otago Exercise Program (OEP) in Batchelor 2012; the Weight‐bearing Exercise for Better Balance (WEBB) programme in Dean 2012; and the SilverSneakers national fitness programme designed for older people in Taylor‐Piliae 2014, which was merged with the Tai Chi intervention condition for reasons stated in the Results section of this review. None of these trials showed significant between‐group differences for reducing the rate of falls and the number of fallers, again indicating that specific interventions may be required for preventing falls in people after stroke, with strategies aimed at particular deficits that people have after their stroke.

Other reviews concerning fall prevention in elderly people published contradicting results regarding the fall‐preventing capacity of exercises (Cameron 2018; Grossman 2018; Sherrington 2019; Tricco 2017). Sherrington 2019 substantiates the value of Tai Chi, balance‐oriented and functional exercises. Tricco 2017 found exercises, either alone or part of a multifactorial approach, to be associated with lower risk of injurious falls. Grossman 2018 found a small effect of multifactorial interventions, whereas Cameron 2018 express their uncertainty of the effect of exercises on rate of falls or risk of falling. We note, however, that the latter study was conducted in an institutionalised setting, opposed to the other three reviews assessing community‐dwelling older people. In general, literature therefore concludes that exercises have a beneficial fall‐preventing effect in elderly people, which was reflected in the stroke population in this review update. Nevertheless, we stress the importance that the exercise‐related studies included in our review did not find any beneficial effect on number of fallers. The overall inconsistency of significant fall‐reducing effects of the investigated interventions, combined with issues outlined earlier in this discussion, requires us to be cautious and thus express low confidence in our conclusion. If future updates could reach sample sizes comparable to fall research in elderly people, this cautiousness regarding concluding statements might turn into more robust evidence more effectively informing clinical practice.