The population of interest included older adults with a sample mean age of 65 years or above with no restriction on the injury or disorder type and with no history of falls studied in various research settings (e.g., community, clinics, and long-term care facilities). Studies were included, if at least two experimental groups participated in each of the different exercise interventions or if there was at least one exercise intervention group with a no-exercise control group. Studies involving any non-exercise interventions (e.g., medication, electrical stimulation, or nutritional supplement) were excluded. Details regarding the exercise interventions must have been provided in the reports. The studies must have included at least one reactive balance assessment, which is defined in this study as an assessment using mechanical postural perturbation during a static or dynamic steady-state task. The studies included in this review were restricted to randomized controlled trials (RCTs) and written in the English language.

The following electronic databases were initially searched by one reviewer (Y.K.) from the inception to February 2021: MEDLINE, EBSCO, CINAHL, SPORTDiscus, PsycINFO, PubMed. WorldCat.org, OpenGrey.eu, and PROQUEST were additionally searched for unpublished trials. To keep this search up to date, an updated search followed in August 2021 by two reviewers (YK and MV). Earlier reviews and bibliographies of included studies were reviewed for additional potentially relevant trials. The combination of the following keywords was employed for the database searches: (aged OR aging OR old^* OR elder^* OR senior^*) AND (exercise OR train^* OR activit^* OR rehabilitat^* OR therap^* OR physiotherapy OR hydrotherapy OR conditioning OR exertion OR recreation^* OR aerobic^* OR stretch^* OR strengthen^* OR walk^* OR jog^* OR run^* OR cycl^* OR pilates OR yoga OR tai chi OR ai chi OR dance OR swim^*) AND (reactive postural response OR stepping response OR perturbation OR slip perturbation OR reactive balance OR reactive stepping OR protective stepping OR compensatory stepping OR anticipatory postural adjustment^* OR compensatory postural adjustment^* OR anticipatory postural response^* OR compensatory postural response^* OR anticipatory adjustment^* OR compensatory adjustment^* OR postural adaptation^* OR postural stabili^*ation OR automatic postural response^* OR postural stepping response^*) AND (random^*).

After exporting the references and removing duplicates, titles and abstracts of records were screened independently by two reviewers (YK and MV) according to the eligibility criteria. Full texts of all potentially relevant trials were subsequently retrieved and reviewed to confirm the final eligible trials. Any disagreements were resolved via consensus, and when any disagreement was elusive, a third reviewer (EB) acted as an arbiter.

A total of 46 eligible studies were reviewed and coded in REDCap (https://www.projectredcap.org/) by one reviewer (YK) and confirmed by a second reviewer (M.V.). Any disagreements were resolved via consultation with a third reviewer (EB). The extracted data included: (1) study characteristics; (2) baseline demographics of participants; (3) exercise interventions; (4) reactive balance outcome measures; and (5) results. Exercise categorizations developed by Howe et al. and Sibley et al. were modified in consideration of the purpose of the current research and applied to the coding (Table 1) (Howe et al., 2011; Sibley et al., 2021).

Means (M) and standard deviations (SD) for all eligible outcomes of reactive balance measures at baseline and post-intervention were extracted for the analysis. Missing data related to eligibility and study outcomes (i.e., data not reported either in a text or on publicly accessible data repositories) were requested to the corresponding authors via email. In the case of no response after one month, a second request was sent, if another month elapsed without response, the data were considered irretrievable. If the requested, but not retrieved data were presented in a graphical format rather than numeric data (e.g., tabular format), Engauge Digitizer 12.1 (http://markummitchell.github.io/engauge-digitizer/) was applied for data digitization and extraction.

To ascertain an overall and study-level risk of bias for each trial, a pair of reviewers (YK and MV) independently determined the bias arising from the following domains using the Cochrane risk of bias tool (RoB 2): (1) randomization process; (2) deviations from the intended interventions; (3) missing outcome data; (4) measurement of the outcome; and (5) selection of the reported result (Sterne et al., 2019). Each domain was assigned a judgement of “low risk,” “some concerns,” or “high risk.” Disagreements were resolved through discussion or referral to a third reviewer (EB).

Considering indeterminate baseline similarities of reactive balance measures in several studies, change values from baseline to post-intervention were calculated or directly extracted from the published data. If there were more than one post-intervention measure (e.g., post-intervention and follow-up), only the data immediately following the termination of the intervention phase was used. SDs for changes from baseline (pre) to post-intervention (post) were calculated using the following formula (Higgins et al., 2019):

Corr in the SD_change equation is the correlation coefficient describing how similar the pre and post-interventions were across participants. When the correlation coefficient was not reported, it was set as 0.5 (Fu et al., 2008; Bruderer-Hofstetter et al., 2018; Lai et al., 2018; Wu et al., 2021). In the case of a lower score signifying better performance in reactive balance measures (e.g., reaction time), scale directions were adjusted by multiplying −1 to the M_change data, which led to a greater effect size indicating an improvement. Missing SDs were imputed from standard errors (SE), 90 or 95% confidence intervals (CI). Using the M_change and SD_chang data, standardized mean differences (SMD) and standard errors (SE) were calculated.

To include multi-arm trials, two approaches were adopted to avoid a unit-of-analysis error (Rücker et al., 2017; Higgins et al., 2019). First, all relevant experimental intervention groups composed of the same categories of exercises were combined into a single group. This step enabled a single pairwise comparison between a combined group and a comparison group in each study. Second, in the case of heterogeneous exercise types across all intervention groups, we included all relevant comparisons as a series of two-arm comparisons and reflect the fact that comparisons within multi-arm studies are correlated (Schwarzer et al., 2015). Accordingly, adjusted SEs of the two-arm comparisons in each multi-arm study were computed using “netmeta” package in R software. The majority of the eligible trials consisted of multiple outcomes in each trial. When multiple SMDs were estimated in a single study, therefore, a pooled SMD with SE was computed.

To estimate the comparative effectiveness of exercise-based interventions on reactive balance, we implemented NMA, which incorporates both direct (i.e., head-to-head comparison from pairwise meta-analysis) and indirect comparisons (i.e., from network meta-analysis) in one statistical model. A Bayesian framework of NMA was conducted using Markov chain Monte Carlo simulations, and non-informative prior distributions for treatment effects were adopted (Lunn et al., 2000; Dias et al., 2018). A random-effects model was used considering the clinical and methodological between-study heterogeneity (Sutton et al., 2000; Borenstein et al., 2009). The NMA was conducted for all available exercise interventions included in at least two trials. The analyses utilized a burn-in period (50,000 iterations) and a follow-up period (100,000 iterations) to minimize bias of initial values when the chain reached its target distribution (Brooks and Gelman, 1998). The convergence was assessed using the trace plot, density plot, and Brooks-Gelman-Rubin diagnostic statistics (Brooks and Gelman, 1998).

The overall geometry of the network was presented in a network graph. Based on Bayesian posterior rank probabilities, the ranking of exercise interventions was estimated using a hierarchical tool, the surface under the cumulative ranking curve (SUCRA) score, measured on a scale from 0 (theoretically the worst) to 1 (the best). In addition, a network forest plot was produced with the “no exercise (NE)” as a reference intervention. The posterior distribution of the SMDs was reported using the mean differences fto the reference intervention with 95% credible intervals (CrI), which indicate that there is a 95% probability that the unobserved (unknown) effect estimates would fall within the intervals (Hespanhol et al., 2019). If a 95% CrI contains zero (i.e., null effect representing the null hypothesis), the effect can be considered statistically insignificant (Hespanhol et al., 2019). The relative effects with 95% CrI of all pairs of exercise interventions were reported in a matrix. Consistency, which is the most important assumption underlying a NMA and indicates agreement between direct and indirect estimates in the network (Salanti et al., 2014), was checked using the node-splitting analysis. The original intention of the first subgroup analysis was to conduct a network meta-analysis stratified by characteristics of participants (i.e., healthy and disease-specific). However, due to the insufficient number of exercise interventions to establish a network in each disease category (e.g., Parkinson's disease), the first subgroup analysis was performed by the inclusion of studies with healthy older adults only (78% of all studies) who did not have any disease, injury, or disability at the time of the studies. The second subgroup analysis was conducted by grouping the outcome measures by the types of reactive balance outcomes: (1) simulated slip or trip while walking; (2) simulated forward falls; (3) being pushed or pulled; (4) movable platform; and (5) balance test battery. A sensitivity analysis was carried out using a frequentist framework NMA to appraise the robustness of the results. Sources of statistical heterogeneity and small study bias were not explored due to the insufficient number of trials (k ≤ 5) for each comparison. All data syntheses and statistical analyses were conducted using “Gemtc” (version 1.0-1), “rjags” (version 4-10), and “netmeta” (version 1.4-0) packages in R (Version 4.1.0, R Foundation for Statistical Computing, Vienna, Austria).

A total of 7,394 records were retrieved from electronic databases and two additional records were obtained from other sources, of which 384 studies remained after removing duplicates and screening titles and abstracts. Based on the full-text screening, 46 records fulfilled the eligibility criteria and thus were included for qualitative analysis (i.e., systematic review), whereas seven studies were additionally excluded from the quantitative analysis (i.e., network meta-analysis) due to data not being reported and not irretrievable (Kim and Lockhart, 2010; Okubo et al., 2019; Wang et al., 2019), exercise types not included in the network (Allin et al., 2020; Cabrera-Martos et al., 2020), exercise intervention included in only one trial (Lacroix et al., 2016), and no continuous data reported (Beling and Roller, 2009), resulting in a total of 39 studies for NMA. The schematic flow chart for the selection process is presented in Figure 1, and all included studies are listed in Supplementary Table 1.

The eligible studies represented a total of 1,745 older adults, included in both pre and post-intervention analyses, with the mean age of 71.9 ± 3.9 years (ranged from 65.3–80.9 years). The majority of the studies exclusively included community-dwelling healthy older adults (k = 36). Ten studies reported on older adults selected for a specific disease or medical condition, such as Parkinson's disease (k = 6), post-surgical interventions for knees, hips, or backs (k = 2), postmenopausal women with osteopenia (k = 1), and chronic stroke (k = 1).

The duration and frequency of the exercise interventions ranged from 1 week to 1 year, 1–5 sessions/week, and 15–90 min/session. Of the 46 studies, 16 executed multicomponent (i.e., multifaceted) exercise interventions in at least one group. Reactive balance was assessed before and after the exercise interventions by use of laboratory-induced slip, trip, and falls, external impacts (e.g., pulling or pushing a body part), platform translation, and treadmill perturbation (e.g., rapid change of the speed) while participants were performing a steady-state task, such as standing or walking. Twenty studies provided training with a postural perturbation while standing or walking, and 11 of which implemented task-specific training (i.e., comparable reactive balance task included in the assessment and training) (Wolf et al., 1997; Bieryla et al., 2007; Beling and Roller, 2009; Mansfield et al., 2010; Parijat and Lockhart, 2012; Jagdhane et al., 2016; Morat et al., 2019; Okubo et al., 2019; Wang et al., 2019; Arghavani et al., 2020; Rieger et al., 2020). The characteristics of the participants and exercise interventions are presented in Tables 2, 3, respectively. Outcome measures and main findings are summarized in Supplementary Table 2.

The summary of the risk of bias assessment across all included studies is presented in Figure 2. Detailed results of the assessment are reported in Supplementary Table 3. Overall, the majority of outcomes were at some concerns (50%) and high risk (48%), and only one study was rated as at low risk. Missing outcome data (46%) was the most influential source of high risk of bias. Selection of the reported result (83%), randomization process (76%), and deviations from intended interventions (61%) were also common sources of bias.

Data from a total of 39 studies (n = 1388, age = 71.5 ± 3.9 years) were included in the NMA. Of the 15 exercise types reported in Table 1, 14 types were included in the NMA as functional training was implemented in only one study and consequently included in a disconnected network (Cabrera-Martos et al., 2020). There were 11 multi-arm trials, and three of which consisted of two groups sharing the same exercise type and the third group with another type (Hatzitaki et al., 2009; Ni et al., 2014; Lacroix et al., 2016); thus, data in these two groups were combined into a single group. Two exercise groups in studies by Gatts and Woollacott (2007) and Gatts (2008), str and NE groups in studies by Granacher et al. (2006, 2009), and two exercise groups in studies by Parijat et al. (2015a,b) shared the same participants, respectively. Thus, each of the aforementioned pairs of studies was combined as a single study in NMA. Overall, 17 exercise interventions with either single or multiple exercise components were included in the NMA. The geometric distribution of the network is depicted in Figure 3. When a study involves a trial arm with a combination of the pre-categorized exercise types, the combination was considered as another distinct exercise intervention.

Estimates of all exercise programs against all others in NMA were reported in a matrix (Figure 4). In the 17 exercise programs, SBR displayed the highest probability of being the most effective exercise intervention (SUCRA score = 0.90) for improving reactive balance, followed by pw (SUCRA score = 0.67) and gaitR (SUCRA score = 0.62) (Table 4).

The relative treatment effect estimates of each exercise program with the no-exercise program being the mutual contrast for comparison are presented in a forest plot (Figure 5). SBR, pw, and gaitR demonstrated the largest mean difference vs. NE; however SBR only demonstrated a statistically significant difference when compared to the no-exercise program (mean difference = 2.7, 95% CrI = 1.0 to 4.3). The trace plot, density plot, and Brooks-Gelman-Rubin diagnostic statistics showed good convergence, which signifies our data has converged to a reasonable distribution. Relatively reliable evidence was derived from the statistical consistency between direct and indirect evidence demonstrated by the node-splitting model (p > 0.05). According to the sensitivity analysis using a Frequentist framework of NMA, the ranking based on the P-scores showed identical results (Table 4). The results suggest that our main findings regarding the relative effectiveness of each exercise intervention are robust for future decisions.

In the subgroup analysis for healthy older adults (k = 29, n = 1120, age = 71.5 ± 3.7 years), effects of 12 exercise programs were compared (Figure 6).

According to the SUCRA scores, SBR was the highest-ranked exercise program (0.90), followed by pw (0.71), which was consistent with the ranking in the complete sample (Table 5). The other exercise programs ranked slightly differently from the NMA for the complete sample; however, the rankings based on the SUCRA scores were consistent with those estimated by P-scores in the frequentist framework (Table 5). The relative effects of all exercise interventions compared to NE were presented in Figure 7. A relative effect matrix was additionally created for all comparisons in the healthy older adults (Figure 8). Too few trials in other disease categories (Parkinson's disease: 6 trials, arthritis: 2 trials, osteopenia: 1 trial) and types of exercise interventions were available to establish a network in each category and conduct further disease-specific subgroup analyses.

For the second subgroup analysis regarding the types of reactive balance outcomes, the first type (simulated slip or trip while walking) was analyzed for gaitR vs. gaitNR using a multilevel MA due to insufficient trials in other treatment comparisons (i.e., only one direct comparison). The second (simulated forward falls), third (being pushed or pulled), and fourth (movable platform) types were analyzed using NMA. The fifth type (balance test battery) was not analyzed due to the insufficient number of exercise interventions and direct comparisons to establish a network. When a slip or trip was simulated while walking, participants showed greater improvements in measures of balance recoveries after gaitR training vs. gaitNR training (SMD = 0.60; 95% CI, 0.33 to 0.88). In other types, SBR presented the first or second highest probability of being the best intervention for improving each reactive balance task. The ranking and relative effects of each exercise vs. NE are reported in Table 6 and Figure 9, respectively.

Considering the findings of this study, it would be advisable for clinicians to preferentially include reactive balance training in line with specifically targeted context, direction, and type of postural perturbations, and power training as a secondarily or complementary approach to improve reactive balance in older adults irrespective of their clinical classifications. Multicomponent exercise interventions not including a reactive balance component may not bring as marked changes in reactive balance as a single reactive balance training does, whereas they still have benefits regarding general health and physical functioning. The possibility of task-specific training adaptations with balance training using external mechanical perturbations has far-reaching clinical and research implications. In fact, beyond simply training one specific type of balance reaction (e.g., a slip), future trials may wish to include multiple types of reactive balance tasks in various simulated contexts that are likely to occur in daily life and appraise the generalizability and ecological validity of the trained tasks from a long-term perspective. Moreover, the addition of power training may synergize the effects on functional reflex activities as well as general functional capabilities needed for daily tasks and reducing falls in older adults.

One of the major strengths of the current study is the use of a NMA. The notable advantage of a NMA over a conventional pairwise meta-analysis is the ability to allow for indirect comparisons, accounting for the effects of multiple interventions in a single statistical model (Schwarzer et al., 2015). Thus, a NMA concurrently summarizes both direct and indirect comparisons between multifarious interventions and enables more complex statistical models and broader interpretation. Random-effects models attempt to generalize the results beyond the trials included in the NMA with an assumption that the selected trials are random samples from a larger population (Cheung et al., 2012). Accordingly, the use of a NMA with a random-effects model in this study enhances the applicability and generalizability of study findings. It should be noted that in general, the indirect estimates tend to have greater variance than direct estimates, and the reliability of the indirect estimates are influenced by the number of direct estimates in the network (Dias et al., 2018). Future meta-analyses may wish to assess publication bias and heterogeneity with a greater number of trials in each direct comparison.

The interpretations of the results in the current study are limited due to small sample sizes and the existence of the probable risk of bias in the included studies. For example, only two trials included more than 100 total participants (Bogaerts et al., 2007; Wang et al., 2019). Furthermore, there was heterogeneity in participants and exercise interventions. For example, there were several distinct disease groups, and the frequency and duration were set differently for various exercise interventions pooled together. With further trials, future reviews may wish to break down the analyses on the basis of hypothetical effect modifiers, such as detailed age and disease groups, baseline functional capacities, or dosage of intervention, for more specific clinical decisions. Also, the low number of trials per comparison precluded investigating sources of publication bias and heterogeneity, and the overall risk of bias was appraised as some concern or high-risk level. Thus, a comprehensive search of published and unpublished works of literature with a paired screening process was conducted to guarantee all available literature was identified to reduce the potential risk of publication bias. Considering the number of trials per each direct comparison, sample sizes, and overall risks of bias, thei results of our analyses may as such guide future research. Despite the aforementioned limitations, we believe that this systematic review with a NMA shed light on better understanding effective interventions for reactive balance in older adults via more comprehensive and inclusive analyses of available literature.