Influence of Reactive Balance Training Program Characteristics on Reactive Balance Control and Fall Risk: A Systematic Review and Meta-Regression Authors: Hadas Nachmani, Laura Kathleen Langer, Augustine Joshua Devasahayam, Avril Mansfield Published: 2026 ABSTRACT AbstractIntroductionReactive balance training (RBT) programs have been developed to address age-related deterioration in reactive balance control and increased fall risk. Despite the demonstrated effectiveness of those programs, there is significant variability in intervention characteristics (e.g., perturbation type, total volume and intensity of training) and study findings. It is likely that intervention effectiveness depends on the intervention features; however, little is known about the optimal way to deliver RBT. The purpose of this systematic review and meta-regression was to determine the optimal intervention characteristics for RBT for improving reactive balance control and preventing falls in daily life.MethodsWe searched MEDLINE ALL (July 2023), Embase (July 2023), Physiotherapy Evidence Database (August 2023), and Cochrane (July 2023) for randomized controlled trials of RBT that reported measures of reactive balance control and/or falls in daily life. Results were screened by two reviewers independently to determine eligibility. The following details were extracted: study population; intervention characteristics (number, frequency and duration of sessions, type, intensity and number of perturbations; description of the control), number of participants in each group; reactive balance outcomes pre- and post-intervention, and number of falls post-intervention. Risk of bias (RoB) and certainty of evidence (GRADE) were assessed. Meta-regressions were performed to explore the influence of different study components on reactive balance control and falls.ResultsAfter screening 7,677 records, 32 studies were included; 25 reported a reactive balance outcome, and 19 reported falls in daily life. RoB of reactive balance control revealed main concerns arising from selection of reported results. RoB of falls in daily life had high or some concerns in the measurements of the outcome and selection of reported results. RBT programs that included manual perturbations were associated with reduced fall rates compared to waist-pull perturbations (rate ratio: 0.45; 95% confidence interval: 0.22, 0.91, p = 0.042). There were no other significant relationships between any other training parameters and falls in daily life or reactive balance control. Quality of evidence (GRADE) was low for reactive balance control and very low for falls in daily life.ConclusionOptimal RBT training characteristics remain unclear due to substantial variability in training protocols and underreporting of key intervention details across studies, which limited meaningful synthesis and interpretation of findings. Although programs incorporating manual perturbations were associated with reduced fall rates, no other training parameters showed consistent relationships with reactive balance control or falls in daily life. Future RBT studies should provide more detailed and standardized descriptions of RBT protocols, include head-to-head comparisons of training parameters, and report outcomes for both reactive balance control and real-world falls to strengthen the evidence base. Introduction Falls among older adults contribute to morbidity and mortality, with significant economic costs. Falls are a major contributor to trauma-related hospitalizations, and only about half of those admitted to hospital after a fall will be alive a year later [1, 2]. Balance recovery abilities resulting from unexpected loss of balance decline with age and, as a result, there is an increased risk of falling in older adults [3]. Diverse fall prevention exercise programs have been developed to address this issue. In a meta-analysis, Sherrington et al. [4] found that older adults who participated in balance training programs reported 23% fewer falls and were 15% less likely to fall than those who completed a control intervention that did not include balance training. Reactive balance training (RBT) is a type of balance training that is focused on improving control of rapid balance reactions that are necessary to avoid falling after a loss of balance [5]. With RBT, participants repeatedly experience a loss of balance (or balance perturbation) so that they can practice and improve control of balance reactions, such as step reactions [6]. It is possible that this balance recovery-specific type of balance training may be even more effective for fall prevention. Indeed, another systematic review [7] found that people who participated in RBT reported 40% fewer falls than those who participated in a control intervention. Despite the demonstrated effectiveness of RBT for preventing falls in daily life, there is significant variability in intervention characteristics (e.g., type of perturbations, total volume of training) and study findings [7]. It is likely that the effectiveness of the intervention depends on its specific features. For example, the type of perturbation can be demonstrated in various ways: waist pull involves applying a pull to the participant’s waist while standing or walking, manual perturbations are delivered randomly by the trainer, overground perturbations are described as slips on the tile or unexpected obstacles raised above the ground during walking on a runway, and surface perturbations which are delivered by a moving platform on which the participant stands or walks. The features of these perturbations may influence the overall effectiveness by targeting reactive balance in different ways. However, little is known about the optimal way to deliver RBT. The purpose of this systematic review and meta-regression was to determine the optimal intervention characteristics for RBT (e.g., volume of training, perturbation type, perturbation intensity) for improving reactive balance control and preventing falls in daily life. Methods Study Design This study is a systematic review and meta-regression, based on a previous review by Devasahayam et al. [7], with an updated search conducted in July 2023. The review of literature was conducted according to Cochrane guidelines [8] and reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [9], the PRISMA Checklist and PRISMA Abstract Checklist are included in the online supplementary Materials (for all online suppl. material, see https://doi.org/10.1159/000552354). Eligibility Criteria Studies were included in this review if they met the following criteria: (1) written in English; (2) included an RBT intervention, either alone or as part of a multicomponent exercise program; (3) included a population at increased risk of falls (e.g., older adults; people with neurological conditions; orthopedic conditions; any other condition that increases the risk of falling, as identified by the study authors); (4) included a control group that did not practice RBT; (5) random allocation of participants to RBT or control groups (randomized control trials [RCTs]); and (6) outcomes included reactive balance control and/or falls in daily life after the intervention. We define RBT as a training method with a goal of improving reactive balance control; for example, RBT targets reactions elicited by slip- or trip-like perturbations and incorporates diverse challenges designed to train individuals to recover balance and prevent a fall after a loss of balance. With RBT, participants may experience internal and/or external balance perturbations that cause loss of balance and evoke balance reactions (e.g., step or grasp responses). Information Sources and Search Strategy The following search engines were used for data collection: MEDLINE ALL (in Ovid, including Epub Ahead of Print, In-Process & Other Non-Indexed Citations, Ovid MEDLINE Daily; 1946 to 25 July 2023), Embase (in Ovid, including Embase Classic; 1947 to 25 July 2023), Physiotherapy Evidence Database (PEDro; searched on 1 August 2023), and Cochrane 2014 to 25 July 2023). The search strategy was developed and implemented by an information specialist. The full search strategy is provided in online supplementary material 1. Selection Process After the initial search, duplicates were removed, and records identified through other methods were added (e.g., reference lists of systematic reviews (n = 281); reference list of other records (n = 3), see Figure 1). Titles were independently screened by two reviewers, and ineligible titles were removed (n = 4,341 excluded; n = 60 conference abstracts/trial registrations; n = 9 not retrieved). A total of 675 reports were screened to determine eligibility. When eligibility was unclear from the abstract, the full text was reviewed. Of these, 622 reports were excluded for several reasons (e.g., not RBT (n = 504); not balance-impaired population (n = 26); not randomized (n = 22); no non-RBT control group (n = 22); no falls or reactive balance outcome (n = 20); see Figure 1). Disagreements regarding inclusion or exclusion were discussed and resolved between the two reviewers. Authors of publications that met all but the last inclusion criterion (reported reactive balance measurements or fall data) were contacted to determine if these data existed. We used Endnote (X5.0.1, Thomson Reuters, Toronto, ON, Canada) for bibliographic reference management. PRISMA flow diagram of study selection. Data Collection Process and Data Items The following details were extracted: population of the study; RBT and control interventions (number, duration, and frequency of the sessions; type, intensity, and number of perturbations; or description of the control intervention; and program duration), number of participants in each group; reactive balance outcomes pre- and post-intervention, and data on falls in daily life post-randomization or post-training. If additional information was needed (e.g., when relevant data were referenced but not fully reported, or when clarification was required to interpret the published results), it was obtained through secondary publications or by contacting the corresponding authors. Data were compiled into a Microsoft Excel® spreadsheet. All RBT types were included, including slip devices, treadmills, moving platforms, treadmills on tilting platforms, treadmills on translating platforms, waist-pull systems, manual perturbations, trip-over-obstacle and slip-and-trip walkways. Similar training types were grouped into subgroups (surface translation, manual, overground, and waist-pull). Due to variability in perturbation methods, intensity of training was not consistently defined across studies [7]. Therefore, we classified relative perturbation intensity as follows:Low – the goal was for participants to recover balance without stepping, or participants did not need to take a reactive step more than half of the time;Moderate – the goal was for participants to recover balance with just one reactive step, or participants took a single reactive step more than half of the time;High – the goal was for participants to recover balance with more than one reactive step, or participants took more than one reactive step more than half of the time; Very high – the goal was for participants to fall, or participants fell more than half of the time.Absolute – if all participants experienced the same perturbation magnitude (i.e., the intensity was not relative to participants’ abilities).Subjective – if participants’ perceived difficulty of the perturbations was used to determine the intensity. Low – the goal was for participants to recover balance without stepping, or participants did not need to take a reactive step more than half of the time; Low – the goal was for participants to recover balance without stepping, or participants did not need to take a reactive step more than half of the time; Moderate – the goal was for participants to recover balance with just one reactive step, or participants took a single reactive step more than half of the time; Moderate – the goal was for participants to recover balance with just one reactive step, or participants took a single reactive step more than half of the time; High – the goal was for participants to recover balance with more than one reactive step, or participants took more than one reactive step more than half of the time; High – the goal was for participants to recover balance with more than one reactive step, or participants took more than one reactive step more than half of the time; Very high – the goal was for participants to fall, or participants fell more than half of the time. Very high – the goal was for participants to fall, or participants fell more than half of the time. Absolute – if all participants experienced the same perturbation magnitude (i.e., the intensity was not relative to participants’ abilities). Absolute – if all participants experienced the same perturbation magnitude (i.e., the intensity was not relative to participants’ abilities). Subjective – if participants’ perceived difficulty of the perturbations was used to determine the intensity. Subjective – if participants’ perceived difficulty of the perturbations was used to determine the intensity. Additionally, perturbation direction (backward, forward, single axis, or multidirectional) and posture during the intervention (walking, standing, or various postures) were extracted. Study and Reporting Risk of Bias Assessment and Certainty Assessment A Risk of bias (RoB) evaluation was performed by two reviewers using the revised Cochrane Risk of Bias (RoB2) Tool [10]. The assessment has five domains (risk of bias): (1) arising from the randomization process, (2) due to deviations from the intended interventions, (3) due to missing outcome data, (4) in measurement of the outcome, and (5) in selection of the reported results [10]. The assessment rated the studies as low risk (green)/some concerns (yellow)/high risk (red). Any disagreement between assessors was discussed and resolved. Reporting risk of bias results was through a traffic light plot of the RCTs; the plot illustrates the risk of bias across all five domains along with the overall risk of bias assessment for each study. Studies judged to be at high risk of bias in one or more domains were assigned an overall judgment of high risk of bias. An assessment of certainty of evidence [11, 12] was performed considering five domains of limitations (risk of bias, inconsistency, indirectness, imprecision, other considerations). Studies were graded by two reviewers, and limitations were labeled “not serious/serious/very serious” for each domain. The overall assessment (high, moderate, low, and very low certainty of evidence) was determined based on the confidence in the estimated effects. Synthesis Methods To explore the influence of different study components on results of the meta-analyses, exploratory meta-regressions were performed using a restricted maximum likelihood estimation mixed effects random intercepts model for standard mean difference meta-analysis and a minimum variance quadratic unbiased estimation mixed model with random effects for log rate ratio meta-analysis. Perturbation type (waist-pull, manual, overground, and surface translation), intensity (subjective, absolute, low, moderate or high), posture (various, standing, and walking), number of perturbations (log-converted continuous variable), number of sessions (continuous variable), and number of directions (1, 2, 4, or 6) were independent variables in the meta-regression. Multivariable models were fitted using backward variable selection methods; however, these models failed to achieve correctly inverted Hessian matrix and/or G-matrix was not definite positive and univariable meta-regressions were then performed for each identified study component. Multicollinearity between total number of perturbations and program duration was assessed. These variables were moderately correlated (r = −0.71), but variance inflation factors were low (VIF = 2.01), indicating no severe multicollinearity. Given the conceptual overlap, only one variable was retained in the model. Total number of perturbations was selected as it more accurately reflects training dose and is more valid and consistent as a measure of training exposure. SAS (Version 9.4, SAS Institute, NC, USA) for Windows was used. Results Study Selection A total of 7,677 records were identified from database searching. After removing duplicates (n = 2,878), 5,085 records were screened for eligibility. A total of 32 studies were included in the review; 21 additional reports of these studies, such as protocols and secondary publications, were also used in the review to provide details not included in the primary report (Fig. 1). Note that one study was available as an online-only publication at the time of our search and was included in the review despite a 2024 print date [13]. Details for all reports are provided in online supplementary materials 2 and 3. Study Characteristics The characteristics of the studies are presented in Table 1. Samples included in this review were as follows: healthy older adults [13–25], older adults with mixed diagnoses [26–30], people with Parkinson’s disease [31–37], people with chronic [38–40] and sub-acute [41] stroke, people with multiple sclerosis [42, 43], and people with incomplete spinal cord injury [44]. Twenty studies used surface translation perturbations, 5 used manual perturbations, 5 used overground perturbations (slip-and-trip walkway), and 2 studies used waist-pull perturbations. Perturbation directions varied from a single direction (i.e., backward [n = 4] or forward [n = 1] only), a single axis (antero-posterior [n = 13] or mediolateral [n = 1]), and multidirectional (n = 13). Perturbations were delivered during walking in 13 studies, standing in 5 studies, and various postures in 9 studies. Perturbation intensities were low (n = 1), moderate (n = 11), high (n = 5), absolute (n = 7), and subjective (n = 4). Total training duration (calculated as session duration × frequency × program duration in weeks) ranged from 15 to 1,800 min (mean: 573 min, standard deviation: 484 min), and the total number of perturbations was 16–1440 (mean: 539, standard deviation: 493). Table 2 presents the individual results of reactive balance control outcomes from each study. Table 3 summarizes the rate ratios of falls in daily life as reported in the studies. Table 1.Study characteristics of included papersStudyCountrySampleReactive balance training detailsDuration, min (program)Number of perturbationsType of control groupReactive balance control outcomeFall ratetypedirectionpostureintensityShimada et al. [30] (2004)JapanOlder adults, mixed diagnosesSurface translationBackwardWalkingAbsolute600Usual care–YesProtas et al. [31] (2005)USAParkinson’s diseaseSurface translationMultidirectionalStandingHigh1,440720No intervention–YesMarigold et al. [40] (2005)CanadaChronic strokeManualMultidirectionalStandingModeratea1,800Stretching/weight-shifting exercisesIn-laboratory fallsYesBieryla et al. [14] (2007)USAHealthy older adultsOvergroundForwardWalkingModerate1520Treadmill walkingMaximum trunk angle following trip during walkingNoMaki et al. [28] (2008)CanadaOlder adults, mixed diagnosesSurface translationMultidirectionalVariousModerate540864Voluntary stepping and grasping movements–YesSmania et al. [36] (2010)ItalyParkinson’s diseaseManualMultidirectionalVariousLow1,050Joint mobilization, muscle stretching, and motor coordination exercisesUPDRS total scoreYesMansfield et al. [18] (2010)CanadaHealthy older adultsSurface translationMultidirectionalVariousModerate5401,008Flexibility exercises and guided meditationFrequency of multi-step reactionsYesParijat et al. [21] (2012)USAHealthy older adultsOvergroundBackwardWalkingAbsolute37.524Unperturbed walkingIn-laboratory fallsNoShen et al. [34] (2012)ChinaParkinson’s diseaseSurface translationMultidirectionalStandingAbsolute720480Resistance trainingUPDRS-PG (posture and gait sub-score)NoLurie et al. [26] (2013)USAOlder adults, mixed diagnosesSurface translationAntero-posteriorWalkingUsual care (balance and gait training)–YesPai et al. [29] (2014)USAOlder adults, mixed diagnosesOvergroundBackwardWalkingAbsolute24No intervention–YesSchlenstedt et al. [33] (2015)GermanyParkinson’s diseaseManualMultidirectionalVarious840Resistance trainingFullerton Advanced Balance scaleNoShen et al. [35] (2015)ChinaParkinson’s diseaseSurface translationAntero-posteriorWalkingModerate1,440Resistance trainingLatency of balance reactionsYesKurz et al. [17] (2016)IsraelHealthy older adultsSurface translationMulti-directionalWalkingModerate4801,100Treadmill walkingPerformance-Oriented Mobiity Assessment – balance sub-scoreYesSteib et al. [37] (2017)GermanyParkinson’s diseaseSurface translationAntero-posteriorWalking640Treadmill walkingMini-BEST reactive sub-scoreNoMansfield et al. [39] (2018)CanadaChronic strokeManualMulti-DirectionalVariousHigh720720Conventional balance trainingMini-BEST reactive sub-scoreYesAviles et al. [13] (2019)USAHealthy older adultsSurface translationAntero-posteriorStandingHigh360480Tai ChiCustom reactive balance ratingYesHandelzalts et al. [41] (2019)IsraelSub-acute strokeSurface translationAntero-posteriorVariousHigh360432Weight shifting and gait trainingFall threshold (in-lab)NoOkubo et al. [20] (2019)AustraliaHealthy older adultsOvergroundAntero-posteriorWalkingSubjective12060Overground walkingIn-lab fallsNoEsmaeili et al. [38] (2020)CanadaChronic strokeSurface translationAntero-posteriorWalkingModerate315–630a234Treadmill walkingMini-BESTNoLurie et al. [27] (2020)USAOlder adults, mixed diagnosesSurface translationAntero-posteriorWalking562.5Usual care (balance and gait training)–YesRieger et al. [22] (2020)NetherlandsHealthy older adultsSurface translationAntero-posteriorWalkingAbsolute8a16aTreadmill walkingDeviation in trunk velocity from unperturbed walkingNoRogers et al. [24] (2021)USAHealthy older adultsWaist pullMedio-lateralStandingModerate1,440Hip strengtheningMulti-step thresholdYesUnger et al. [44] (2021)CanadaIncomplete spinal cord injuryManualMulti-directionalVariousModerate1,4401,440“Conventional” intense balance trainingBehavioural response score following perturbationsYesMonjezi et al. [42] (2022)IranMultiple sclerosisWaist pullBackwardVariousHigh240600Standing balance exercises and Treadmill walkingLatency of postural responsesYesWang et al. [25] (2022)USAHealthy older adultsSurface translationAntero-posteriorWalkingAbsolute3040Treadmill walking–YesBrull et al. [15] (2023)GermanyHealthy older adultsSurface translationMulti-directionalWalkingModerate4321,296No interventionStepping Threshold TestNode Souza et al. [32] (2023)BrazilParkinson’s diseaseSurface translationMulti-directionalStandingAbsolute4401,024Resistance trainingIn-lab near fallsNoGerards et al. [16] (2023)NetherlandsHealthy older adultsSurface translationMulti-directionalVariousSubjective90Usual careMini-BEST reactive sub-scoreNoNørgaard et al. [19] (2023)DenmarkHealthy older adultsSurface translationAntero-posteriorWalkingSubjective4080Treadmill walkingIn-lab fallsYesOkubo et al. [43] (2023)AustraliaMultiple SclerosisOvergroundAntero-posteriorWalkingSubjective10048Sham obstacle avoidance trainingIn-lab fallsNoRieger et al. [23] (2024)NetherlandsHealthy older adultsSurface translationAntero-posteriorWalkingModerate240256Treadmill walking with cognitive taskMini-BEST reactive sub-scoreYes aDetails obtained from correspondence with study authors.BEST, Balance Evaluation Systems Test; UPDRS, Unified Parkinson’s Disease Rating Scale. Study characteristics of included papers aDetails obtained from correspondence with study authors. BEST, Balance Evaluation Systems Test; UPDRS, Unified Parkinson’s Disease Rating Scale. Table 2.Results of studies reporting on reactive balance control outcomesStudyOutcomeRBT groupControl group n pre-trainingpost-training n pre-trainingpost-trainingMarigold et al. [40] (2005)Falls following in-laboratory perturbations (number)261.6 (3.1)1.0 (2.5)220.8 (2.1)1.2 (2.7)Bieryla et al. [14] (2007)aMaximum trunk angle following trip during overground walking (degrees)646.5 (26.2)37.7 (22.1)535.9 (16.2)43.1 (13.8)Mansfield et al. [18] (2010)Frequency of multi-step reactions in response to waist-pull perturbations (% of trials)1644.8 (49.8)32.9 (47.0)1456.0 (49.7)40.5 (49.2)Smania et al. [36] (2010)UPDRS total score2846.1 (11.5)43.5 (14.6)2743 (16.9)43.9 (18.3)Parijat et al. [21] (2012)Falls following in-laboratory slips (% of participants)12420125025Shen et al. [34] (2012)UPDRS-PG score144.0 (1.2)2.7 (1.4)144.2 (1.5)3.3 (1.9)Schlenstedt et al. [33] (2015)Fullerton Advanced Balance scale score1524.5 (4.6)24.9 (5.3)1722.2 (4.8)24.5 (5.4)Shen et al. [35] (2015)Pre- to post-training change in latency of balance reactions to external perturbation (ms)22​−9 (14)23​2 (9)Kurz et al. [17] (2016)Performance-Oriented Mobility Assessment balance sub-score2214.9 (1.3)15.4 (0.9)2014.8 (1.4)15.3 (1.0)Steib et al. [37] (2017)Mini-BEST reactive sub-score184.8 (1.1)5.2 (0.8)205.2 (1.1)5.0 (1.3)Mansfield et al. [39] (2018)aPre- to post-training change in mini-BEST reactive sub-score39​1.5 (1.9)38​0.8 (1.4)Aviles et al. [13] (2019)aCustom reactive balance rating based on behavioral responses to treadmill perturbations164.8 (3.5)8.2 (3.2)164.3 (2.0)6.3 (2.4)Handelzalts et al. [41] (2019)Fall threshold for in-laboratory perturbations (intensity level)164.5 (2.0)6.5 (1.3)163.5 (2.4)4.6 (2.3)Okubo et al. [20] (2019)Fall rate following overground slips and trips in the laboratory (proportion of trials)19​0.1722​0.44Esmaeili et al. [38] (2020)Mini-BEST total score1020.0 (2.1)23.0 (1.9)816.5 (4.7)17.0 (2.4)Rieger et al. [22] (2020)Deviation in trunk velocity following mediolateral perturbations to walking (sum of deviations)152,578 (834)2,055 (622)152,310 (594)1,940 (534)Rogers et al. [24] (2021)Multi-step threshold (intensity level)144.8 (1.0)6.1 (0.8)154.5 (1.0)4.8 (1.0)Unger et al. [44] (2021)Pre- to post-training change in behavioral response score following lean-and-release perturbations10​0.3 (0.4)10​0.1 (0.4)Monjezi et al. [42] (2022)Latency of postural responses to external perturbations (ms)17216 (25)200 (31)17205 (45)203 (42)Brull et al. [15] (2023)Stepping threshold test (score)2324.2 (6.1)33.5 (5.0)2124.8 (3.4)27.1 (5.5)de Souza et al. [32] (2023)In-laboratory near falls (proportion of participants)90.670.22100.600.13Gerards et al. [16] (2023)Mini-BEST reactive sub-score344.0 (1.5)5.0 (1.5)375.0 (1.5)5.0 (1.5)Nørgaard et al. [19] (2023)Falls following in-laboratory perturbations (proportion of trials)630.69 (0.77)0.14 (0.36)680.73 (0.84)0.69 (0.84)Okubo et al. [43] (2023)In-laboratory falls (proportion of trials)14​0.2216​0.42Rieger et al. [23] (2024)Mini-BEST reactive sub-score332.3 (1.8)2.7 (1.3)342.5 (1.5)3.4 (1.5)Values presented are means with standard deviations in parentheses. aData obtained from correspondence with the study authors.RBT, reactive balance training; BEST, Balance Evaluation Systems Test; UPDRS, Unified Parkinson’s Disease Rating Scale; PG, Posture and Gait sub-score; ms, milliseconds. Results of studies reporting on reactive balance control outcomes Values presented are means with standard deviations in parentheses. aData obtained from correspondence with the study authors. RBT, reactive balance training; BEST, Balance Evaluation Systems Test; UPDRS, Unified Parkinson’s Disease Rating Scale; PG, Posture and Gait sub-score; ms, milliseconds. Table 3.Results of studies reporting falls in daily lifeStudyFalls collectedRBT groupControl groupRate ratio95% CI n number of falls n number of fallsShimada et al. [30] (2004)Post-randomization15811110.5330.21–1.33Marigold et al. [40] (2005)Post-randomization192521750.3680.23–0.58Protas et al. [31] (2005)aPost-training9109230.4350.21–0.91Maki et al. [28] (2008)aPost-training44441.0000.25–4.00Mansfield et al. [18] (2010)aPost-training1661541.4060.40–4.98Smania et al. [36] (2010)Post-training2836271110.3130.21–0.46Lurie et al. [26] (2013)Post-randomization261033320.3970.20–0.81Pai et al. [29] (2014)Post-training1021796320.5000.28–0.90Shen et al. [35] (2015)Post-randomization221023180.5810.27–1.26Kurz et al. [17] (2016)aPost-training211219180.6030.29–1.25Mansfield et al. [39] (2018)Post-training415342640.8480.59–1.22Aviles et al. [13] (2019)Post-training1641650.8000.22–2.98Lurie et al. [27] (2020)Post-randomization2186862126031.1060.991–1.235Rogers et al. [24] (2021)Post-training1871541.4440.425–4.910Unger et al. [44] (2021)Post-training101010210.4760.224–1.010Monjezi et al. [42] (2022)Post-training174617640.7190.493–1.049Wang et al. [25] (2022)Post-training703763281.1890.729–1.940Nørgaard et al. [19] (2023)Post-trainingb636268780.7800.559–1.089Rieger et al. [23] (2024)Post-training331034130.7930.347–1.808Rate ratios <1 indicate that the RBT group reported fewer falls than the control group.RBT, reactive balance training; CI, confidence interval. aData obtained from correspondence with the study authors. bFalls data were collected starting from the 3rd of 4 training sessions. Results of studies reporting falls in daily life Rate ratios <1 indicate that the RBT group reported fewer falls than the control group. RBT, reactive balance training; CI, confidence interval. aData obtained from correspondence with the study authors. bFalls data were collected starting from the 3rd of 4 training sessions. Risk of Bias Figure 2 presents the risk of bias for studies reporting reactive balance outcomes (n = 25). Six studies were rated as high risk, 14 had some concerns, and 5 were assessed as low risk of bias. The main concerns arose from the selection of reported results (20 of 25) and deviations from the intended interventions (8 of 25). In contrast, most studies showed low risk of bias in randomization (19 of 25), missing data (24 of 25), and outcome measurement (21 of 25). Risk of bias for studies reporting a reactive balance control outcome. Figure 3 presents the risk of bias for the 19 studies reporting falls in daily life. Nine studies were rated as high risk, eight had some concerns, and only two were assessed as low risk. These findings primarily reflect high or some concerns in the randomization process (7 of 19 studies), outcome measurement (12 of 19), and selection of reported results (15 of 19). In contrast, most studies showed low risk of bias for missing data (18 of 19) and deviations from intended interventions (15 of 19). Risk of bias for studies reporting falls in daily life. Meta-Regression Table 4 describes the meta-regression analysis of the influence of each training parameter on reactive balance outcomes and falls in daily life. RBT programs that included manual perturbations were associated with reduced falls in daily life compared to the reference (waist-pull perturbations; rate ratio: 0.45; 95% confidence interval: 0.22, 0.91, p = 0.042). There were no other significant relationships between any other training parameters and falls in daily life or reactive balance control outcomes. Table 4.Univariable meta-regressions for intervention componentsFalls in daily lifeReactive balance outcomes n rate ratiop value n estimatep valuePerturbation type Waist-pull2Reference2Reference Manual20.45 [0.22, 0.91]0.0425−0.71 [−2.38, 0.96]0.42 Overground10.48 [0.18, 1.31]0.174−0.061 [−1.73, 1.61]0.94 Surface translation70.71 [0.38, 1.33]0.3014−0.11 [−1.50, 1.29]0.88Intensity category Subjective1Reference4Reference Absolute30.84 [0.24, 2.96]0.794−0.51 [−2.15, 1.14]0.51 Low10.40 [0.08, 1.93]0.271−0.62 [−2.74, 1.51]0.58 Moderate81.03 [0.32, 3.32]0.97100.19 [−0.98, 1.35]0.7 High40.86 [0.25, 2.93]0.8140.085 [−1.37, 1.54]0.91Posture Various6Reference8Reference Standing40.93 [0.51, 1.69]0.8150.38 [−0.78, 1.53]0.53 Walking90.87 [0.52, 1.44]0.60120.45 [−0.47, 1.37]0.35Number of perturbations Number of perturbations131.04 [0.87, 1.25]0.66190.12 [−0.23, 0.46]0.52Number of directions 13Reference3Reference 281.43 [0.70, 2.94]0.34120.64 [−0.73, 2.01]0.38 481.04 [0.51, 2.14]0.9160.14 [−1.31, 1.59]0.85 60––6−0.22 [−2.40, 1.96]0.84Quality of the evidence (GRADE)⊕⊝⊝⊝ very low1,2,3⊕⊕⊝⊝ low2,3Values presented are estimates (rate ratio for falls in daily life, slope for reactive balance outcomes) with confidence intervals in brackets and associated p values; n is the number of studies included in each category. 1Downgraded due to high risk of bias or some concerns due to selection of reported results and measurement of the outcome. 2Downgraded due to indirectness – there are no head-to-head comparisons for any training parameters, so we only have indirect comparisons. 3Downgraded due to some imprecision as the confidence intervals (CI) are wide, suggesting evidence for both harm and benefit, for almost all comparisons. The only comparison with a narrow CI is for manual perturbations vs. waist-pull perturbations and fall risk; however, the sample size for this comparison is low. Univariable meta-regressions for intervention components Values presented are estimates (rate ratio for falls in daily life, slope for reactive balance outcomes) with confidence intervals in brackets and associated p values; n is the number of studies included in each category. 1Downgraded due to high risk of bias or some concerns due to selection of reported results and measurement of the outcome. 2Downgraded due to indirectness – there are no head-to-head comparisons for any training parameters, so we only have indirect comparisons. 3Downgraded due to some imprecision as the confidence intervals (CI) are wide, suggesting evidence for both harm and benefit, for almost all comparisons. The only comparison with a narrow CI is for manual perturbations vs. waist-pull perturbations and fall risk; however, the sample size for this comparison is low. GRADE Quality of Evidence The evidence quality for reactive balance outcome and fall in daily life outcome were rated as low and very low, respectively, according to the GRADE approach. The studies included in this meta-regression are RCTs, which provide a baseline rating of high-quality evidence. However, there were some limitations that downgraded the quality of the evidence (Table 4). Downgrading for risk of bias was applied due to high risk or some concerns related to selective reporting and outcome measurement for falls in daily life. The certainty of evidence was downgraded for indirectness, as no direct head-to-head comparisons were available for any of the training parameters, resulting in reliance on indirect comparisons for both outcomes. Downgrading for imprecision was also applied, as confidence intervals were wide for almost all comparisons, indicating effects compatible with both benefit and harm for both variables. The only comparison with a narrow confidence interval was between manual perturbations and waist-pull perturbations for fall risk; however, this estimate was based on a small sample size. Discussion RBT has been widely shown to improve reactive balance control [45] and reduce falls in daily life [7]. This review aimed to determine the optimal RBT intervention characteristics for improving reactive balance control and preventing falls in daily life. However, we were unable to identify key training program characteristics that contribute to improved training effects. Our analysis suggests that RBT using manual perturbations may be superior to waist-pull perturbations for preventing falls in daily life. Among people with significant balance impairments, such as those with chronic stroke [46] or long-term care residents [47], a large number of falls occur due to “incorrect weight transfer” (i.e., loss of balance during voluntary movement), rather than due to external forces or environmental hazards (e.g., slips or trips). Manual RBT includes both external and internal perturbations (i.e., “incorrect weight transfers”); if this training mimics how people typically lose balance prior to a fall in daily life, this may explain the apparent benefits of manual RBT for preventing falls. However, it is possible that other differences between studies account for the apparent significant reduction in fall rates for manual RBT. For example, all manual RBT studies included people with neurological conditions (stroke, Parkinson’s disease, and spinal cord injury); it is possible that large training effects for falls in daily life were observed for these studies due to the high baseline fall rate among these groups. When exploring evidence from single comparative studies, different types of perturbations may influence fall prevention outcomes differently [16, 48]; however, these effects were not detected in our analysis. Moreover, some studies reported no effect of one perturbation type over another [16, 49], similar to our analysis. No studies in our review compared different RBT intensities, and few studies [14, 18, 19, 22, 23, 25, 28–30, 32, 41] reported the intensity of training. Two previous studies that included healthy young adults found that higher intensity RBT led to greater improvements in reactive balance control than lower intensity training [50, 51]. Conversely, a more recent study comparing different RBT intensity schedules among healthy young adults found no difference in training effect between the low- to high-intensity training schedule and the group that always completed RBT at a high intensity [52]. For other types of exercise and components of fitness, shorter high-intensity training sessions are often found to lead to similar or improved outcomes compared to longer sessions of moderate intensity training [53, 54]. Future studies should compare RBT of different intensities to determine if the same is true for RBT. In our analysis, we expected that a higher training volume (i.e., more perturbations) would lead to better outcomes. This expectation is based on previous studies showing that more practice generally leads to better outcomes [55]. However, some RBT studies have demonstrated significant effects with 1–3 relatively short training sessions [21, 23, 56, 57]. While these studies used absolute (all participants experienced the same intensity) and subjective RBT intensities, the intensity of training in these studies may have been high to very high (i.e., evident by reports of falls into the safety harness in these studies). Therefore, high-intensity training may explain the significant training effects with a short duration of training. This hypothesis will need to be tested in future studies. We anticipated that a more variable practice, incorporating a variety of postures, directions, and fall risk situations, would be more effective for improving reactive balance control and reducing fall rates [58]. However, we did not observe greater training effects for programs that included greater variability in postures or directions. Walking is one of the most frequent activities at the time of a fall [39, 47]. However, in our study, training programs that focused on walking were not more effective than programs that provided perturbations to stance only or that included various postures. The presence of evidence supporting opposing outcomes may explain the difficulty in establishing the best characteristics for RBT interventions. Our findings are similar to those of Farlie et al. [59], who conducted a meta-analysis of general balance training programs, and found no association between the components of the “FITT” principle (frequency, intensity, type, and time) and improved balance outcomes following balance training. Similar to this previous analysis, we also found that no evidence that specific RBT program characteristics led to superior training effects. Farlie et al. [59] speculated that their findings may be partially explained by incomplete reporting of training characteristics in published studies, which may also be the case for our review. Lack of reporting key study characteristics also reduced the available sample size for our analyses [59]. Less than half of the studies (14/32) reported all training characteristics in a way that could be meaningfully incorporated into meta-analysis; specifically, progression of perturbation intensity or number, or complexity of training, was not consistently applied or reported across all included studies. Due to this small sample size and high variability in study designs (e.g., in participant characteristics, control group interventions, and reactive balance measurements), we were unable to conduct multivariable meta-regression to examine the effect of multiple training characteristics in combination. The Template for Intervention Description and Replication (TIDieR) [60] and Consensus on Exercise Reporting Template (CERT; [61]) are guidelines for reporting intervention characteristics in clinical trials. Theses reporting guidelines should be used by authors of RBT studies to ensure that intervention characteristics are reported fully, and use of these guidelines should be enforced by journal reviewers and editors. More consistent use of these reporting guidelines will not only facilitate meta-analyses, like ours, but also establishing clinical practice guidelines and facilitating translation of effective interventions into clinical practice. This review included a wide range of populations, intervention designs, with different types of training methods, substantial variability in training parameters, and diverse outcome measures; however, despite its comprehensive scope, definitive conclusions could not be drawn due to this variability. Despite including only RCTs, the quality of evidence for reactive balance and falls in daily life was low to very low due to wide confidence intervals, lack of direct comparisons, high risk of bias, and the dominance of a few larger studies [13, 17, 20, 25, 27, 29, 39] that may have potentially impacted the results, posing an additional limitation, as the true effect may have been masked. Another limitation of this review is that, in all included trials, reactive balance was incorporated as part of the assessment protocol. Given that improvements in reactive balance may occur even after a single session [56], the assessment itself may have influenced outcomes in the control groups, thereby potentially reducing the differences between intervention and control conditions. Additionally, included studies did not consistently report intensity with clearly defined thresholds (e.g., 50%), which may have affected the precision of our intensity categorizations. Additional limitations include the possibility that relevant trials published after the final search date were not included, as well as potential language bias due to the exclusion of studies not published in English. We suggest that future studies develop training programs based on existing training protocols, with appropriate modifications. Given the wide variety of training programs available, it is important to design studies that rely on existing research to validate their effects; while some studies have directly compared different training parameters [16, 23, 49, 51, 54, 57], further head-to-head comparisons are needed. There is a need to conduct an in-depth investigation into the guiding principles for designing RBT, for example, identifying the most effective perturbation parameters for fall prevention or determining the optimal intervention duration for improving balance reactions. Future studies should provide more detailed reports of the training protocols. Statement of Ethics A Statement of Ethics is not applicable because this study is based exclusively on published literature. Conflict of Interest Statement The authors have no conflicts of interest to declare. Funding Sources This study was not supported by any sponsor or funder. Author Contributions A.M. conceived the concept for this review. H.N. and A.M. equally contributed to the literature searches, article selection, and drafting of the manuscript. L.K.L. was responsible for the statistical analysis and data interpretation. A.J.D. contributed to the literature searches. All authors actively participated in writing the original drafts, as well as reviewing and editing the final manuscript. All authors have read and approved the published version of the manuscript.