Effects of hydrokinesitherapy on balance and walking ability for stroke survivors: update of a systematic review and meta-analysis of randomized controlled studies

Abstract

BACKGROUND

Balance dysfunction often affects stroke patients. While our earlier research indicated that aquatic exercise therapy could help, the evidence was initially rated as very low. With new studies emerging, this research aims to update the evidence on the effectiveness of water-based exercise in improving balance for stroke survivors.

METHODS

The study followed PRISMA guidelines, retrieving records from five databases (Medline, Embase, CENTRAL, CINAHL, and SPORTDiscus) up to July 1, 2025. Two researchers independently conducted a two-step screening of titles, abstracts, and full texts to select studies. Quality was assessed using the ROB2 tool, and data extraction followed a set protocol. All steps were independently performed by two authors, with a third resolving any disagreements. Data extraction followed a set protocol, with two authors independently handling literature screening, bias assessment, and data extraction. Discrepancies were resolved by consulting a third party. Data analysis was conducted using RevMan 5.4 and Stata, calculating combined effect sizes using either SMD or MD based on the Q test results, and choosing between random- or fixed-effects models. Sensitivity analysis was done with Stata 15.0, and publication bias was evaluated using funnel plots and Egger’s test. Evidence quality was assessed with the GRADE system.

RESULTS

A total of 29 RCTs involving 1075 stroke patients were included in this study. The results indicated that the water-based exercise can improve the balance function of stroke survivors based on BBS score with moderate effect size (N.=523, SMD=0.71, 95% CI: 0.28-1.14, P=0.001) and FRT(n=107, SMD=0.97, 95% CI: 0.32-1.61, P=0.003), while test time for TUGT in the intervention group was not improved (SMD=-0.46, 95% CI: -0.96 to 0.04, P=0.07). The quality of the evidence for BBS was assessed as moderate, and the quality of evidence for TUGT and MBI was assessed as very low.

CONCLUSIONS

Moderate strength of evidence suggests that aquatic exercise therapy may improve the balance function of stroke patients. However, due to methodological limitations, inconsistent results, and clinical heterogeneity, this result should be interpreted with caution. Future research employing more rigorous study designs is necessary to substantiate these results.

Introduction

Stroke is a major global public health burden, being one of the leading causes of long-term disability and mortality. Clinically, it is characterized by high incidence, high disability rate, and high recurrence.¹ Stroke survivors often develop a range of post-stroke sequelae, including impaired balance function, gait dysfunction, reduced muscle strength, and poor motor coordination.^{2, 3} Approximately 50% of stroke survivors suffer from persistent motor impairments, which increase fall risk and severely impair their ability to perform activities of daily living (ADLs) independently, thereby reducing their quality of life.^{4, 5}

Balance relies on sensory inputs from the visual, vestibular, and somatosensory systems, which are integrated by the central nervous system to regulate motor responses for postural control. Impaired balance is highly prevalent in stroke survivors,^{6, 7} primarily due to disrupted sensory-motor networks essential for maintaining balance. It severely compromises patients’ ability to perform ADLs and sustain independent living safely. Falls are a common serious complication after stroke, with around 70% experiencing at least one fall within six months post-discharge, especially in the early recovery period.^8-12 Stroke survivors who fall have a twofold higher risk of hip fracture than non-stroke individuals.^13-15 Moreover, balance deficits reduce confidence and physical activity levels,¹⁶ potentially leading to severe adverse outcomes if unaddressed.^{17, 18}

While balance training benefits stroke survivors, no single optimal approach has been validated.¹⁹ Emerging modalities, including virtual reality and rehabilitation robotics, integrated within balance training programs, have shown efficacy for post-stroke,^{20, 21} but are limited by substantial costs and technical requirements.^{22, 23}

Aquatic therapy utilizes water’s unique properties to facilitate patient exercise, aiming to alleviate symptoms and improve functional capacity.²⁴ It improves post-stroke balance function through multiple mechanisms:²⁵ buoyancy reduces body weight and alleviates joint stress, enabling easier balance training than land-based training;²⁶ water viscosity offers natural resistance to strengthen muscles and enhance neuromuscular control;²⁷ water flow a stimulate proprioceptors to enhance tactile feedback for better posture control;²⁸ the safe aquatic environment reduces fall risk, fostering patient engagement;²⁵ and a water temperature of 30-34 °C aids muscle relaxation, and training comfort.^{29, 30} This, in turn, enables patients to more effectively engage in balance-focused training programs.

A 2011 Cochrane review found insufficient evidence for its benefit on ADLs,^{31, 32} while our prior study indicated potential benefits for balance and mobility but with low-quality evidence due to limited sample sizes and poor quality.³³ Additionally, the most recent meta-analysis, published in 2021, suggests that aquatic exercise is more effective than no intervention or land-based rehabilitation²⁴ but included substantial clinical limitations (e.g., non-fully randomized controlled trials [RCTs], passive water interventions instead of active aquatic exercises.^{34, 35} Since the 2021 meta-analysis, additional studies have emerged,^{30, 36-39} highlighting the need to update the evidence by incorporating these recent findings to provide more current, reliable clinical guidance for the application of aquatic therapy in post-stroke balance rehabilitation. Specifically, this update expands the evidence base by including more RCTs and a larger sample size, which enhances the statistical power of the results and conducts a more detailed exploration of clinical heterogeneity, including subgroup analyses and meta-regression to identify potential moderators, addressing the limitations of insufficient heterogeneity exploration in prior studies.

Methods

Trial register

The methodology of this study was rigorously executed in accordance with the guidelines outlined in the Cochrane Handbook for Systematic Reviews of Interventions and the prospectively developed protocol.^{33, 40} The study protocol was registered prospectively, and the registration details are available on the PROSPERO website (ID: CRD42018110787). Furthermore, the complete protocol has been published in a peer-reviewed academic journal.⁴¹

Data sources

Records were conducted in line with previously reported methods. A comprehensive search of five electronic databases (Medline, Embase, CENTRAL, CINAHL, and SPORTDiscus) was performed according to the pre-established search strategy,³³ updated to July 1, 2025, without language restrictions.

Search strategy

The literature retrieval process was executed utilizing the finalized search strategy, which had been validated through pilot searches and repeated retrieval validations. This search strategy was meticulously developed based on the foundational principles of the PIS framework. Within this framework, P (Population) encompasses stroke-related population terms such as “stroke” and “cerebrovascular accident.” I (Intervention) include terms related to aquatic exercise interventions, such as “aquatic exercise therapy” and “water-based exercise”. S (Study Design) specifies RCTs as the eligible study type, incorporating terms like “randomized controlled trial”, “random allocation”, and “randomised trial” to ensure the inclusion of only RCTs. These terms, representing the P, I, and S categories, were systematically combined using Boolean logical operators (AND/OR/NOT). Additionally, both controlled vocabulary terms (e.g., MeSH terms, Emtree terms) and free-text terms were utilized. This approach maximizes the coverage of relevant literature and minimizes retrieval bias. The comprehensive search strategy, encompassing detailed database-specific adaptations and term lists, is provided in Supplementary Digital Material 1, Supplementary Text File 1 of this manuscript. Alternatively, it can be accessed in our previously published article.³³

Study eligibility criteria

Studies were incorporated into this analysis only if they satisfied the following predefined inclusion criteria: 1) participants: individuals diagnosed with either hemorrhagic or ischemic stroke, with no restrictions imposed on age or gender. Neither the initial severity of functional impairment nor the time elapsed since stroke onset was constrained; 2) intervention: any individual or combined intervention involving structured, standardized, and replicable water-based exercise treatment for participants. Notably, this intervention had to be administered by a healthcare professional with specialized training; interventions conducted in non-therapeutic settings (e.g., spa pools or residential bathtubs) were excluded. No limitations were set on the total duration or weekly frequency of hydrokinesitherapy sessions. 3) Control group: the control arm of the study was required to receive either land-based (non-water-based) exercise interventions or no intervention at all. 4) Outcome measures: no restrictions were placed on the types of outcome measures reported in the studies. 5) Study design: only clinical RCTs were considered eligible for inclusion. Only studies with full-text articles published in the English language were included.

Study selection

All collected literature was systematically imported into the reference management system (EndNote X7.4). Following this, an initial deduplication process was conducted using the software’s automatic duplicate elimination feature. Subsequently, a manual verification of titles, authors, and publication details was performed to remove any remaining duplicate records, ensuring that no redundant literature was retained for the subsequent screening stages. An initial screening of the deduplicated literature was independently conducted by two researchers, utilizing the pre-established PICOS framework. During this phase, studies that clearly did not meet the PICOS criteria were promptly excluded based on their titles and abstracts. Literature with incomplete information or ambiguous details, which precluded a definitive eligibility assessment, was earmarked for full-text evaluation. Following the retrieval of full texts for both the literature that passed the initial screening and the ambiguous records, a comprehensive full-text review was conducted by the two researchers independently to preliminarily identify eligible records. In instances where discrepancies arose between the two researchers (F.Y.L. and L.J.L.) regarding eligibility judgments, a third independent researcher (G.L.X.) was consulted to facilitate discussion and reach a final consensus, through which the list of literature included in the analysis was confirmed.

Data extraction

Data extraction was initially performed by a single researcher (Z.L.Z.) utilizing a pre-constructed Excel spreadsheet in accordance with the PICOS framework, and subsequently verified by a second researcher (G.Y.Y.). The extracted data encompassed the following elements: the first author’s name, year of publication, total sample size, mean age of participants, gender distribution, and detailed characteristics of the intervention measures administered to both the experimental and control groups. These characteristics included type, dosage, intervention duration, primary and secondary outcome indicators, follow-up period, and dropout rates, among others. In RCTs involving multiple groups, data were extracted in accordance with the recommendations of the Cochrane Handbook for Systematic Reviews of Intervention.³² The two independent intervention arms were combined into a single intervention group to avoid the statistical issues associated with reusing the shared control group.³² In instances where essential data were missing or the report lacked sufficient detail, attempts were made to obtain the necessary information by contacting the corresponding author via email. If no response was received from the original author, despite two-week intervals between follow-up attempts, the data were converted in strict accordance with the methodological guidelines outlined in the “Cochrane Handbook for Systematic Reviews” (6.5 edition).⁴² Any discrepancies in data extracted by the two researchers were resolved through consultation with a third experienced researcher before inclusion in data analysis.

Risk of bias assessment

The risk of bias for all RCTs included after PICOS-framework screening was evaluated using the Cochrane Collaboration’s Risk of Bias 2.0 (ROB2) tool by two researchers (F.Y.L. and J.L.L.) with formal ROB2 training, who worked independently and focused on the tool’s five core domains (Randomization process, Deviations from intended interventions, Completeness of outcome data, Objectivity of outcome measurement, Selective reporting) and one optional domain (Other sources of bias); during assessment, researchers answered domain-specific “signaling questions” in ROB2, categorized each domain’s bias risk as “low risk,” “high risk,” or “some concerns” based on study methodologies and results, then cross-verified their findings – resolving discrepancies by revisiting full texts, discussing with reference to the ROB2 Tool Handbook, and consulting a third evidence-based medicine expert (LL) if needed – until consensus was reached.

Data analysis

All data analyses were conducted utilizing Stata version 15.0 or RevMan 5.4 software, with statistical significance for primary outcomes established at a threshold of P<0.05. For continuous outcomes, the pooled effect size was determined using either mean differences (MD) or standardized mean differences (SMD), accompanied by their respective 95% confidence intervals (CIs). Based on Cohen’s d effect size classification criteria, the pooled effect size was interpreted as follows: a SMD of less than 0.2 indicates a negligible effect; an SMD ranging from 0.2 to less than 0.5 denotes a small effect; an SMD ranging from 0.5 to less than 0.8 represents a moderate effect; and an SMD of 0.8 or greater signifies a large effect.⁴² Heterogeneity among the included studies was evaluated using the I² statistic and the Q-test. A significant level of heterogeneity was considered present if the I² value exceeded 50% or if the P value was less than 0.10; under these conditions, a random-effects model was employed for effect size pooling. Conversely, a fixed-effects model was applied when heterogeneity was not significant.

Subgroup analysis

A subgroup analysis was performed according to the pre-established grouping methodology. Within this study, variations in assessment tools were noted for evaluating primary outcome indicators, including balance function and activities of daily living. Specifically, the included studies employed diverse scales such as the Berg Balance Scale (BBS), the Timed Up and Go Test (TUGT), the Functional Reach Test (FRT), and the Five Times Sit-to-Stand Test (FTSST) to measure balance function. Consequently, subgroups were delineated based on the different assessment scales utilized.

Meta-regression analysis

To explore sources of between-study heterogeneity, meta-regression was performed for outcomes with P<0.10 in the Chi-squared Test, I²>75%, and with more than 10 included studies according to the guidelines.⁴³ Pre-specified covariates for the meta-regression were selected based on clinical and methodological rationale, including intervention of the experimental group, intervention duration, type of stroke, duration of stroke, follow-up time, and country. The restricted maximum likelihood (REML) method was employed to estimate the between-study variance (τ²). The proportion of between-study variance explained by the covariates (adjusted R²) was calculated. The Knapp-Hartung modification was applied to adjust for the small number of studies, and the results were considered statistically significant at P<0.05.

Sensitivity analysis

The reliability of the study’s findings was assessed through comprehensive sensitivity analyses. These analyses were systematically conducted at each decision node, including literature search, study inclusion, data extraction, and data analysis, to evaluate the influence of these decisions on the study outcomes. Detailed procedures and corresponding handling methods are documented in Supplementary Digital Material 2, Supplementary Table I. During the selection of statistical methods, either a random-effects model or a fixed-effects model was employed to estimate the overall effect size. Furthermore, utilizing Stata software (Version 15.0), studies were sequentially excluded one by one, followed by re-conducting the meta-analysis to assess the robustness of the results. A significant change in the pooled effect size upon the exclusion of a particular study indicated that the study exerted a substantial influence on the overall findings.

Publication bias

The evaluation of potential publication bias was assessed for primary outcomes that included a minimum of 10 studies or had a total sample size exceeding 250 participants.^{40, 44} Funnel plots were constructed utilizing Stata 15.0, and Egger’s Test was applied to assess asymmetry. A P value of less than 0.05 was considered indicative of potential publication bias.

The quality of evidence

The evidence pertaining to the primary outcome was appraised using the GRADE framework (Grading of Recommendations Assessment, Development, and Evaluation). The quality of evidence was determined by assessing five downgrading criteria (risk of bias, inconsistency, indirectness, imprecision, and publication bias) and three upgrading criteria (large effect size, dose-response gradient, and a low probability of confounding factors).^45-47 Based on these assessments, the evidence was classified of high, moderate, low, or very low quality.

Evidence synthesis

Study identification

As shown in Figure 1, a total of 1236 records were retrieved through systematic searches of five electronic databases (N.=1229) and supplementary sources (N.=7). After removing 267 duplicates, 969 unique records were screened at the title/abstract level. Of these, 924 were excluded due to ineligible patient populations (N.=583), irrelevance to hydrokinetherapy (N.=144), or non-RCT study designs (e.g., reviews, protocols, case reports; N.=197), leaving 45 full-text articles for eligibility assessment. During full-text review, 16 studies were excluded due to non-randomized design (N.=6), ineligible subjects (N.=3), insufficient outcome data (N.=3), inappropriate intervention or comparison (N.=2), non-English publication (N.=1), and unobtainable full text (N.=1). This yielded 29 RCTs for qualitative synthesis. Among these, seven studies were excluded from the quantitative meta-analysis because their outcome measures were reported in fewer than three studies, resulting in a final sample of 22 RCTs for the meta-analysis.

Figure 1.

—Flow diagram of the study selection process.

Study characteristics

As shown in Supplementary Digital Material 3, Supplementary Table II,^{30, 36-39, 48-71} this review included 29 RCTs, comprising 1075 stroke patients, of whom 639 were male, and 434 were female. All included studies were single-center trials, with sample sizes varying between 12 and 160 participants. The average age of participants was 43.8 years to 70 years old. Of the included studies, 16 were conducted in Asia (including 13 in Korea^{49-55, 57, 58, 61, 66, 68, 71} and 3 in China^{39, 62, 67}), 9 in Europe,^{30, 36-38, 48, 63, 64, 69, 70} and others were in Canada, Egypt, Türkiye, or Iran.

In terms of participant settings, 14 trials recruited inpatients,^{30, 39, 50, 52, 53, 55, 57, 58, 63, 65, 66, 68, 71} 5 trials enrolled outpatients,^{51, 60, 61, 69, 71} one trial included both inpatients and outpatients,⁶⁷ and 1 study recruited community-dwelling individuals.⁵⁶ Four trials were conducted in healthcare institutions,^{36-38, 59} while 6 did not report the source of participants.^{48, 49, 54, 62, 64, 70} All studies provided detailed inclusion and exclusion criteria for participant selection.

The intervention methods of experimental group of studies included aquatic exercise training,^{30, 39, 48, 49, 51, 54, 56, 58, 60, 61, 64, 65, 69, 70} aquatic Tai Chi,^{37, 38, 57} aquatic trunk exercise,⁵⁵ motor dual task training in water,⁵⁹ underwater treadmill training,^{50, 52, 53, 62, 66-68} and the Halliwick method.^{57, 63, 71} Control groups of studies included received no specific physical activity,^{48, 52, 65, 70} land-based rehabilitation,^{36-38, 49, 50, 54, 56-58, 60, 62-64, 67-69, 71} conventional therapy,^{30, 39, 51, 55} motor dual task training on land,⁵⁹ or over ground treadmill walking.^{53, 66} The intervention frequency ranged from 2 to 6 sessions per week, with each session lasting 30-60 minutes, and the intervention duration ranged from 2 to 12 weeks.

The instruments in the study for assessing balance function include BBS,^{30, 38, 39, 50, 54-57, 63, 67, 69, 70} TUGT,^{38, 51, 54, 67, 69, 70} FRT,^{54, 55, 63, 67} and FTSST.^{38, 54} The MBI/BI and FIM were used for ADLs. Walking speed,^{59, 64, 71} 2-minute walk test,^{60, 67} 6-minute walk test,⁶⁸ 7.62-meter walk test,⁷⁰ 10-meter walk test,⁵⁴ etc., were used to assess walking ability. Other outcome indicators include VAS^{36, 37}and SF-36.^{37, 69}

Risk of bias assessment

The results of the quality of included studies are shown in Figure 2 and Figure 3.^{30, 36-39, 48-71} Among the studies included, only 19 studies detailed the methodologies employed for generating random sequences, such as the use of sealed envelopes,^{30, 50, 51, 59, 68, 69} random number tables or computer-generated randomization,^{36-38, 55, 57, 59, 62, 63, 67} lottery method,⁷⁰ coin tossing,⁶⁰ drawing lots.⁷¹ Regarding allocation concealment, only 16 studies have offered comprehensive descriptions of their random allocation concealment strategies, which include the utilization of sealed envelopes^{30, 36-38, 50, 57, 60, 63, 65, 67-69} or undertaken by a researcher who blinded the grouping.^{51, 56, 57, 62} One study reported that there was a significant difference between experimental group and control group regarding the outcomes.⁶⁹ Finally, in the “Randomization Process” domain, 14 studies were ultimately assessed as having “Some Concern” due to issues such as the absence of random allocation or concealed random allocation, in addition to baseline data that were not comparable.^{39, 48, 49, 52-54, 58, 60, 64-66, 69-71} Five studies had a dropout rate exceeding 5% and did not conduct an ITT analysis, thus being classified as high-risk in the field of missing data.^{30, 48, 56, 57, 69} Only seven studies reported the research plan, and the statistical plan was made public in advance. Therefore, this category of selective reporting was rated as low risk.^{30, 36-39, 59, 65} The remaining 22 studies were evaluated as having “Some concerns” in the domain of “Selection of the reported result” due to inadequate evidence supporting the existence of a predefined statistical plan.

Figure 2.

—Risk of bias graph: review authors’ judgements about each risk of bias item presented as percentages across all included studies.^{30, 36-39, 48-71}

Figure 3.

—Risk of bias summary: review authors’ judgements about each risk of bias item for each included study.^{30, 36-39, 48-71}

Effects of intervention

Balance function

As shown in Figure 4,^{30, 38, 39, 50, 51, 54-57, 63, 67, 69, 70} 13 studies focused on the impacts of hydrokinesitherapy on balance function in stroke survivors, assessed using the BBS, TUGT, FRT, and FTSST. Due to the diversity of measurement instruments employed across the studies, subgroup analyses were performed rather than an overall pooling of data. Twelve studies reported the effect of water-based exercise for stroke survivors based on the BBS score. The results indicate that the balance function of the intervention group was improved compared with the control group with a moderate effect size (N.=523, SMD=0.65, 95% CI: 0.22-1.08, P=0.003). Six studies reported the effect of hydrokinesitherapy for stroke patients based on TUGT. The results proved the test time for TUGT in the intervention group was not improved (SMD=-0.46, 95% CI: -0.96 to -0.04, P=0.07), although the heterogeneity was considerable (I²=67%). Four studies focused on the effect of water-based exercise for stroke based on FRT. The results indicated that aquatic exercise therapy can improve the FRT performance of stroke patients (N.=107, SMD=0.97, 95% CI: 0.32-1.61, P=0.003), with moderate heterogeneity (I²=59%). Similar results were also obtained in the outcome indicator FTSST (N.=65, SMD=-1.59, 95% CI: -2.17 to -1.01).

Figure 4.

—Meta-analysis of hydrokinesitherapy on balance ability.^{30, 38, 39, 50, 51, 54-57, 63, 67, 69, 70}

Walking ability

As shown in Figure 5A,^{54, 59, 60, 64, 67, 68, 70, 71} eight studies investigated the effects of hydrokinesitherapy on walking ability, using a variety of instruments such as walking speed,^{59, 64, 71} the 2-minute walk test,^{60, 67} the 6-minute walk test,⁶⁸ the 7.62-meter walk test,⁷⁰ and the 10-meter walk test.⁵⁴ The meta-analysis results indicated there is insufficient evidence to indicated the water-exercise can improve walking ability for stroke survivors.

Figure 5.

—A) Meta-analysis of hydrokinesitherapy on walking ability;^{54, 59, 60, 64, 67, 68, 70, 71} B) sensitivity analysis of hydrokinesitherapy on walking ability.^{54, 59, 60, 64, 67, 68, 70, 71}

ADL

As shown in Figure 6A,^{30, 39, 50, 55, 62, 68, 69} seven studies investigated the effects of hydrokinesitherapy on ADL, employing either the MBI or the FIM as assessment tools. The results indicate that there is not enough evidence to prove that the water exercise can improve the ADLs using MBI/BI or FIM (N.=277, SMD=0.75, 95% CI: -0.40 to 1.90, P=0.20; N.=111, SMD=-0.18, 95% CI: -0.55 to 0.20, P=0.35, respectively).

Figure 6.

—A) Meta-analysis of hydrokinesitherapy on ADL;^{30, 39, 50, 55, 62, 68, 69} B) sensitivity analysis of hydrokinesitherapy on ADL.^{30, 39, 50, 55, 62, 68, 69}

Dropout

Among the studies included in this study, 13 reported both the number of participants and the reasons for these dropouts.^{30, 48, 50, 52, 53, 55-57, 60, 62, 63, 69, 70} Twelve trials explicitly stated that there were no participants lost to follow-up.^{36-39, 49, 51, 59, 61, 65-68} The remaining 4 studies did not provide any information concerning attrition.^{54, 58, 64, 71}

Adverse effects

Twelve studies reported that no participants experienced adverse events.^{36-38, 49-51, 61, 62, 66-69} One study reported that one participant experienced excessive fatigue.⁵⁶ In addition, one study reported one accident-related injury due to a fall in the control group.⁵² Although other studies did not provide information on adverse events, twenty-two studies reported reasons for participant dropout, including hip fracture,⁵⁷ health issues,³⁰ and excessive fatigue.⁶⁰

Meta-regression analysis

For the primary outcome indicator BBS, which included more than 10 studies, meta-regression was conducted to investigate the sources of heterogeneity. The overall model, including all six covariates, explained 43.30% of the between-study variance (adjusted R²=43.30%), but did not reach statistical significance [F (6, 5)=2.06, P=0.2228]. Among the individual covariates, the type of stroke was significantly associated with the effect size (coefficient=-1.19, 95% CI: -2.32 to -0.05, P=0.043), indicating that the intervention effect may vary depending on the type of stroke. The country of study origin showed a borderline significant association (coefficient=-0.38, 95% CI: -0.81 to 0.04, P=0.069), suggesting that regional differences might also contribute to the observed heterogeneity. No significant associations were found for intervention of the experimental group, intervention duration, duration of stroke, or follow-up time (all P>0.05). The residual between-study variance (τ²) was estimated at 0.254, and the residual I² was 66.81%, indicating that a substantial proportion of heterogeneity remained unexplained after accounting for these covariates (details of the results of meta-regression are shown in Supplementary Table I).

Sensitivity analyses

To assess the robustness and reliability of the pooled results in this meta-analysis, a two-stage sensitivity analysis was conducted. Initially, protocol-specific measures were implemented at each study node across the phases of literature retrieval, data extraction, quality assessment, and data analysis, aiming to address potential sources of heterogeneity and bias. These measures specifically included: 1) independent screening of the literature by two researchers, with discrepancies resolved through discussion or consultation with a third senior researcher (if applicable); 2) verification of data accuracy by a second researcher against the original study reports, including full-text articles, supplementary materials, and (when available) study protocols, immediately following data extraction by the first researcher; 3) independent quality assessment of all included studies using the Cochrane Risk of Bias Tool 2 (ROB 2), where each domain was rated and cross-validated by two researchers; and 4) systematic confirmation of consistency in outcome definition and measurement methods across all included studies.

Subsequently, for secondary outcome indicators subgroup with a heterogeneity greater than 75%, which included fewer than 10 studies (such as MBI and walking speed), one of the included studies was removed to investigate the sources of clinical heterogeneity. As shown in Figure 5B,^{54, 59, 60, 64, 67, 68, 70, 71} for walking ability, exclusion of one Park (2016) reversed the direction of the pooled effect and eliminated heterogeneity (I²=0%), suggesting that the overall result was highly influenced by this single study. As shown in Figure 6B,^{30, 39, 50, 55, 62, 68, 69} for MBI, exclusion of Bei (2023) did not change the results of the overall effect, but reduced between-study heterogeneity from high to nearly zero (I² reduced from 94% to 0%), indicating that this study was the primary source of heterogeneity. Sensitivity analysis was carried out for BBS, TUGT, and MBI by one-by-one study exclusion using Stata 15.0 software. The results indicated that the pooled effect size remained statistically stable across all iterations of single-study exclusion (as shown in Figure 7^{30, 38, 39, 50, 51, 54-57, 62, 63, 67-70}). This suggests that no individual study disproportionately influenced the final pooled result, thereby affirming that the overall conclusions of this meta-analysis are robust against potential variations or biases inherent in the individual studies.

Figure 7.

—A) Sensitivity analysis of hydrokinesitherapy for stroke on BBS;^{30, 38, 39, 50, 54-57, 63, 67, 69, 70} B) sensitivity analysis of hydrokinesitherapy for stroke on TUGT;^{38, 51, 54, 67, 69, 70} C) sensitivity analysis of hydrokinesitherapy for stroke on MBI/BI.^{39, 50, 55, 62, 68}

Publication bias

To assess publication bias in this meta-analysis, Egger’s Test and funnel plot visualization were used for BBS. Egger’s test yielded P values of 0.951 (slope is 0.616, 95% CI: -0.774-2.007) for BBS. Funnel plots showed relatively symmetric study distribution around the pooled effect size for both indicators, suggesting low publication bias risk. Collectively, these results confirm that publication bias is unlikely to have influenced the pooled effect sizes or overall conclusions (as shown in Figure 8^{30, 38, 39, 50, 54-57, 63, 67, 69, 70}).

Figure 8.

—Funnel plot for assessing publication bias of BBS in included studies.^{30, 38, 39, 50, 54-57, 63, 67, 69, 70}

GRADE assessment

The quality of evidence for three primary outcomes (BBS, TUGT, and the MBI), was assessed utilizing the GRADE methodology (as shown in Table I). For the outcome indicator BBS, the evidence level was initially downgraded due to the presence of bias risk and inconsistency among the results of the included studies. However, it was subsequently upgraded owing to the substantial effect size observed. Consequently, the evidence level for this outcome indicator was deemed moderate. In contrast, the evidence levels for the outcome indicators TUGT and MBI were downgraded due to three primary concerns: risk of bias, inconsistency in study results, and publication bias. No factors were identified that warranted an upgrade. As a result, the evidence levels for these outcome indicators were assessed as very low.

Table I. —The level of GRADE for outcomes.

Outcomes/subgroup	Comments	N. participants (studies)	Quality of the evidence (GRADE)
Balance - Berg Balance Scale	SMD 0.65 (0.22 to 1.08)	523 (12 studies)	⊕⊕⊕⊝ moderatea, b, e
Balance - Time Up and Go Test	SMD -0.46 (-0.96 to 0.04)	211 (5 studies)	⊕⊝⊝⊝ very lowa, c, d
ADL - Modified Barthel Index	SMD 0.75 (-0.40 to 1.90)	277 (5 studies)	⊕⊝⊝⊝ very lowa, c, d

The basis for the assumed risk (e.g., the median control group risk across studies) is provided in footnotes. GRADE Working Group grades of evidence: - High quality (⊕⊕⊕⊕): Further research is very unlikely to change our confidence in the estimate of effect. - Moderate quality (⊕⊕⊕⊝): Further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate. - Low quality (⊕⊕⊝⊝): Further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate. - Very low quality (⊕⊝⊝⊝): We are very uncertain about the estimate. ^aThere is a risk of bias. ^bThe results included in the study are inconsistent. ^cThe number of participants in the included studies is less than 300. ^dThe number of included studies is less than 10. ^eLarge effect size.

Discussion

This meta-analysis systematically evaluated the efficacy and safety of hydrokinesitherapy in improving balance function, walking ability, and ADLs in stroke survivors, based on 29 eligible RCTs involving 1075 participants. It updates our previous meta-analysis by expanding the evidence base, increasing the number of RCTs from 13 to 29, and the sample size from 381 to 1075 to improve statistical power. At the same time, it includes the FTSST for balance assessment, systematically reports adverse events, conducts in-depth meta-regression to identify heterogeneity sources, and upgrades evidence quality from low to moderate. This upgrade supports the effectiveness of hydrokinesitherapy in improving balance function in stroke patients with balance disorders, providing more reliable evidence for clinical practice.

The core finding is that hydrokinesitherapy exerts a positive effect on balance function in stroke survivors, as reflected by significant improvements in the BBS, FRT, and FTSST (P<0.05), while the TUGT showed only a trend toward significance (P=0.07). The divergent results stem from the distinct balance constructs assessed by each tool. The BBS, FRT, and FTSST primarily evaluate static balance and in-situ functional limits, which are particularly responsive to aquatic intervention.⁷² The BBS focuses on static posture maintenance and voluntary weight shifting,⁷² the FRT assesses standing stability limits, and the FTSST measures lower limb functional strength and fall risk.⁷⁴ These tasks do not require continuous forward progression, enabling patients to focus on postural control mechanisms that benefit from the supportive aquatic environment. In contrast, the TUGT evaluates dynamic mobility and task-switching, involving sequential sit-to-stand, walking, turning, and re-seating.⁷⁵ Total TUGT time is influenced by gait speed, turning efficiency, and fear of falling,^{38, 73} and its turning phase is sensitive to balance and trunk control deficits.⁷⁴

This discrepancy in findings between static and dynamic balance measures also raises important questions about the transferability of aquatic exercise effects to real-world mobility. Notably, balance improvements from aquatic exercise may not immediately translate to complex land-based mobility tasks (e.g., TUGT), as the latter requires integrated neural control that aquatic exercise alone may not fully provide,⁷⁵ with intervention characteristics also affecting this transfer,⁷⁹ and instrumented TUGT studies indicate total time fails to capture subtle but clinically meaningful sub-phase improvements (e.g., turn duration) that better reflect balance gains.⁷⁷

Notably, insufficient evidence was found to support the beneficial effects of hydrokinesitherapy on walking ability or ADL. This may be attributed to three key factors. First, heterogeneity in assessment tools for walking ability (e.g., walking speed, 2-minute walk test) and ADL (MBI/BI) across included studies introduced methodological variability. Second, short intervention duration (2-12 weeks in most studies), which is insufficient to drive meaningful improvements in complex functional outcomes like walking ability and ADL that require long-term rehabilitation. Third, variations in participants’ baseline functional status (acute vs. chronic stroke), which may have influenced response to the intervention.

Meta-regression analysis for BBS identified stroke type as a significant moderator (P=0.043), indicating that the intervention effect may vary by stroke subtype, which has important implications for personalized clinical decision-making. Country of participants’ origin showed a borderline significant association (P=0.069), potentially reflecting regional differences in healthcare systems, intervention protocols, and patient populations (e.g., baseline comorbidities, rehabilitation guidelines). However, the overall meta-regression model was not statistically significant (P=0.2228), with 66.81% of residual heterogeneity unexplained, suggesting unmeasured factors (e.g., patient baseline balance severity, specific hydrokinesitherapy protocols, therapist experience) may also contribute. This variability underscores the importance of future research to explore factors affecting intervention outcomes, especially across different stroke populations.

The sensitivity results indicated that the sources of heterogeneity for the two outcome indicators, walking speed and MBI, may be completely different from those of BBS. For MBI/BI, excluding Bei (2023), the only study that enrolled acute-phase stroke patients,³⁹ decreased from 94% to 0% without changing the effect direction. This is significant because acute-phase patients generally have a higher potential for ADL recovery, while those in later phases tend to have more stable and limited improvement potential.^{36, 67} For the walking speed, a critical finding was that excluding the 2016 Park study not only reversed the direction of the pooled effect but also reduced between-study heterogeneity to 0%, confirming this single study was the core of heterogeneity and the main driver of unstable walking speed results. There are two main reasons for our consideration. Firstly, Park’s study focused on trunk training programs, utilizing underwater interventions to enhance trunk muscle activation and gait function. This trunk-specific focus differs from other studies in the subgroup, which may have exerted a greater impact on walking speed. Secondly, the participants in Parker’s 2016 study could walk independently, not with moderate hemiplegia or partial walking dependence. This higher baseline functional reserve in Parker’s study allowed for greater improvements in walking speed from aquatic training. In contrast, leave-one-out sensitivity analyses for BBS, TUGT, and MBI confirmed stable pooled effect sizes, indicating the overall conclusions of this meta-analysis are robust and reliable for clinical reference.

Limitations of the study

Despite rigorous methodology, several limitations should be acknowledged to guide the interpretation of the results. Firstly, the research quality is generally poor due to unclear randomization schemes and hidden allocation, the absence of pre-defined statistical plans, and the failure to conduct an ITT analysis with a small sample size. Secondly, heterogeneity among studies remains unexplained, which may affect the general applicability of the research results in different clinical settings. It is worth noting that for the walking speed, the differences were entirely caused by Parker’s research, which had unique intervention measures and participant characteristics. This highlights the necessity for more standardized intervention plans in future research. Thirdly, the differences in intervention plans (types, frequencies, durations) among different studies hinder the determination of the optimal parameters of hydrodynamic therapy.

To address the limitations of this study and enhance the applicability of hydrokinesitherapy in stroke rehabilitation, several future research directions are proposed. First, high-quality, multi-center, and long-term follow-up RCTs with large sample sizes and standardized intervention protocols are needed to reduce heterogeneity and improve the reliability of evidence. ITT analysis should be implemented to minimize bias from dropouts, and detailed reporting of randomization, allocation concealment, and blinding should be prioritized. Second, future studies should focus on optimizing intervention parameters, including comparing the efficacy of different types of hydrokinesitherapy (e.g., underwater treadmill training vs. aquatic Tai Chi) and exploring personalized training tailored to stroke subtype, baseline functional status, and comorbidities. Third, subgroup studies should be conducted to identify populations that benefit most from hydrokinesitherapy, which will help guide targeted clinical application. Fourth, objective assessment tools (e.g., motion capture technology, electromyography) should be integrated to complement subjective scales, providing a more comprehensive and accurate evaluation of intervention effects. Instrumented balance assessments are recommended to use, especially the instrumented TUGT with wearable sensors, to capture sub-phase parameters such as turn duration, turn velocity, and sit-to-stand peak angular velocity. These kinematic measures are likely more sensitive to aquatic therapy effects than total completion time alone.^{74, 76} Finally, studies exploring the mechanism of hydrokinesitherapy in stroke rehabilitation (e.g., effects on neuroplasticity, muscle strength, and balance control pathways) are needed to provide a theoretical basis for its clinical application.

Conclusions

In conclusion, this meta-analysis demonstrates that hydrokinesitherapy has a moderate beneficial effect on balance function and is a safe intervention for stroke survivors, while insufficient evidence supports its effects on walking ability or ADL. Significant between-study heterogeneity exists, which may be attributed to stroke type, regional differences, and intervention protocol diversity. By expanding the evidence base, enriching outcome assessment, and conducting in-depth heterogeneity analysis, this study provides more reliable and timely evidence to guide the clinical application of hydrokinesitherapy in stroke rehabilitation. Future high-quality, standardized RCTs are needed to optimize intervention protocols, confirm long-term benefits, and further enhance the clinical value of this intervention.

Supplementary Digital Material 1

Supplementary Text File 1

Search strategy.

Supplementary Digital Material 2

Supplementary Table I

Sensitivity analysis.

Supplementary Digital Material 3

Supplementary Table II

Characteristics of included studies in this systematic review.^{30, 36-39, 48-71}