Comparative efficacy of different exercise interventions in patients with ankylosing spondylitis: a systematic review and network meta-analysis

Search strategy and information sources

A comprehensive literature search was performed in six significant databases: CNKI, Web of Science (WOS), Cochrane Library, Embase, Scopus, and PubMed. The search spanned from the inception of each database to March 5, 2025, using a combination of controlled vocabulary and free-text terms. Keywords included “Ankylosing Spondylitis,” “Spondylarthritis,” “AS,” “Exercise,” “Physical Activity,” “Sport,” and “Training.” The whole search strategy is detailed in Appendix S1. Independently, two trained researchers, L.K. K. and C.W. Y., conducted the literature search and screened the retrieved studies based on predefined inclusion and exclusion criteria. Any disagreements between the two reviewers were resolved through discussion. If consensus could not be reached, a third reviewer, C.X.W., was consulted to make the final decision and ensure selection accuracy and consistency.

Study selection

In this study, we classified the included exercise interventions into six distinct categories based on modality, environment (land vs. aquatic), and core training focus (aerobic, stretching, muscle exercise, or China health campaign). This classification reflects both physiological mechanisms and clinical application patterns described in existing guidelines and literature. Aquatic therapies are categorized separately because of their unique biomechanical and thermal features, which notably change joint loading and patient experience. The primary physiological goal of the exercise determines its category. For example, interventions aimed at improving endurance are classified as aerobic, while those focusing on joint range of motion are labeled as stretching. CHC interventions are distinctly grouped because they combined physical, respiratory, and psychological components specific to Eastern medical traditions.

Initially, all retrieved records were deduplicated using EndNote X9. Subsequently, two researchers (C.W. Y. and Z.P. M.) independently screened the titles and abstracts to identify studies that met the inclusion criteria. A full-text review was conducted to assess eligibility and quality for studies further initially deemed eligible. Any disagreements in the selection process were resolved through discussion, and if consensus could not be reached, a third researcher (C.X.W.) in the field was consulted for arbitration. The specific inclusion criteria were as follows: (a) The study design had to be either an RCT or an nRCT. (b) Participants had to be patients with a confirmed AS diagnosis. (c) The intervention had to involve one of the following exercise types: Land Aerobic Exercise (LAE) (e.g., running, yoga), Land Stretching Exercise (LSE) (e.g., stretching), Aquatic Aerobic Exercise (AAE) (e.g., water yoga, swimming), Aquatic Stretching Exercise (ASE) (e.g., water Pilates), China Health Campaign (CHC) (e.g., Tai Chi, Baduanjin), or Muscle Exercise (ME) (e.g., HIIT, respiratory muscle training). (d) The control group received no exercise intervention or underwent routine care, health education, or sham exercise training rather than active exercise therapy. (e) The study had to report at least one of the following outcome measures: BASDAI, BASFI, BASMI, or ASQoL score. The exclusion criteria were as follows: (a) Studies with erroneous or missing raw data that prevented valid data extraction. (b) Studies where the intervention group was not exclusively composed of AS patients, including other diseases or acute relapsing AS cases. (c) Studies without full-text availability, conference abstracts, review articles, or other secondary research. (d) Duplicate publications: If multiple studies originated from the same cohort with overlapping or redundant results, only the most comprehensive or most recently published study was included. (e) Studies involving exercise simulations or virtual interventions. For example, we excluded the study by Salbaş & Karahan (2023).

Data extraction

Two researchers (K.L.K and C.W.Y) independently extracted key information from the included studies. After initial extraction, they cross-checked each other’s data to ensure accuracy and consistency. Any discrepancies were resolved through discussion, and if needed, a third researcher (C.X.W) provided arbitration. The finalized data were used for the systematic review and subsequent quantitative analysis. The extracted data included: (a) Study details: first author and publication year. (b) Participant characteristics: sample size, average age, and gender distribution. (c) Intervention details: type of exercise intervention, standard treatment description, and intervention duration. (d) Outcome measures: primary study results, including BASDAI, BASFI, BASMI, and ASQoL.

Statistical analysis

This study utilized Stata 18.0 software for statistical processing and analysis. For a detailed description of the procedures used to perform network meta-analysis in Stata, readers are referred to Shim et al. (2017), who provide a comprehensive methodological framework for conducting network meta-analyses using frequentist models. To assess the risk of bias, Cochrane’s Risk of Bias tool for RCTs (Higgins et al., 2011) was applied independently by two reviewers (K.L.K and C.W.Y). Any discrepancies in the assessment were resolved through discussion with a third reviewer (C.X.W). Bias risk levels were classified according to Cochrane standards as “low risk”, “unclear risk”, or “high risk” and recorded accordingly. For continuous outcome variables, the mean deviation (MD) was used to estimate the pooled effect size, Transformation formulas for estimating the mean and standard deviation see Appendix S3. NMA was performed using a frequentist approach, combining direct and indirect comparison evidence. In the network structure diagram, each node represents an intervention, with its size proportional to the number of participants receiving that intervention. The thickness of the connecting lines indicates the number of studies that directly compared two interventions. Network consistency was examined using global and local inconsistency tests (loop inconsistency test). The NMA results were displayed in forest plots, where the zero line represents no effect (0), and results were interpreted based on effect sizes and 95% confidence intervals. To rank the effectiveness of interventions, we used surface under the cumulative ranking curve (SUCRA) values and mean rankings. SUCRA values (ranging from 0% to 100%) indicate the probability of an intervention ranking higher in effectiveness. A higher SUCRA value suggests a more effective intervention. Lastly, a funnel plot was used to evaluate publication bias across the included studies.

Basic features of the included literature

The PRISMA flow diagram outlines the study selection process (Fig. 1). We used two previous reviews to identify potentially eligible studies for inclusion (Luo et al., 2024; Boudjani et al., 2023). A total of 4,014 studies related to AS were identified through database searches. Based on predefined inclusion and exclusion criteria, the research team conducted a rigorous screening process, ultimately including 48 eligible studies—comprising 41 RCTs and seven nRCTs—involving 3,140 diagnosed AS patients, with 1,213 males and 947 females. The basic characteristics of the included studies are provided in Table 1. The included studies examined seven different types of exercise interventions, specifically: four studies introduced AAE (e.g., water walking, aquatic jogging, or water aerobics) in the experimental group (Gandomi et al., 2022; Dundar et al., 2014; Kjeken et al., 2013; Karapolat et al., 2009). Four studies implemented ASE (Gandomi et al., 2022; Gunay, Keser & Bicer, 2018; Gurcay et al., 2008; Ciprian et al., 2013). Twenty-five studies used LAE (e.g., walking, jogging, or aerobic workouts) (Basakci Calik et al., 2018; Sveaas et al., 2020; Dundar et al., 2014; Acharya et al., 2024; Wang et al., 2022; Gunay, Keser & Bicer, 2018; Karapolat et al., 2009; Gurcay et al., 2008; Basakci Calik et al., 2021; Karahan et al., 2016; Demontis et al., 2016; Karapolat et al., 2008; Masiero et al., 2011; So et al., 2012; Analay et al., 2003; Aytekin et al., 2012; Cagliyan et al., 2007; Günendi et al., 2010; Hsieh et al., 2014; Jennings et al., 2015; Karahan et al., 2016; Niedermann et al., 2013; Roşu et al., 2014; Rosu & Ancuta, 2015; Silva, Andrade & Vilar, 2012). Twenty-three studies applied LSE (García et al., 2015; Aksoy et al., 2017; Singh et al., 2023; De Souza et al., 2017; Li et al., 2017; Karapolat et al., 2009; Altan et al., 2012; Analay et al., 2003; Basakci Calik et al., 2021; Günendi et al., 2010; Hsieh et al., 2014; Jennings et al., 2015; Karapolat et al., 2009; Kasapoglu Aksoy et al., 2017; Masiero et al., 2014; Rodriguez-Lozano et al., 2013; Roşu et al., 2014; Rosu & Ancuta, 2015; Silva, Andrade & Vilar, 2012; De Souza et al., 2017; Sweeney, Taylor & Calin, 2002; Acharya et al., 2024; Cagliyan et al., 2007). Five studies examined CHC (e.g., Tai Chi, Baduanjin, and Qigong) (Chen, 2014; Qu et al., 2020; Lee et al., 2008; Levitova et al., 2016; Xie et al., 2019). Five studies investigated ME (e.g., bodyweight training, resistance band exercises, or machine-based resistance training) (Basakci Calik et al., 2018; So et al., 2012; Sveaas et al., 2014; Durmus et al., 2009; Widberg, Karimi & Hafström, 2009).

Network geometry

We assessed the effectiveness of six distinct exercise interventions for patients with ankylosing spondylitis. Figure 2 visually represents the direct comparisons among these interventions, where each node corresponds to a specific intervention. The size of each node reflects the sample size of that intervention, while the thickness of the connecting lines indicates the number of direct comparisons conducted between the two interventions. This network diagram offers a clear perspective on the relationships among interventions and the comprehensiveness of the evidence network. Figure 2A illustrates the network of different exercise interventions in relation to BASDAI. The analysis covers six intervention types: LAE, LSE, AAE, ASE, ME, and CHC. The largest nodes, corresponding to conventional treatment (CG), LAE, and LSE, indicate that these interventions had the most participants. The thickest line connecting CG and LAE suggests this pair was the most frequently compared indirect studies. The network is relatively well-balanced, with most interventions directly compared to the control group and some indirect comparison pathways present. Figure 2B depicts the network of exercise interventions concerning BASFI. Again, six interventions are included, with CG, LAE, and LSE having the largest sample sizes. The thicker connecting lines indicated that direct comparisons were particularly frequent between CG and LSE and CG and LAE. The CHC intervention appears less studied, as shown by its smaller node and limited direct comparisons. Despite this, the network remains robust, making it a suitable framework for a network meta-analysis. Figure 2C presents the network structure for interventions targeting BASMI. Here, LAE, LSE, and CG emerge as the most extensively studied interventions, reflected in their larger node sizes. The strongest direct comparison was between LAE and CG, reaffirming their status as key intervention pairs. The network structure in this diagram highlights the significant interventions and their relationships, providing valuable insights for subsequent effect ranking analyses. Figure 2D offers a more streamlined representation, focusing on exercise interventions in relation to ASQoL. The diagram features fewer nodes, centering on CG, LAE, AAE, and ASE. The most substantial comparison between CG and LSE suggests that there should be greater research attention on this pair. While this network can assess the relative effects of different exercise interventions, the structure is comparatively sparse, limiting the availability of indirect comparisons.

Risk of bias within studies

Figure 3 summarizes the risk of bias assessment for each study. Blinding of participants and staff was rarely reported, leading to a generally high risk of bias across most studies. This could introduce potential biases, such as expectation effects, affecting effect size estimations. Given the high risk of bias in the comparator studies, the certainty rating of all treatment groups was downgraded accordingly.

Main outcome: BASDAI

For the BASDAI index, our network meta-analysis incorporated 38 randomized controlled trials (RCTs) and four non-randomized controlled trials (nRCTs), encompassing 2,852 patients with ankylosing spondylitis. The distribution of studies across different exercise interventions was as follows: AAE (3), ASE (3), LAE (21), LSE (19), CHC (5), and ME (5). Our findings indicate that all exercise interventions outperformed the control group in improving BASDAI, as illustrated in Fig. 4A. More specifically, ASE [−1.42, 95% CI [−2.51 to −0.33], LAE [−0.94, 95% CI [−1.41 to −0.47]], and LSE [−0.49, 95% CI [−0.94 to −0.04]] demonstrated statistically significant benefits over conventional treatment (CG) in reducing ankylosing spondylitis symptoms. Meanwhile, AAE [−1.13, 95% CI [−2.34 to 0.08]], CHC [−0.55, 95% CI [−1.45 to 0.35]], and ME [−0.73, 95% CI [−1.49 to 0.03]] showed some potential for improvement, though their effects were not statistically significant. No significant differences were detected between exercise interventions in indirect comparisons. To determine the most effective intervention, we utilized the SUCRA method, a widely adopted ranking approach in network meta-analysis that accounts for both effect estimates and their uncertainty (Chen, Zhang & Zhu, 2023). Figure 5A and Table 2 reveal that ASE had the highest SUCRA score (85.5), followed by AAE (71.3), LAE (66.5), ME (51.2), CHC (39.6), and LSE (32.6), with the control group scoring the lowest (3.4). These results suggest that ASE is the most effective intervention for improving BASDAI.

Main outcome: BASFI

For the BASFI assessment, our network meta-analysis included 38 randomized controlled trials (RCTs) and five non-randomized controlled trials (nRCTs), with a total of 2,924 participants. The studies were distributed across the following interventions: AAE (four), ASE (three), LAE (22), LSE (21), CHC (three), and ME (five). Results from the network meta-analysis demonstrated that all exercise interventions outperformed the CG in improving BASFI, as shown in Fig. 4B. Further analysis revealed that AAE [−0.90, 95% CI [−1.50 to −0.29]], ASE [−1.74, 95% CI [−2.45 to −1.04]], LAE [−0.74, 95% CI [−1.05 to −0.43]], LSE [−0.54, 95% CI [−0.81 to −0.27]], and ME [−0.48, 95% CI [−0.83 to −0.13]] demonstrated statistically significant improvements compared to CG, whereas CHC [−0.15, 95% CI [−0.68 to 0.38]] did not show a significant effect. Indirect comparisons indicated that ASE significantly outperformed AAE [−0.85, 95% CI [−1.70 to 0.00]], whereas no significant differences were found between AAE and LAE, LSE, CHC, or ME. When compared to ASE, however, LAE, LSE, CHC, and ME all showed statistically significant differences, suggesting that ASE was a more effective intervention. No significant differences were observed when comparing LAE with LSE, CHC, or ME, and similar results were found in pairwise comparisons between LSE, CHC, and ME. According to the SUCRA ranking results (Fig. 5B, Table 2), ASE ranked the highest (99.6), followed by AAE (74.4), LAE (66.9), LSE (46.5), ME (40.2), CHC (17.5), and CG (4.9). These findings further confirm that ASE is the most effective intervention for improving BASFI.

Main outcome: BASMI

For the BASMI assessment, 25 randomized controlled trials (RCTs) and three non-randomized controlled trials (nRCTs) involving 1,459 participants were included. The number of studies for each intervention was as follows: AAE (two), ASE (three), LAE (16), LSE (12), CHC (three), and ME (three). As shown in Fig. 4C, the results of the network meta-analysis showed that compared to the CG, ASE [−1.06, 95% CI [−2.10 to −0.02]], LAE [−0.51, 95% CI [−1.01 to −0.01]], and CHC [−1.15, 95% CI [−2.18 to −0.13]] exhibited statistically significant improvements. However, while AAE [−0.88, 95% CI [−2.05 to 0.29]], LSE [−0.04, 95% CI [−0.54 to 0.45]], and ME [−0.77, 95% CI [−1.66 to 0.11]] showed a trend toward improvement, their effects did not reach statistical significance. Additionally, no significant differences were observed in indirect comparisons. The SUCRA ranking results (Fig. 5C, Table 2) indicated that CHC had the highest SUCRA score (78.7), followed by ASE (75.6), AAE (65.9), ME (60.6), LAE (45.4), LSE (14.0), and CG (9.9). These findings suggest that CHC holds the greatest potential for improving BASMI.

Main outcome: ASQoL

For the ASQoL assessment, a total of nine randomized controlled trials (RCTs) and one non-randomized controlled trial (nRCT) were included, involving 1,165 participants. The number of studies for each intervention was as follows: AAE (one), ASE (three), LAE (five), and LSE (four). As illustrated in Fig. 4D, the results of the network meta-analysis indicated that compared to the CG, ASE [−3.67, 95% CI [−6.17 to −1.18]] and LAE [−2.64, 95% CI [−4.50 to −0.79]] demonstrated statistically significant advantages. Although AAE [−2.59, 95% CI [−6.82 to 1.64]] and LSE [−0.63, 95% CI [−1.86 to 0.61]] showed a trend toward improvement, their effects did not reach statistical significance. Indirect comparisons revealed a significant difference between LSE and ASE [3.05, 95% CI [0.23–5.87]], while no other significant differences were observed. The SUCRA ranking results (Fig. 5D, Table 2) showed that ASE ranked highest (88.4), followed by LAE (65.3), AAE (62.4), LSE (26.7), and CG (7.2), further supporting its superior efficacy in improving quality of life.

Heterogeneity and inconsistency assessment

To ensure the reliability of our findings, we evaluated the inconsistency across all outcome measures. As shown in Appendix S4, the p-values for global, node, and loop inconsistencies in BASDAI, BASFI, BASMI, and ASQoL were all above 0.05, suggesting that the results were consistent and free from significant inconsistencies.

Publication bias

Because forty or more studies were included in the meta-analysis of the effects of exercise interventions on BASDAI (disease activity), BASFI (functional impairment), BASMI (spinal mobility), and ASQoL (quality of life), we used a funnel plot to test their publication bias. Appendix S5 shows that the scatter is upward distributed, with a fundamental balance between left and right and no significant publication bias among the studies.