Effects of Mind-Body Exercises on Pain Intensity in Individuals with Chronic Nonspecific Low Back Pain: A Systematic Review and Network Meta-Analysis of Randomized Controlled Trials

Search Strategy

We conducted a thorough literature search in the following databases: CINAHL, Cochrane Library, Embase, MEDLINE, PEDro, PsycINFO, SPORTDiscus, CNKI, Wanfang, and VIP in order to identify all relevant randomized controlled trials (RCTs). The search covered the period from database inception to July 11, 2026. The search strategy combined keywords and Medical Subject Headings (MeSH) terms related to “mind-body exercise,” “chronic nonspecific low back pain,” and “randomized controlled trials” (Table 1). Additionally, we manually screened the reference listings of the included studies to identify additional publications, thereby reducing the risk of omitting potentially eligible studies. The literature search was conducted by the first and the second researchers (YX and TF), who independently screened records and selected studies. In order to achieve consensus, discrepancies were resolved through discussion with a third researcher (ZL).

Eligibility Criteria

The inclusion criteria were established in accordance with the participants, interventions, comparators, outcomes, and study design (PICOS) framework. Participants (P): Adults (≥ 18 years) with a diagnosis of CNLBP, defined as pain located between the lower rib margin and the gluteal folds, persisting for ≥ 3 months, and without an identifiable specific pathological cause. Intervention (I): MBEs, such as Yoga, Pilates, Qigong (eg, Baduanjin, Yijinjing, Liuzijue, Wuqinxi), and TC, delivered by certified instructors using standardized protocols, primarily in group settings. Depending on the conditions of the individual, training intensity, posture, duration, and rest intervals may be adjusted as necessary. Comparison (C): Control groups receiving no intervention, usual care, health education, or conventional exercise therapy. Outcome (O): Pain intensity measured with the visual analog scale (VAS) as the primary outcome. Study design (S): RCTs.

Exclusion criteria were as follows: (1) Studies not published in English or Chinese; (2) conference abstracts, protocols, books, or theses; (3) studies without full-text access or sufficient data; (4) studies involving acute, subacute, or recurrent LBP with an unspecified duration; (5) LBP attributable to specific pathologies (eg, pregnancy, infection, neoplasms, osteoporosis, vertebral fractures, structural deformities, inflammatory diseases, nerve root compression, or cauda equina syndrome); and (6) animal studies.

Data Extraction

Two independent researchers (YX and TF) extracted data using a standardized data extraction form. Extracted data included: (1) study characteristics, including first author’s surname, year of publication, country, and sample size; (2) intervention details, including duration, frequency, and single-session length; (3) participant demographics; and (4) outcome measures (VAS). We contacted the corresponding authors twice to request missing information for studies with incomplete intervention data. If the required data could not be retrieved, the relevant studies were excluded from the analyses. Any discrepancies were resolved through discussion with a third researcher (ZL) until consensus was reached.

Risk of Bias Assessment

Risk of bias in the included studies was assessed using the Cochrane Risk of Bias 2 (RoB 2) tool. The tool evaluates five domains: bias arising from the randomization process; bias due to deviations from intended interventions; bias due to missing outcome data; bias in measurement of the outcome; and bias in selection of the reported result. Each domain was judged as low risk of bias, some concerns, or high risk of bias. Two reviewers (YX and TF) independently assessed the risk of bias. Disagreements were first discussed between the two reviewers; if consensus could not be reached, a third reviewer (ZL) was consulted, and the final judgment was reached through discussion among all three reviewers.

Statistical Analysis

All analyses were conducted using Stata version 15 (StataCorp, College Station, TX, USA). Because all pain outcomes included in the quantitative synthesis were VAS scores harmonized to a common 0–10 cm scale, treatment effects were expressed as mean differences (MDs) with 95% confidence intervals (CIs); negative MDs indicated lower pain intensity in the MBE intervention group. No other pain scales were included in the quantitative synthesis; therefore, standardized mean differences were not used. Random-effects models were used for pairwise meta-analysis, subgroup analyses, and meta-regression because true effects were expected to vary across studies. Between-study variance (τ²) was estimated using the DerSimonian-Laird method for pairwise and subgroup analyses. Heterogeneity was assessed using Cochran’s Q, the I-squared statistic (I²), and τ², and an approximate 95% prediction interval (PI) was calculated for the overall pooled effect. For multi-arm trials in pairwise analyses, the shared control-group sample size was divided equally across relevant comparisons, while the control mean and standard deviation were retained unchanged, to avoid double-counting. Exploratory subgroup analyses examined intervention duration, exercise frequency, session duration, weekly exercise volume, and comparator type; formal tests for subgroup differences were used to compare categories. Leave-one-out sensitivity analyses assessed the influence of individual comparisons on the pooled estimate and τ². Separate univariable random-effects meta-regression models examined study-level associations between effect estimates and intervention characteristics, participant characteristics, comparator type, continent, and exercise modality. Residual τ² was estimated using restricted maximum likelihood (REML). These analyses were considered exploratory and were not interpreted as establishing causal dose-response relationships. Funnel plots and Egger’s regression test were used to explore publication bias.

A random-effects consistency NMA was conducted for post-intervention VAS pain. The network comprised five nodes: Yoga, Pilates, TC, Qigong, and a broad control node. No intervention, usual care, health education, and conventional exercise therapy were combined as the control node to provide a common reference for comparing the four MBE modalities. This operational grouping did not imply clinical equivalence among control conditions; thus, interpretation of indirect comparisons depended on the assumption that these differences did not materially distort relative treatment effects. Transitivity was assessed descriptively by comparing age, baseline VAS score, intervention duration, exercise frequency, session duration, weekly exercise volume, and comparator type across the MBE nodes. A common between-study variance was estimated using REML, and multi-arm trials were modelled with their original arms while accounting for correlations between comparisons sharing a control group. Approximate comparison-specific 95% PIs were calculated for each MBE modality versus control. Global and local inconsistency were explored using the inconsistency model, node-splitting, and loop-specific approaches. Treatments were ranked using the surface under the cumulative ranking curve (SUCRA), which was interpreted alongside league-table estimates and their CIs and PIs.

Quality of the Evidence

The certainty of evidence was evaluated using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach, which classifies evidence into four levels: high, moderate, low, or very low. The assessment was conducted using the following domains: publication bias, inconsistency, indirectness, imprecision, and risk of bias.

Study Selection

Initial searches retrieved 3623 records. Upon the removal of duplicates, 1215 unique records were screened by title and abstract, leading to the exclusion of 1,037 records that did not qualify for inclusion. Full-text screening then eliminated 153 studies for the following reasons: (1) non-randomized controlled trials (n = 10); (2) conference abstracts, protocols, books, or theses (n = 19); (3) no full text available (n = 16); (4) inadequate outcome data (n = 28); (5) missing data (n = 36); (6) ineligible study populations (n = 14); (7) mismatched intervention-control group designs (n = 22); and (8) not published in English or Chinese (n = 8). A total of 25 studies met the inclusion criteria and were included in the final analysis (Figure 1).^13–37

Characteristics of the Included Studies

Table 2 summarizes the key characteristics of the included studies. The trials comprised 938 participants in 27 intervention groups and 737 in 25 control groups. In this study, each intervention group was categorized into one of four exercise modalities: Yoga, Pilates, Qigong, and TC. The intervention durations ranged from 3 to 24 weeks. The VAS was the primary outcome measure in all 25 studies.^13–37 Of the studies that were included, one was conducted exclusively on women, 16 included both sexes, and eight did not specify gender distribution. Participants’ ages ranged from 25.8 to 78.6 years, and the duration of LBP symptoms ranged from 0.25 to 35.9 years. Of the 27 intervention groups, six included Yoga, eight included Pilates, nine included Qigong, and four included TC interventions. Weekly intervention frequency ranged from 1 to 7 sessions, with a mean of 3 sessions per week. Session duration ranged from 30 to 120 minutes, with an average of 56.15 minutes per session. Weekly intervention duration ranged from 60 to 600 minutes.

Pairwise Meta-Analysis Results

The overall random-effects pairwise meta-analysis showed lower pain intensity with MBEs than with control conditions (MD = −1.564 cm, 95% CI −1.982 to −1.147; p < 0.001; Figure 2). Between-study heterogeneity was substantial (Q = 355.02, degrees of freedom [df] = 26, p < 0.001; I² = 92.7%; τ² = 1.0211 cm²). The approximate 95% PI ranged from −3.69 to 0.56 cm. The pooled MD was numerically close to the reported minimal clinically important difference (MCID) of 1.5 cm on a 0–10 cm VAS for individuals with CNLBP.³⁸

Subgroup Analysis

No statistically significant between-subgroup differences were observed for intervention duration (p = 0.328), exercise frequency (p = 0.682), session duration (p = 0.106), or weekly exercise volume (p = 0.618). In contrast, pooled effects differed significantly by comparator type (p < 0.001). The estimated effect was largest in comparisons with no intervention (MD = −2.722 cm, 95% CI −3.132 to −2.312). Detailed subgroup-specific effect estimates and heterogeneity statistics are presented in Table 3.

Meta-Regression Analysis

In univariable random-effects meta-regression, comparator type was associated with variation in study-level effect estimates (overall test for the categorical covariate, p = 0.0036; residual τ² = 0.9144 cm²). Compared with usual care, no intervention was associated with a more negative estimated MD (β = −1.649 cm, 95% CI −2.647 to −0.651; p = 0.002). No statistically significant associations were observed for the remaining examined covariates. Regression coefficients, 95% CIs, p values, and residual heterogeneity estimates are presented in Table 4.

Assessment of Transitivity

The descriptive assessment of potential effect modifiers is presented in Supplementary Table S1. Age, baseline VAS scores, intervention duration, exercise frequency, and session duration showed overlapping ranges across the MBEs nodes. In contrast, mean weekly exercise volume ranged from 119.17 min/week for Qigong to 250.00 min/week for Yoga. Comparator composition also differed across nodes: all summarized TC intervention arms were compared with no intervention, whereas 66.7% of Qigong intervention arms were compared with usual care. Thus, although several observed characteristics overlapped across nodes, variation in weekly exercise volume and comparator composition introduced uncertainty regarding the transitivity assumption.

Network Meta-Analysis Results

Risk of Bias

Risk-of-bias assessments for the VAS pain outcome in the 25 included studies are presented in Figure 5. For bias arising from the randomization process, 15 studies^{13,16,19,20,23,24,27,28,31–37} were judged to have some concerns, mainly because allocation concealment was insufficiently described and, in some studies, the method used to generate the random sequence was unclear. The remaining 10 studies^{14,15,17,18,21,22,25,26,29,30} were judged to be at low risk of bias. For bias due to deviations from intended interventions, 15 studies^{13–16,19–22,24,27–32} were judged to have some concerns because of insufficient information on trial-context-related deviations or uncertainty about whether the analyses appropriately estimated the effect of assignment to intervention. Ten studies^{17,18,23,25,26,33–37} were rated at low risk of bias, and none was rated at high risk in this domain.

For bias due to missing outcome data, four studies^16,20,23,27 were judged to be at high risk of bias because attrition was substantial and, based on the reported reasons for missingness, differences in attrition between groups, or other study-specific evidence, missingness was judged likely to depend on participants’ true pain outcomes. Eight studies^{13,21,22,24,26,29,30,36} were judged to have some concerns because the reasons for missing data, or the relationship between missingness and the true pain outcomes, could not be established with sufficient certainty. The remaining 13 studies^{14,15,17–19,25,28,31–35,37} were judged to be at low risk of bias.

For bias in measurement of the outcome, 22 studies^{13–16,18–22,24–27,29–37} were judged to have some concerns. In these studies, participants knew which intervention they had received and reported their own VAS pain scores; this could have allowed awareness of treatment allocation to influence outcome assessment, but the available study-specific information was insufficient to judge such influence likely. Three studies were rated at high risk. In Liu¹⁷ and Zou,²⁸ the authors reported that participants in the exercise groups had greater treatment expectations than those in the inactive control groups. In Silva,²³ the same researcher delivered the intervention and conducted the outcome assessments. In these studies, the reviewers judged that these circumstances, together with participants’ awareness of allocation and the subjective nature of VAS pain ratings, made it likely that outcome assessment was influenced by knowledge of the intervention received. No study was judged to be at low risk of bias in this domain.

For bias in selection of the reported result, 23 studies^{13,15–25,27–37} were judged to have some concerns, mainly because the available documentation was insufficient to verify prespecification of the reported VAS analyses. Two studies^14,26 were rated at low risk. Overall, 19 studies^{13–15,18,19,21,22,24–26,29–37} were judged to have some concerns and six^{16,17,20,23,27,28} were rated at high risk. No study was rated at low overall risk of bias.

Sensitivity Analysis

In leave-one-comparison-out sensitivity analyses, pooled MDs ranged from −1.685 to −1.475 cm. All pooled estimates remained negative, and all 95% CIs excluded zero (Supplementary Figure S1), indicating that no single comparison materially altered the direction or statistical significance of the pooled pairwise estimate favoring MBEs over control.

Publication Bias

The funnel plot appeared asymmetric (Supplementary Figure S2). However, Egger’s regression test did not detect statistically significant effects in the VAS pain analysis (p = 0.901; Supplementary Table S4).

GRADE Evidence Quality Classification

Table 6 presents the GRADE certainty assessments for VAS pain. The certainty of evidence was rated as low for all network comparisons assessed.

Safety and Adverse Events

Among the four studies reporting adverse events, the non-serious events were mainly muscle soreness and transient dizziness and were mild. One study also reported a serious adverse event that was judged unrelated to the intervention.

Main Findings

Across 25 RCTs, pairwise meta-analysis suggested that MBEs may reduce pain intensity in individuals with CNLBP compared with control conditions (MD, −1.56 cm; 95% CI, −1.98 to −1.15) on a 0–10 cm VAS, although substantial between-study heterogeneity was present. In the five-node NMA, point estimates favoured yoga, Pilates, TC, and qigong over the broad control node, but the NMA did not demonstrate clear differences among MBEs. The certainty of evidence was low for all network comparisons assessed according to GRADE. Although TC had the highest SUCRA value, this ranking did not establish its superiority. Comparator type was associated with variation in study-level effect estimates, and exploratory dose analyses did not identify an optimal regimen. Overall, the findings support a possible analgesic role for MBEs, but they do not establish a preferred intervention or an evidence-based standard exercise dose for individuals with CNLBP.

Clinical Interpretation of the Overall Pain Effect

Although the pooled MD was numerically close to a commonly cited MCID of 1.5 cm on a 0–10 cm VAS,³⁸ its clinical interpretation should remain cautious. The pooled estimate represents the average effect across the included trials rather than a guarantee that each intervention, comparator condition, or clinical setting will yield a similar benefit. The high I² value (92.7%) and between-study variance (τ² = 1.0211 cm²) indicate considerable variability in observed effects. In addition, the approximate 95% PI (−3.69 to 0.56 cm) suggests that the true effect in a new comparable study could range from a clinically important reduction in pain to little or no benefit, or a small increase in pain. Thus, the average effect may be clinically meaningful, but its magnitude and consistency across settings remain uncertain.

Comparative Effects and Treatment Rankings

Although TC had the highest SUCRA value, the present NMA does not establish TC as the preferred intervention. SUCRA describes the probability of achieving a high relative rank within the specified network rather than providing direct evidence of treatment superiority. The league-table comparisons did not demonstrate clear differences between TC and yoga, Pilates, or qigong, and the approximate 95% PIs for each modality versus the broad control node included the null. Thus, the comparative estimates and their expected variation across future comparable settings do not support a claim that any one intervention is superior.

This interpretation is reinforced by variation across previous NMAs. Zhao et al, using a broader exercise network, also ranked TC highest for pain (SUCRA, 77.4%),⁶ whereas Shi et al ranked Pilates highest for pain reduction.⁵ The concordance with Zhao et al does not confirm TC superiority, and the difference from Shi et al is not a contradiction. Instead, these findings show that treatment rankings depend on the available evidence, node definitions, and model specification. Rankings should therefore be interpreted alongside relative effect estimates, CIs, and PIs, rather than being used alone to select a “best” intervention.

Comparator Type and Interpretation of the NMA

Comparator type was associated with study-level effect estimates in exploratory subgroup and univariable meta-regression analyses, suggesting that it was a potential contributor to the observed heterogeneity. This association does not establish causal effect modification, but it indicates that the observed effects of MBEs may depend on the conditions against which they are evaluated. To provide a common reference for comparing yoga, Pilates, TC, and qigong, the five-node NMA grouped no intervention, usual care, health education, and conventional exercise therapy into a broad control node. Similar broad control nodes have been used in previous NMAs of exercise interventions for CNLBP.^6,7

This operational choice did not imply clinical equivalence among the control conditions. Comparator distributions differed across intervention nodes, and weekly exercise volume varied across modalities, introducing uncertainty regarding transitivity. Accordingly, the NMA provides conditional relative estimates within the prespecified five-node framework rather than evidence that the four control conditions are exchangeable; the indirect comparisons and treatment rankings should therefore be interpreted cautiously.

Exercise Dose and Implications for Practice

Neither formal subgroup tests nor exploratory meta-regression identified clear differences in pain reduction according to intervention duration, exercise frequency, session duration, or weekly exercise volume. These dose-related variables were evaluated separately in subgroup and univariable meta-regression analyses. The available evidence therefore does not support a validated optimal exercise prescription for individuals with CNLBP. Importantly, the absence of statistically significant differences should not be interpreted as evidence that exercise dose is unimportant or that all dose categories are equivalent. Rather, the available study-level evidence did not identify a sufficiently clear dose-related pattern to support a specific regimen. Exercise intensity was not evaluated because it was insufficiently reported and should be regarded as an important evidence gap.

In practice, MBEs may be considered as one of several non-pharmacological options for pain management in CNLBP, consistent with guideline support for exercise-based care.^2,3 Yoga, Pilates, TC, and qigong may be selected according to patient preferences, ability and willingness to participate, local availability, and clinical judgement, rather than SUCRA rankings alone. The present findings should not be used to prescribe a uniform or modality-specific exercise dose. Confirmation requires adequately designed head-to-head randomized trials that directly compare prespecified MBE modalities and exercise-dose regimens. Such trials should also provide complete descriptions of comparator care, intervention content, dose, intensity, adherence, co-interventions, and adverse events, using established exercise-reporting guidance.³⁹