Acupuncture and Moxibustion for Subthreshold Depression: a systematic review and Bayesian network meta-analysis

Abstract

Objectives

Subthreshold depression (SD), a precursor stage of depression, has gained considerable interest in recent years. Acupuncture and moxibustion have unique advantages in the treatment of depressive disorders. This systematic review and Bayesian network meta-analysis (NMA) aimed to evaluate the comparative effects of acupuncture and moxibustion therapies on SD and provide clinical recommendations.

Methods

Eight databases were searched from their inception to January 1, 2026, for randomized controlled trials (RCTs) on acupuncture and moxibustion for SD. We included 17 studies with six different therapies. The primary and secondary outcomes were scores on the Hamilton Depression Rating Scale (HAMD) and the Center for Epidemiologic Studies Depression scale (CES-D), respectively. A random-effects Bayesian NMA was conducted using the gemtc package in R.

Results

Acupuncture + moxibustion and acupuncture alone were associated with greater reductions in HAMD scores than placebo, no treatment, and Western medicine. In addition, acupuncture + moxibustion showed greater reductions in HAMD scores than acupoint herbal patching and psychological treatment. Acupuncture + moxibustion and moxibustion were associated with greater reductions in CES-D scores than placebo. Acupuncture + moxibustion and acupuncture alone resulted in greater score reductions than no treatment. Among the included therapies, acupuncture + moxibustion therapy was ranked highest in terms of surface under the cumulative ranking curve values for both HAMD (93.56%) and CES-D (96.21%).

Conclusion

This Bayesian NMA suggests that acupuncture + moxibustion may have a relatively favorable ranking for improving SD symptoms, providing preliminary evidence to inform future clinical research and practice. However, the generalizability and evidence quality were limited by the single-center, small-sample design of most included studies. Further high-quality RCTs are warranted to validate these findings.

Systematic review registration PROSPERO, CRD42023483384 (version 2.0)

INTRODUCTION

Depressive disorders are highly prevalent and represent the most common mental health disorders in primary care worldwide [1]. Their global prevalence is estimated at 27.6% and continues to rise, posing a major public health challenge [2]. Depressive disorders are also a leading cause of nonfatal health loss and are projected to become the second-largest contributor to the global disease burden by 2030 [3]. Subthreshold depression (SD) has attracted increasing attention as a precursor to depressive disorders. It is characterized by clinically relevant depressive symptoms such as depressed mood and loss of interest; however, it does not meet the diagnostic criteria for full depressive disorder [4]. The 5th edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-V) defines SD as the presence of at least two concurrent criteria for depressive symptoms lasting for at least 2 weeks, without meeting the full criteria for major depressive disorder (MDD) [5]. As a critical prodromal stage of MDD, SD is particularly common in adolescence, with a prevalence nearly twice that of MDD [6, 7]. A longitudinal study revealed that approximately 12% of individuals with SD progressed to diagnosable depressive disorders within 3 years, and the rates were even higher over longer follow-up periods [8-10]. Considering its prevalence and effects, SD poses a public health burden comparable to that of MDD [11]. However, despite its clinical significance, no effective pharmacological treatment exists for SD. Antidepressants and benzodiazepines have shown no proven efficacy against SD, whereas psychological treatments have shown only short-term benefits [12, 13]. Although cognitive behavioral therapy (CBT) is the preferred psychological treatment, it is expensive and complex [14, 15]. Therefore, the identification of cost-effective and durable interventions is necessary to prevent the progression of SD to MDD.

Skepticism toward pharmacological and psychological treatments for depressive disorders has resulted in a growing interest in complementary therapies such as acupuncture [16], exercise [17], meditation [18], and natural products [19]. As a representative traditional Chinese medicine, acupuncture has demonstrated substantial clinical benefits and is considered as an effective and safe option for managing the symptoms of depression [20, 21]. The use of acupuncture and moxibustion in SD treatment has expanded owing to their fewer side effects and lower cost than pharmacological and psychological interventions. However, despite evidence for their efficacy in SD, their optimal application has not yet been elucidated. This systematic review and Bayesian network meta-analysis (NMA) aims to integrate clinician-rated and patient-reported outcomes to enable direct and indirect comparisons of different acupuncture and moxibustion therapies and to provide evidence-based recommendations for optimal SD treatment.

METHODS

1. Registration

This NMA adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines [22]. The systematic review was registered with the International Prospective Register of Systematic Reviews (PROSPERO) (registration no. CRD42023483384, version 2.0).

2. Data sources and search strategy

We searched four Chinese and four English databases for relevant randomized controlled trials (RCTs) evaluating acupuncture and moxibustion for SD published from database inception to January 1, 2026. The databases included the China National Knowledge Infrastructure Database, China Science and Technology Journal Database, Wanfang Database, China Biomedical Literature Database, PubMed, Cochrane Library, Embase, and Web of Science. We also conducted supplementary searches for ongoing trials in the clinical trial registries such as ClinicalTrials.gov, WHO International Clinical Trials Registry Platform, and the Chinese Clinical Trial Register and for gray literature. Our search strategy followed the guidelines of the Cochrane Handbook, using Boolean operators (AND, OR, and NOT) to combine Medical Subject Headings and free-text keywords. This approach was adapted to the specific characteristics of each database. We manually retrieved references from identified studies to supplement the relevant literature. Our search terms included “acupuncture,” “moxibustion,” “electroacupuncture,” “body acupuncture,” “auricular acupuncture,” “scalp acupuncture,” “fire needling,” “warm acupuncture,” “subcutaneous needling,” “acupoint herbal patching,” “acupoint injection,” “acupoint catgut embedding,” “auricular point sticking,” “subthreshold depression,” “minor depression,” “subsyndromal depression,” “subclinical depression,” and other related terms. Both Chinese and English terms were used. Our search included no restrictions on language or publication year. The retrieval records were managed using EndNote 20 (Clarivate Analytics, Philadelphia, PA, USA). Details of the PubMed search strategy are provided in Table 1, and the strategies for the Chinese databases are presented in Supplementary Table S1.

Table 1.

PubMed database search strategy

Number	Search terms
#1	"Acupuncture"[Mesh Terms]
#2	"Acupuncture"[Title/Abstract] OR "Acupuncture Therapy"[Title/Abstract] OR "Acupuncture Treatment*"[Title/Abstract] OR "Treatment, Acupuncture"[Title/Abstract] OR "Therapy, Acupuncture"[Title/Abstract] OR "Pharmacoacupuncture Treatment"[Title/Abstract] OR "Acupotom*"[Title/Abstract] OR "Acupuncture"[Title/Abstract]
#3	#1 OR #2
#4	"Moxibustion"[Mesh Terms]
#5	"Moxibustion"[Title/Abstract] OR "Moxibustion Therapy"[Title/Abstract]
#6	#4 OR #5
#7	"Electroacupuncture"[Mesh Terms]
#8	"Acupuncture, Ear"[Mesh Terms]
#9	"Electroacupuncture"[Title/Abstract] OR "Body Acupuncture"[Title/Abstract] OR "Auricular Acupuncture"[Title/Abstract] OR "Scalp Acupuncture"[Title/Abstract] OR "Fire Needle"[Title/Abstract] OR "Warm Acupuncture"[Title/Abstract] OR "Subcutaneous needles"[Title/Abstract] OR "Acupoint herbal patching"[Title/Abstract] OR "Acupoint injection"[Title/Abstract] OR "Acupoint injection therapy"[Title/Abstract] OR "Acupoint catgut embedding"[Title/Abstract] OR "Acupoint embedding"[Title/Abstract] OR "Auricular point sticking"[Title/Abstract]
#10	#7 OR #8 OR #9
#11	#3 OR #6 OR #10
#12	"Depressive Disorder"[Mesh Terms] OR "Depression"[Mesh Terms]
#13	"Subthreshold Depression"[Title/Abstract] OR "Minor Depression"[Title/Abstract] OR "Subsyndromal Depression"[Title/Abstract] OR "Sub-clinical Depression" [Title/Abstract] OR "Mild Depression"[Title/Abstract]
#14	#12 OR #13
#15	"Randomized Controlled Trial" [Publication Type]
#16	"Randomized Controlled Trial"[Title/Abstract] OR "Controlled Clinical Trial"[Title/Abstract] OR "RCT"[Title/Abstract] OR "Randomly Assigned"[Title/Abstract]
#17	#15 OR #16
#18	#11 AND #14 AND #17

3. Selection criteria

1) Participants

Because of the lack of clear, unified diagnostic criteria for SD, participants were included if they met either the DSM-V or the International Classification of Diseases criteria and fell within pre-established cutoff scores on the Hamilton Depression Rating Scale (HAMD; HAMD-17: 7-17; HAMD-24: 8-20) and the Center for Epidemiologic Studies Depression scale (CES-D ≥ 16). This means of defining SD was applied consistently across all participants, regardless of age, gender, or nationality.

2) Intervention

The treatment group included all participants with SD treated with any form of acupuncture, moxibustion, or a combination of both.

3) Comparator

The control group included all participants with SD who received no treatment or were treated with any therapies other than those used in the treatment group, including psychological treatment, Western medicine, placebo, or no treatment. In multiarm trials, only those control arms meeting predefined comparator criteria were selected.

4) Outcomes

This study was initially registered with PROSPERO. The rate of clinical efficacy, HAMD-24 scores, CES-D scores, and subthreshold depressive symptoms (SDS) were specified as the study outcome measures. However, during the literature screening process, several prespecified outcomes were found to be insufficiently reported in the eligible studies. Therefore, to ensure the feasibility and clinical relevance of the review, we modified the outcome measures based on the data reported in the included studies, with HAMD designated as the primary outcome and CES-D as the secondary outcome. Clinical efficacy rates and SDS were excluded. We formally amended the details of our PROSPERO registration and updated it to version 2.0.

5) Study type

We included only RCTs in this review and NMA.

4. Study selection

Two independent reviewers (HL and CH) screened the database search results (primary and secondary screenings) according to our inclusion and exclusion criteria, extracted the required data, and cross-checked the results. Any disagreement was resolved by discussion until consensus was reached or through consultation with a third author (TM). The search process began with screening of titles and abstracts to exclude obviously irrelevant studies and duplicates (primary screening). Subsequently, the full texts of potentially eligible studies were assessed against the inclusion and exclusion criteria (secondary screening). When necessary, we contacted study authors to obtain uncertain but crucial information.

5. Data extraction

Microsoft Excel 2016 (Microsoft Corp., Redmond, WA, USA) was used to create a database of the required information extracted from the final included RCTs. The extracted data included study reference information (title, authors, and publication year), participant characteristics (diagnostic criteria, sample size, age, sex distribution, and disease duration), intervention details, outcome measures, study design characteristics (randomization methods, allocation concealment, and blinding), and other relevant information.

6. Assessment of the risk of bias and quality of evidence

The methodological quality of the included studies was assessed using the Cochrane risk-of-bias tool. Two reviewers (XM and YW) independently performed this assessment. We assessed each study in seven domains. These included random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessors, completeness of outcome data, selective reporting, and other biases. Each domain was classified as having a high, low, or unclear risk of bias. The greater the number of domains judged to have a low risk of bias, the higher the credibility of the study data, and vice versa. Any disagreement between reviewers was resolved through discussion.

The Confidence in Network Meta-Analysis (CINeMA) framework was used to assess the quality of evidence. CINeMA evaluates confidence in the following six domains: within-study bias, reporting bias, indirectness, imprecision, heterogeneity, and incoherence [23, 24].

7. Data synthesis

The effects of different interventions for SD were compared using a Bayesian NMA. RevMan (version 5.4.1; The Cochrane Collaboration, Oxford, UK) was used to assess the quality of the included studies, and R (version 4.2.0; R Foundation for Statistical Computing, Vienna, Austria) was used to construct plots and conduct the Bayesian NMA. Because of variations in the version of HAMD used across the included studies, we used the standardized mean difference (SMD) to pool effect sizes. For the CES-D, we used the mean difference (MD). Effect sizes were calculated based on the mean score change from baseline to post-treatment and the corresponding standard deviations. Effect estimates were reported with 95% credible intervals (CrIs) derived from posterior distributions. Positive values in league tables indicated greater improvement in depressive symptoms with the row intervention than with the column intervention, whereas negative values indicated greater improvement with the column intervention.

We performed Bayesian NMA using a random-effects model and the gemtc package in R. Default noninformative priors were used, with a normal (0, 106) prior for relative treatment effects and a uniform (0, 5) prior for the between-study heterogeneity standard deviations (denoted as τ). We used the Markov chain Monte Carlo method with four chains, 10,000 burn-in iterations, 50,000 sampling iterations, and a thinning interval of one. We evaluated model convergence using trace plots, density plots, and Brooks–Gelman–Rubin diagnostic plots, with the potential scale reduction factor (PSRF) as the convergence criterion. PSRF values between 1.00 and 1.05 indicated satisfactory convergence, with values closer to 1.00 reflecting greater model stability. We assessed overall inconsistency by comparing the deviance information criterion between consistent and inconsistent models. For closed loops in the network, we used the node-splitting method to evaluate inconsistency between direct and indirect evidence, with p-values < 0.05 indicating significant inconsistency. Treatment rankings were generated by calculating the surface under the cumulative ranking curve (SUCRA). We assessed between-study heterogeneity using τ derived from the posterior distribution, with larger values indicating greater variability in treatment effects. We also described statistical heterogeneity using the global I2 statistic, with values > 50% indicating substantial heterogeneity. We performed meta-regression and sensitivity analyses to explore potential sources of heterogeneity. Publication bias was assessed using comparison-adjusted funnel plots.

RESULTS

1. Study screening and selection

Using the search strategy described in section 2, we retrieved 702 records from the eight databases. We removed 57 duplicates using EndNote. A further 354 records were excluded after screening the titles and abstracts. The remaining 291 full-text articles were assessed, and 17 studies were finally included. Details of the study selection process are provided in Fig. 1.

Figure 1.

Study selection flowchart.

2. Characteristics of included studies

This review and NMA included a total of 17 RCTs, involving 1063 participants, with 509 in the treatment group and 554 in the control group. The interventions used included acupuncture, electroacupuncture, moxibustion, acupoint catgut embedding, acupoint herbal patching, acupuncture + moxibustion, psychological treatment, Western medicine, placebo, and no treatment. Among the reported outcomes, 13 studies reported HAMD scores [25-37] and 11 reported CES-D scores [25, 27-30, 33, 35, 38-41]. Baseline participant data showed no significant differences between the two groups (p > 0.05). The characteristics of the included studies are presented in Supplementary Table S2.

3. Risk-of-bias assessment

For random sequence generation, 11 studies adopted the random number table method [25-33, 35, 40], 5 adopted simple randomization [34, 35, 38, 39, 41], and 1 adopted block randomization [36]. For allocation concealment, 10 studies used opaque sealed envelopes [26-31, 36, 37, 39, 41], whereas the remaining studies did not report allocation concealment. For blinding, seven studies implemented blinding of outcome assessors [27-29, 31, 37, 39, 40], one blinded participants [36], two were double-blinded [32, 41], one was unblinded [30], and the remaining studies did not report blinding. All studies had a low risk of incomplete reporting [25-41]. For selective outcome reporting, only one study prospectively registered its protocol and reported all predefined outcomes [36]. It was unclear whether there were other potential biases. The results of the risk-of-bias assessment are shown in Fig. 2.

Figure 2.

Percentage of items included in the studies that produced a risk of bias.

4. Network meta-analysis of Hamilton Depression Rating Scale scores

A total of 13 studies reported the effects on HAMD scores of acupuncture and moxibustion treatment in participants with SD, including 11 two-arm and 2 three-arm studies, forming two closed loops. The network plot for HAMD scores is shown in Fig. 3a. The NMA showed that acupuncture + moxibustion (SMD = 1.51, 95% CrI: 0.32, 2.70) and acupuncture alone (SMD = –0.94, 95% CrI: –1.69, –0.19) significantly reduced HAMD scores compared with Western medicine. Acupuncture + moxibustion (SMD = 2.04, 95% CrI: 1.02, 3.09), acupuncture alone (SMD = –1.46, 95% CrI: –2.12, –0.86), moxibustion alone (SMD = –1.27, 95% CrI: –2.25, –0.30), and acupoint catgut embedding (SMD = –1.81, 95% CrI: –3.16, –0.48) significantly reduced HAMD scores compared with a placebo. Acupuncture + moxibustion (SMD = 2.28, 95% CrI: 1.03, 3.52), acupuncture alone (SMD = –1.71, 95% CrI: –2.49, –0.91), moxibustion alone (SMD = –1.51, 95% CrI: –2.78, –0.20), and acupoint catgut embedding (SMD = –2.05, 95% CrI: –3.46, –0.61) significantly reduced HAMD scores compared with no treatment. Acupuncture + moxibustion significantly reduced HAMD scores compared with acupoint herbal patching (SMD = 1.79, 95% CrI: 0.24, 3.36) and psychological treatment (SMD = 1.61, 95% CrI: 0.08, 3.13). No significant differences were observed between other treatment comparisons (Table 2).

Figure 3.

Results of the NMA for HAMD. (a) Evidence network map of HAMD scores. (b) SUCRA rank of HAMD scores. ACU, acupuncture; WE, Western medicine; ACU + MOX, acupuncture + moxibustion; EA, electroacupuncture; MOX, moxibustion; ACE, Acupoint catgut embedding; AHP, acupoint herbal patching; PT, psychological treatment; PLA, placebo; NT, no treatment.

Table 2.

Results of a network meta-analysis of pre- and post-treatment scores on the Hamilton Depression Rating Scale and the Center for Epidemiologic Studies Depression scale among participants with subthreshold depression receiving different treatments

ACU	2.26 (–3.63, 8.29)	–0.48 (–7.82, 7.34)	-	-	0.37 (–4.81, 5.43)	6.05 (–0.03, 12.38)	7.73 (4.10, 11.67)	3.00 (–3.24, 9.39)	–5.84 (–12.95, 1.73)
–1.01 (–2.41, 0.42)	EA	–2.71 (–12.27, 7.03)	-	-	–1.88 (–6.59, 2.53)	3.83 (–4.80, 12.40)	5.49 (0.10, 10.91)	0.77 (–7.28, 8.76)	–8.10 (–17.38, 1.48)
–0.19 (–1.25, 0.83)	0.81 (–0.97, 2.55)	MOX	-	-	0.84 (–8.61, 9.76)	6.53 (0.08, 12.76)	8.19 (–0.35, 16.63)	3.48 (–6.45, 13.20)	–5.39 (–11.95, 1.21)
0.34 (–0.85, 1.52)	1.35 (–0.51, 3.17)	0.54 (–1.03, 2.12)	ACE	-	-	-	-	-	-
–1.21 (–2.55, 0.09)	–0.20 (–2.18, 1.70)	–1.02 (–2.54, 0.50)	–1.55 (–3.35, 0.19)	AHP	-	-	-	-	-
–1.04 (–2.20, 0.14)	–0.03 (–0.84, 0.77)	–0.84 (–2.38, 0.74)	–1.38 (–3.04, 0.30)	0.17 (–1.56, 1.97)	PT	5.69 (–2.21, 13.89)	7.37 (2.68, 12.39)	2.63 (–4.83, 10.35)	–6.21 (–14.87, 2.99)
–1.46 (–2.12, –0.86)	–0.46 (–2.04, 1.06)	–1.27 (–2.25, –0.30)	–1.81 (–3.16, –0.48)	–0.25 (–1.42, 0.91)	–0.42 (–1.79, 0.88)	PLA	1.65 (–5.60, 8.99)	–3.05 (–11.93, 5.70)	–11.90 (–19.50, –4.12)
–1.71 (–2.49, –0.91)	–0.70 (–2.33, 0.92)	–1.51 (–2.78, –0.20)	–2.05 (–3.46, –0.61)	–0.50 (–1.98, 1.05)	–0.67 (–2.09, 0.75)	–0.24 (–1.20, 0.77)	NT	–4.71 (–11.12, 1.52)	–13.58 (–21.72, –5.28)
–0.94 (–1.69, –0.19)	0.07 (–1.55, 1.65)	–0.74 (–1.96, 0.49)	–1.28 (–2.68, 0.12)	0.27 (–1.16, 1.75)	0.10 (–1.30, 1.48)	0.53 (–0.33, 1.41)	0.77 (–0.16, 1.68)	WE	–8.86 (–18.32, 0.95)
0.57 (–0.40, 1.54)	1.58 (–0.14, 3.29)	0.77 (–0.19, 1.75)	0.23 (–1.29, 1.76)	1.79 (0.24, 3.36)	1.61 (0.08, 3.13)	2.04 (1.02, 3.09)	2.28 (1.03, 3.52)	1.51 (0.32, 2.70)	ACU + MOX

The HAMD league table is on the lower left, with data presented as standard mean differences and their 95% credible intervals. The CES-D league table is on the upper right, with data presented as mean differences and their 95% credible intervals. The differences between baseline and post-treatment scores were used as the effect sizes. Positive values indicate greater improvement in depressive symptoms with the row intervention than with the column intervention. Negative values indicate greater improvement with the column intervention. Significant results are shown in bold.

ACE, Acupoint catgut embedding; ACU, acupuncture; AHP, acupoint herbal patching; CES-D, Center for Epidemiologic Studies Depression scale; EA, electroacupuncture; HAMD Hamilton Depression Rating Scale; MOX, moxibustion; NT, no treatment; PLA, placebo; PT, psychological treatment; WE, Western medicine.

SUCRA ranking of HAMD score improvements was the highest for acupuncture + moxibustion (93.56%). The remaining interventions were ranked as follows: acupoint catgut embedding (85.69%), acupuncture (76.38%), moxibustion (68.17%), Western medicine (42.41%), electroacupuncture (39.03%), psychological treatment (37.10%), acupoint herbal patching (30.66%), placebo (17.48%), and no treatment (10.0%) (Fig. 3b).

5. NMA of CES-D scores

A total of 11 studies reported CES-D scores for acupuncture and moxibustion therapies in participants with SD, including 8 two-arm and 3 three-arm studies, forming three closed loops. The network plot for CES-D scores is shown in Fig. 4a. The NMA results showed that acupuncture + moxibustion (MD = –11.90, 95% CrI: –19.50, –4.12) and moxibustion alone (MD = 6.53, 95% CrI: 0.08, 12.76) significantly reduced CES-D scores compared with a placebo. Acupuncture + moxibustion (MD = –13.58, 95% CrI: –21.72, –5.28), acupuncture alone (MD = 7.73, 95% CrI: 4.10, 11.67), electroacupuncture (MD = 5.49, 95% CrI: 0.10, 10.91), and psychological treatment (MD = 7.37, 95% CrI: 2.68, 12.39) significantly reduced CES-D scores compared with no treatment. No significant differences were observed among the other treatment comparisons (Table 2).

Figure 4.

Results of the NMA for CES-D. (a) Evidence network map of CES-D scores. (b) SUCRA rank of CES-D scores. ACU, acupuncture; WE, Western medicine; ACU + MOX, acupuncture + moxibustion; EA, electroacupuncture; MOX, moxibustion; PT, psychological treatment; PLA, placebo; NT, no treatment.

SUCRA ranking of CES-D score improvements was highest for acupuncture + moxibustion (96.21%). The remaining interventions were ranked as follows: acupuncture (67.23%), moxibustion (66.77%), psychological treatment (63.91%), electroacupuncture (44.20%), Western medicine (38.99%), placebo (16.98%), and no treatment (5.70%) (Fig. 4b).

6. Consistency and heterogeneity analyses

In the Bayesian framework, both fixed-effects and random-effects models were fitted. The random-effects model was selected based on model fit. Model convergence was achieved, with PSRF values approaching 1.00 after 50,000 iterations (Supplementary Fig. S1). Convergence was further supported by trace and density plots for all parameters (Supplementary Fig. S2). Consistency between the direct and indirect evidence was assessed using the node-splitting method for closed loops within the network. For both HAMD and CES-D scores, no significant inconsistencies were identified, with all pairwise comparisons having p-values > 0.05 (Supplementary Fig. S3).

Heterogeneity analysis showed that the estimated τ was 0.43 (95% CrI: 0.05, 1.09) for HAMD scores and 3.07 (95% CrI: 1.16, 6.53) for CES-D scores, indicating greater between-study heterogeneity for CES-D than for HAMD scores. The global I2 exceeded 50% for both outcomes in the network-level assessment (Supplementary Fig. S4). To further explore potential sources of heterogeneity, we conducted meta-regression analyses. Considering the use of different HAMD versions (17/24 items) across the included studies, we performed separate random-effects meta-regression analyses of HAMD and CES-D scores. In the HAMD analyses, the HAMD version and the mean participant age (young or middle-aged) were included as covariates. For the CES-D analysis, only the mean participant age was included as a covariate. The meta-regression analyses revealed no statistically significant associations, indicating that differences in the version of HAMD used and the mean participant age did not significantly moderate the treatment effects (Supplementary Table S3). However, because only 17 RCTs were included, the statistical power of the meta-regression analyses was limited by the sample size. Therefore, the possibility that different HAMD versions and mean participant age contributed to heterogeneity cannot be entirely ruled out.

7. Sensitivity analysis

Sensitivity analysis for HAMD showed that exclusion of studies with the greatest influence on the pooled estimates reduced heterogeneity and stabilized the effect estimates [27, 33, 35-37]. Similarly, for CES-D, the exclusion of influential studies did not materially change the overall findings, supporting the robustness of the results (Supplementary Fig. S5) [27, 33, 40].

8. Quality of evidence

Based on GRADE principles, we used the CINeMA approach to assess the certainty of evidence from direct and indirect comparisons. Overall, the quality of evidence ranged from very low to moderate. Most comparisons received very low certainty ratings, primarily due to within-study bias, reporting bias, imprecision, and heterogeneity (Supplementary Tables S4, S5).

9. Publication bias

We used comparison-adjusted funnel plots to assess publication bias for both the HAMD and CES-D studies. All funnel plots showed good symmetry, suggesting no evident publication bias (Supplementary Fig. S6).

DISCUSSION

NMA differs from traditional pairwise meta-analyses, which include only direct comparisons between two interventions. This approach synthesizes direct and indirect evidence to estimate the relative treatment effects of multiple interventions. It improves the precision of the effect estimates, ranks the different interventions, and enhances the generalizability of findings [42-44]. Frequentist and Bayesian approaches are the primary statistical frameworks for NMA [45]. The frequentist method provides point estimates with 95% confidence intervals, whereas the Bayesian approach incorporates prior and posterior distributions to estimate the probability rankings of interventions, providing a probabilistic framework to support evidence-based clinical decisions [46-48].

To our knowledge, this is the first Bayesian NMA to specifically evaluate the effects of different acupuncture and moxibustion therapies on HAMD and CES-D scores in participants with SD. HAMD and CES-D are both common measures of depressive symptom severity; however, they assess different dimensions of depression. HAMD is the most widely used clinical evaluation scale, whereas—as a self-reported depression symptom scale—CES-D is mainly used in epidemiological research [49, 50]. Our results showed that acupuncture + moxibustion, acupuncture, moxibustion, and acupoint catgut embedding were associated with greater reductions in HAMD scores compared with placebos and no treatment. Furthermore, acupuncture + moxibustion and acupuncture alone resulted in greater reductions in HAMD scores than Western medicine.

A previous meta-analysis reported no statistically significant difference between acupuncture and antidepressant medications in reducing HAMD scores [51]. Currently, selective serotonin reuptake inhibitors are the first-line pharmacological treatment for depressive disorders and are widely prescribed worldwide [52]. However, no standardized pharmacotherapeutic regimen has been established for SD, and a substantial proportion of patients are insensitive to antidepressant therapy. Moreover, long-term treatment, even with low-dose antidepressants, increases the risk of adverse gastrointestinal effects [53].

Previous studies have suggested that the antidepressant effects of acupuncture may be mediated by multiple pathways. It may regulate the hypothalamic–pituitary–adrenal axis, suppress stress responses, and/or enhance the neuroplasticity and functional connectivity of the limbic system, including the hippocampus and the amygdala-anterior cingulate cortex circuit [54-56]. Acupuncture stimulation is relatively mild, with fewer adverse effects than pharmaceutical interventions, making it more suitable for long-term treatment. These characteristics may partially account for its more favorable effects compared with Western medicine. We found that acupuncture + moxibustion exerted greater relative effects than acupoint herbal patching and psychological treatment in reducing HAMD scores. This may be attributable to its stronger direct effects than the longer onset time for noticeable improvements from psychological treatment [57]. Both acupuncture + moxibustion and moxibustion alone were associated with greater reductions in CES-D scores than placebos, whereas acupuncture + moxibustion, acupuncture alone, and electroacupuncture showed greater score reductions compared with no treatment.

Our results showed that acupuncture + moxibustion ranked highest in both the HAMD and CES-D studies among the included interventions in the alleviation of depressive symptoms. The therapeutic effects of acupuncture and moxibustion are complex processes, and their combination may produce synergistic effects [58]. More specifically, through local thermal stimulation, moxibustion may potentiate the effects of acupuncture by augmenting anti-inflammatory and neuroprotective mechanisms [59]. The SUCRA ranking for CES-D was slightly different from that for HAMD, with psychological treatment ranking higher on CES-D scores. This discrepancy may be attributed to the differences in the assessment dimensions and sensitivity of the two scales. HAMD emphasizes clinician-rated somatic and psychological symptoms, whereas CES-D captures self-reported psychosocial distress. Therefore, patient-centered interventions may demonstrate greater effects when evaluated using self-reported measures. In our study, CBT was classified as a psychological treatment. Cognitive restructuring, which helps reduce negative thoughts, has been reported to potentially mediate the improvement of depressive symptoms [60]. The favorable ranking of psychological treatment for CES-D scores may be linked to its targeted advantage in cognitive behavioral remodeling.

In contrast to the findings of a previous NMA that focused exclusively on CES-D scores [61], our study found a lower probability ranking for the effects of electroacupuncture. This may be related to the inclusion of more indirect comparisons with other acupuncture and moxibustion therapies as well as the use of active interventions (e.g., psychological treatment) in some control groups rather than only no treatment or waitlist controls. These factors may have diluted the relative effect size of electroacupuncture.

Notably, we found that heterogeneity varied across the outcome measures, with CES-D showing larger between-study variability than HAMD. This difference may be attributed to differences in participant characteristics, intervention protocols, baseline symptom severity, and the measurement properties of the outcome scales across included studies. Although meta-regression analyses were performed to explore potential sources of heterogeneity, residual heterogeneity remained. This persistence may be partly explained by the inadequate sample size and statistical power of these analyses, which limited the certainty and interpretability of the findings. Such residual heterogeneity can also arise from clinical and methodological diversity across included studies. Clinical heterogeneity can manifest as variations in treatment frequency, intervention duration, and diagnostic definitions. Such factors are potential effect modifiers, with implications for the transitivity assumption underlying NMA. Nevertheless, all included studies enrolled participants with SD and evaluated acupuncture-related interventions, providing a reasonable degree of clinical comparability across studies. The node-splitting method revealed no significant inconsistencies between the direct and indirect evidence, which is consistent with the plausibility of the transitivity assumption.

Inadequate implementation of blinding and allocation concealment may have caused methodological heterogeneity, introduced bias, and resulted in an overestimation of treatment effects. Consequently, the findings of this study should be interpreted with caution. To avoid such limitations, future trials should be prospectively registered, adhere strictly to CONSORT reporting guidelines, and implement standardized intervention protocols.

This review had some limitations. First, this review reflects a potential benefit of acupuncture + moxibustion for SD. However, nearly all the included studies were conducted in China, which may limit the generalizability of the findings to populations in other countries and regions. Thus, treatment selection should be individualized according to patients’ specific conditions, as therapeutic effects may vary across different populations. Second, the interpretation of treatment rankings was constrained by the sparse network structure and the generally small sample sizes of the included studies. Because the overall certainty of evidence ranged from low to moderate, the treatment rankings obtained should be considered exploratory. Additionally, although our comparison-adjusted funnel plots showed no obvious publication biases, the limited number of studies may have reduced the power of such assessments, and small-study effects cannot be fully ruled out. Third, the outcome measures analyzed in this study diverged from those prespecified in the original PROSPERO protocol. Although we formally amended the registered protocol, it was a deviation from the original protocol and should be acknowledged as a limitation. Fourth, although the included studies used diverse acupuncture and moxibustion therapies, only 17 trials were included. Finally, the lack of long-term follow-up data prevented the analysis of the sustained effects of acupuncture and moxibustion therapies.

CONCLUSION

This Bayesian NMA suggests that acupuncture + moxibustion may rank favorably among interventions for improving the symptoms of SD, providing preliminary evidence that may inform future clinical research and practice. However, because most included studies were conducted at a single center with small sample sizes, the generalizability and quality of evidence were limited. Further multicenter, high-quality, large-sample RCTs with direct head-to-head comparisons of different acupuncture and moxibustion therapies are necessary to validate these findings and strengthen the level of evidence.

SUPPLEMENTARY MATERIALS

Supplementary data is available at https://doi.org/10.3831/KPI.2026.29.3.245.