To comprehensively collect relevant studies, systematic searches were conducted across the following databases: Cochrane Library, Web of Science, PubMed, Embase, China Science and Technology Journal Database (VIP), China National Knowledge Infrastructure (CNKI), Wanfang Database (WF), and China Biomedical Literature Database (CBM). The search period spanned from the inception of each database up to 6 September 2025. The search strategy combined Medical Subject Headings (MeSH) and free-text terms, restricted to publicly published studies. Specific search terms are presented in Table 1 (using PubMed as an example), while Supplementary Table S2 provides the complete search strategies for each database.

Two researchers (Du and Mamuke) independently extracted data. They then cross-checked and validated the results. Discrepancies were resolved through discussion or by consulting an independent third party (Wang) to reach a consensus. Extracted data included: (a) basic information; (b) participant baseline characteristics and treatment methods, including electroacupuncture parameters when applicable (frequency/waveform/intensity when reported, session duration, and total course; see Supplementary Table S4.1); (c) key elements for risk of bias assessment; and (d) outcome measures and corresponding data.

Du and Mamuke independently assessed the risk of bias in the included RCTs and cross-checked their assessments for consistency. When they disagreed, Wang re-evaluated the studies. They used the tools recommended in the Cochrane Handbook for Systematic Reviews (Version 5.1.0).

This study employed Stata 17.0 software and its network package to conduct NMA. All continuous outcomes were summarized as mean differences (MDs) with 95% confidence intervals (CIs). Since lower values indicate improvement for all outcomes in this review, effect directions were aligned so that positive MDs consistently favored ARTs. The overall consistency of the evidence network was verified using a global inconsistency model (When P > 0.05, it indicates that global inconsistency is not significant; the consistency model was ultimately adopted throughout this study.) Local inconsistency was evaluated using node splitting (P < 0.05 indicating local inconsistency). Loop inconsistency was analyzed by calculating the 95% CI for the inconsistency factors (IF) using the ifplot command (a zero inclusion indicates no significant loop inconsistency). Simultaneously, the surface under the cumulative ranking curve (SUCRA) was calculated for each intervention, with pairwise comparisons presented in a league table to rank the probability of intervention effectiveness. SUCRA values (0%–100%) probabilistically rank intervention effectiveness, with higher values indicating a greater likelihood of being the optimal treatment. The methodological quality and risk of bias of included studies were assessed using Rev Man 5.3 software, as recommended by the Cochrane Collaboration.

The initial search across multiple databases yielded 7,711 relevant articles, including those from the Cochrane Library (n = 216), Web of Science (n = 267), PubMed (n = 225), Embase (n = 516), VIP (n = 989), CNKI (n = 1,566), WF (n = 2,028), and CBM (n = 1,904). Following multiple rounds of rigorous screening, 53 studies (24–76) were ultimately included in this NMA. These studies were published between 2007 and 2025. A detailed literature screening flowchart and results are presented in Figure 1.

This NMA included 53 studies involving 4,406 patients, all of which were RCTs (24–76). The studies encompassed 11 distinct treatment modalities: conventional treatment (CT), acupuncture (Acu), laser acupuncture (LA), abdominal acupuncture (AA), electroacupuncture (EA), catgut implantation at acupoint (CIAA), warm needle acupuncture (WNT), auricular seed therapy (AST), acupoint application therapy (AAT), moxibustion (Moxi), acupoint injection therapy (AIT). In addition to single-modality ARTs, some trials evaluated combined ART regimens (e.g., Acu + Moxi and EA + AST), which were treated as separate nodes in the network when applicable. Supplementary Table S3 details the abbreviations and definitions for the different therapies. Supplementary Table S4 provides further details on the included RCTs. A comprehensive network diagram illustrating all comparisons of outcomes is presented in Figure 2.

This study included 53 RCTs and conducted a comprehensive quality assessment. Regarding methods for random sequence generation, one study (31) allocated patients according to the order of diagnosis after admission, potentially introducing selection bias, and was therefore judged to be at high risk. Fourteen studies (33, 36, 38, 43, 44, 48–50, 58, 61, 65, 67, 74, 75) mentioned randomization but failed to describe the randomization method, and were therefore categorized as having an unclear risk. Thirty-eight studies (24–30, 32, 34, 35, 37, 39–42, 45–47, 51–57, 59, 60, 62–64, 66, 68–73, 76) were considered low risk as they employed random number tables or computer-based random sequence generators for subject allocation. Regarding allocation concealment, 5 studies (24–28) were low risk, whilst all remaining studies (29–76) failed to clearly describe their allocation concealment methods, resulting in an unclear risk rating. Only three studies (25, 27, 28) implemented blinding for both subjects and investigators and were assessed as low risk, whilst the remaining studies (24, 26, 29–76) made no mention of blinding and were rated as having unclear risk of bias. Regarding data integrity, seven studies (32, 34, 49, 51, 53, 66, 76) were deemed high risk for failing to explicitly analyze and report reasons for and data from patients who withdrew. All remaining studies (24–31, 33, 35–48, 50, 52, 54–65, 67–75) retained complete datasets. The presence of selective reporting bias was determined by verifying whether these methods were consistent with the reported outcomes. All studies (24–76) fully reported their findings and were thus considered low risk in this regard. All studies (24–76) were of unclear risk for other potential sources of bias and were therefore classified as unclear risk. Figure 3 displays the results of the bias risk assessment conducted using Rev Man 5.3. Specific details of the evaluation are provided in Supplementary Table S5.

The results of global inconsistency tests for all study outcome measures showed P > 0.05, indicating no significant inconsistency across the overall research. The inconsistency test results are presented in detail at the bottom of Supplementary Figures S1-S11. Local inconsistency tests employing node splitting yielded P > 0.05 across all analyses, indicating no significant local inconsistency. Detailed results are presented in Supplementary Table S7. Analysis of loop consistency revealed that the 95% confidence intervals for all loop IF encompassing closed-loop studies contained zero, suggesting non-significant loop inconsistency. Detailed results are presented in Supplementary Table S8.

Effect estimates and SUCRA rankings for all outcomes are presented in Table 2 and Figure 4. Detailed pairwise comparisons (league tables) are provided in Supplementary Tables S6.1-S6.11.

Adverse events were reported in 18 studies (24, 28, 32, 38, 40, 44, 48, 52, 53, 55, 56, 60, 63, 66, 67, 69, 70, 72). Of these, six studies (32, 48, 53, 63, 66, 67) reported no adverse events. The remaining 12 studies (24, 28, 38, 40, 44, 52, 55, 56, 60, 69, 70, 72) reported adverse events, most of which were minor. Most other reported events were mild and transient, and generally resolved after cessation of the intervention or with symptomatic management. However, a small number of serious events, such as tuberculosis, were reported, and relatedness to the interventions was unclear or not assessed. In the true acupuncture group, one participant withdrew during treatment because of tuberculosis; the original report did not provide an assessment of relatedness to the intervention. Most studies reported no statistically significant difference in adverse-event incidence between the intervention and control groups. Detailed results are presented in Table 3.

All funnel plots show that the study points are approximately symmetrically distributed on either side of the central axis. This symmetry indicates a relatively low likelihood of publication bias in this study. These findings are illustrated in Figure 5.

We found that ARTs were associated with reductions in IR-related metabolic outcomes, reproductive endocrine outcomes, and ovarian morphology in women with PCOS. In this NMA of 53 studies involving 4,406 participants, ARTs were compared with conventional medications and/or placebo. Several ARTs demonstrated benefits across key outcomes, particularly reductions in FINS and AFC. Reported adverse events were generally mild; however, adverse-event reporting was limited across trials.

Based on SUCRA rankings, AIT showed the highest probability of improving IR–related markers, including HOMA-IR and FINS. LA appeared particularly beneficial for reducing T and LH and improving the LH/FSH ratio. CIAA and AA showed favorable performance for anthropometric outcomes (BMI and WHR). Overall, Acu ranked highly across multiple outcomes.

Notably, the improvement in the LH/FSH ratio in our network appears to be driven predominantly by reductions in LH rather than increases in FSH, as FSH did not show statistically significant changes across interventions (all 95% CIs crossed zero). This pattern is biologically plausible. PCOS is characterized by accelerated GnRH pulse frequency, which preferentially increases pituitary LH secretion and contributes to ovarian androgen excess; in contrast, lower GnRH pulse frequencies tend to favor FSH secretion (77, 78). Therefore, interventions that modulate hypothalamic–pituitary signaling may yield a clearer and more consistent effect on LH than on FSH. In addition, LH (and the LH/FSH ratio) may be more sensitive to short-term changes in GnRH pulsatility, whereas any small effects on FSH could be diluted by heterogeneity in hormone sampling time points and assays across trials (78).

There has been a recent increase in NMA studies on PCOS, but few focus on acupuncture therapies to improve glucose metabolism in these patients. For example, WEI et al. (79) assessed acupuncture's effect on ovulation rates and outlined dosage parameters, such as the number of acupoints, treatment frequency, and duration, but did not examine glucose metabolism indicators. WANG et al.'s meta-analysis (80) addressed glucose metabolism by comparing acupuncture and metformin for IR in PCOS across 11 RCTs. Metformin was more effective for HOMA-IR, while acupuncture reduced fasting plasma glucose (FPG) more. However, this study had few trials and did not compare different acupuncture therapies. YU et al. (81) used NMA to assess ART effects on IR-related outcomes in obese PCOS, finding that Moxi combined with CIAA was most effective for FBG, FINS, and IR. This study, however, focused solely on obese patients and combined ARTs with metformin, making it difficult to assess the true effectiveness of each ART.

This study analyzed 53 articles covering all major ARTs for PCOS. It systematically assessed how these therapies improve glucose metabolism, body mass index, waist-to-hip ratio, sex hormone levels, and polycystic ovarian morphology. The results showed that different acupuncture modalities have distinct strengths with these indicators and may offer potential benefits compared with conventional treatments, while safety appeared acceptable based on reported adverse events, although adverse-event reporting was limited. Unlike traditional meta-analyses, network meta-analyses (NMAs) use direct and indirect comparisons across multiple interventions to identify the best treatments.

This study has several limitations. Firstly, the literature selection criteria were restricted to research published in Chinese and English, excluding papers in other languages. This limitation may result in an incomplete research perspective and a lack of global representativeness. Secondly, although the search strategy primarily focused on peer-reviewed journals, non-journal literature, such as conference abstracts, was excluded due to considerations regarding data quality and reproducibility. This may also affect the completeness of the results. Furthermore, for certain acupuncture therapies, the number of studies available for analysis was limited, with only 1 or 2 meeting the criteria. This scarcity of data may introduce potential bias and restrict the generalizability of the findings. Thirdly, EA parameters were not standardized across trials and were incompletely reported in some studies (Supplementary Table S4.1), which may contribute to clinical heterogeneity and affect comparability. Fourthly, due to the inherent characteristics of acupuncture treatment, blinding is difficult to implement. Treatment typically relies on interaction and communication between practitioner and patient; this non-blinded status may influence patients' subjective perceptions and responses to treatment. Fifthly, conventional treatment comparators varied across trials, which may challenge the transitivity assumption and contribute to clinical heterogeneity. Lastly, adverse event reporting is crucial for assessing the safety of acupuncture. However, most original studies fail to systematically record or report relevant adverse event information. Future research should place greater emphasis on safety indicators and evaluate them. In summary, we recommend conducting subsequent large-sample, long-course, multicenter, standardized RCTs with extended follow-up periods to further validate the efficacy and safety of acupuncture treatment.