We searched the PubMed, Embase, and Cochrane Library databases for relevant studies from inception to July 21, 2025. The search strategies for each database are shown in Supplementary Material 1.

Inclusion criteria: (1) Population: Adult patients diagnosed with type 2 diabetes mellitus (T2DM); (2) Intervention: Pharmacological interventions for T2DM as defined by the 2025 ADA Standards of Care, including: metformin, sodium-glucose cotransporter-2 inhibitors (SGLT2i), glucagon-like peptide-1 receptor agonists (GLP-1 RAs), dual glucose-dependent insulinotropic polypeptide and GLP-1 receptor agonists (GIP/GLP-1 RAs), dipeptidyl peptidase-4 inhibitors (DPP-4i), thiazolidinediones (TZDs, e.g., pioglitazone), and sulfonylureas (SU), human insulin, and insulin analogs; (3) Studies were designed as network meta-analyses (NMAs).

Exclusion criteria: (1) Population: Studies focusing on the treatment of diabetic complications (e.g., diabetic nephropathy) using non-antidiabetic medications, such as angiotensin-converting enzyme inhibitors (ACEIs), as these fall outside the scope of antidiabetic drugs specified in our protocol; (2) Interventions: Studies investigating Traditional Chinese Medicine, Chinese patent medicines, dietary supplements, fish oil, or probiotics, as these are not classified as core antidiabetic drugs per our predefined standards; (3) NMAs that did not perform a quantitative synthesis of data, as these cannot be adequately assessed using our methodological quality appraisal tools; (4) Studies not indexed in the Science Citation Index, as this review aims to synthesize evidence from analyses with established international recognition and influence; (5) Conference abstracts, posters, or studies for which only an abstract was available, due to insufficient methodological detail for comprehensive assessment.

The study selection process was conducted according to the pre-specified protocol. Initially, all identified studies were imported into ZOTERO 7.0.26 reference management software for duplicate removal. Two reviewers (Dai and Liu) independently screened the titles and abstracts of the remaining records to exclude those clearly not meeting the inclusion criteria. The potentially eligible studies were subsequently retrieved for detailed assessment. When full texts were unavailable, the corresponding authors were contacted to obtain them. The full texts were then thoroughly reviewed to finalize study inclusion, with reasons for exclusion being documented. The selection process was summarized using a PRISMA flow diagram, and a list of excluded studies was provided in the Supplementary Materials. Any disagreements between the two reviewers were resolved through consultation with a third reviewer (Hu).

Data extraction was independently performed by two reviewers (Dai and Liu) using a predesigned data collection form. The following data were extracted from each included study into Microsoft Excel 16.101: number of authors, publication year, country of the first author, continent of the country, economic development level, journal impact factor, Journal Citation Reports (JCR) category, number of primary studies included, network meta-analysis methodology, study design(s), interventions, reporting of funding sources, and protocol publication status. Disagreements were resolved by a third reviewer (Hu).

The methodological quality of the included studies was using the AMSTAR 2 tool (18). This tool is specifically designed for the critical appraisal of systematic reviews of healthcare interventions and comprises 16 items, 7 of which are considered critical. Each item was judged as “Yes”, “partial Yes”, or “No” based on the specific methodological details of each study. Subsequently, the overall grade of each study was categorized into one of four grades: “High”, “Moderate”, “Low”, or “ Critically low”. The detailed assessment criteria are provided in the Supplementary Material 2. The methodological quality assessment was performed using the AMSTAR 2 online tool (https://amstar.ca/Amstar_Checklist.php). This tool automatically generates an overall rating upon completion of all item assessments.

The reporting quality of the included studies was using the PRISMA-NMA statement, which comprises a checklist of 32 items (22). To evaluate the degree of compliance with the PRISMA-NMA statement, each item was assigned one of the following responses: “yes” for full compliance; “partial yes” for partial compliance; and “no” for noncompliance. Responses were scored as “1” point for each “yes,” “0.5” points for each “partial yes, “ and “0” points for each “no” (28, 29). The scores for all items were summed to derive a total quality score. Based on the total score, studies were categorized into three groups: low (up to the 25th percentile), moderate (the interquartile range), and high (the 75th percentile and above) (30). Additionally, for each item, percentage scores were arbitrarily classified as: good (>90%), medium (50%–90%), and poor (<50%) (31).

The methodological and reporting quality of the included studies was assessed independently by two reviewers (Dai and Liu). Disagreements were resolved by a third reviewer (Hu).

Data were analyzed using descriptive and inferential statistical methods. Categorical variables, including economic development level (developed/developing countries), continent, publication year (Up to and including 2017/After 2017, based on AMSTAR 2 publication date), JCR category (Q1–Q4, classified according to the JCR status at the time of this study), network meta-analysis methodology (Bayesian, frequentist, both, or unclear) (32), study design(s) (randomized clinical trials (RCTs) and non-RCTs, or mixed), interventions (DPP-4is, GLP-1 RAs, Insulin, SGLT2is, and Comparison of multiple antidiabetic drugs), outcomes (Glycemic, Cardiovascular, Kidney, Cardiovascular and kidney, Safety outcomes, and Other outcomes), reporting quality (high, moderate, low),reporting of funding sources (reported/unreported), and protocol publication status (pre-registered/not pre-registered), are presented as frequencies and percentages. Continuous variables—including journal impact factor, number of primary studies included, and number of authors—were described using medians (interquartile ranges) due to non-normal distribution. Group comparisons were performed using the chi-square test for categorical variables and the Kruskal–Wallis test for continuous variables. We further examined whether methodological quality differed across studies stratified by type of outcome, drug class, or analytical framework.

A binary logistic regression analysis was performed using R Studio version 4.3.1 to identify factors associated with methodological quality. The dependent variable was study quality, dichotomized into “Low” and “ Critically low” categories. Independent variables included economic development level (binary, with developing countries as reference), publication year (binary, with pre-AMSTAR 2 publication as reference), JCR category (binary, with Q3/Q4 as reference), protocol publication status (binary, with non-registered as reference), along with journal impact factor, number of primary studies included, and number of authors as continuous variables. The statistical significance was set at P < 0.05. All tables in this study were performed using Microsoft Word 16.101, while all figures were generated using Origin 2024.

Prior to performing multivariable logistic regression analysis, we assessed multicollinearity among the independent variables. Diagnostic measures, including the variance inflation factor (VIF) and tolerance, were calculated to evaluate potential collinearity. Generally, a VIF value exceeding 5 or 10 is considered indicative of substantial multicollinearity (33). In the present study, all independent variables exhibited VIF values ranging from 1.08 to 1.96, all well below the threshold of 5, suggesting no substantial multicollinearity concerns. Specific details are shown on Table 1.

The study selection process is detailed in Figure 1. A total of 1,003 records were identified through database searching: 215 from PubMed, 746 from EMBASE, and 42 from the Cochrane Library. After removing 186 duplicates, 817 records underwent title and abstract screening, resulting in the exclusion of 548 records. The remaining 269 full-text articles were assessed for eligibility, leading to the exclusion of 84 studies for the following reasons: not specific to type 2 diabetes (n=1), not SCI-indexed (n=4), not an NMA (n=13), duplicate publication (n=4), abstract only (n=48), ineligible interventions (n=12), and lack of quantitative analysis (n=2). Consequently, 185 studies were included for final data analysis and quality assessment. The list of excluded studies with specific reasons and the list of included studies are provided in the Supplementary Material 3.

The characteristics of the included studies are summarized in Supplementary Material 4. The annual number of published network meta-analyses (NMAs) demonstrated a general upward trend, with notable growth periods between 2014–2015 and 2020–2021, though a decline was observed after 2023. Publication frequency increased substantially following the release of AMSTAR 2 in 2017, with a statistically significant difference before and after the AMSTAR 2 publication (p = 0.02) (Figure 2). 108 studies (58.38%) originated from developing countries, compared to 77 (41.62%) from developed countries (p = 0.01). Asia contributed the majority of studies (122, 65.95%), with China accounting for 91 publications, while Africa had the fewest (1, 0.54%). Methodologically, most studies were published in high-impact journals, with 109 (58.92%) in JCR Q1 journals versus only 2 (1.08%) in Q4 journals (p = 0.03). The Bayesian approach was employed in 98 studies (52.97%), followed by frequentist methods (44, 23.78%), both methods (4, 2.16%), and unspecified methods (39, 21.08%), with significant differences among these categories (p < 0.001). The vast majority of NMAs (180, 97.30%) exclusively synthesized randomized controlled trials (RCTs), while 5 (2.70%) incorporated both RCTs and non-RCTs. The majority of studies involved comparisons of multiple antidiabetic drugs (122, 65.95%), while the remainder examined comparisons within the same drug class. Among these, studies on GLP-1 RAs were more commonly investigated (27, 14.59%), followed by SGLT2is (22, 11.89%), while DPP-4is (6, 3.24%) and Insulin (8, 4.32%) were the least studied. No statistically significant difference was found across the various intervention categories (p = 0.91). Nearly half of the included studies evaluated glycemic (88, 47.57%). A substantial number also focused on cardiovascular (33, 17.84%), kidney (9, 4.86%), and cardiovascular and kidney events (12, 6.49%). A smaller proportion examined safety outcomes (24, 12.97%), and other outcomes (neurological complications and gastrointestinal cancers) were also reported in a limited number of studies (19, 10.27%). No statistically significant difference was observed among the outcome categories (p = 0.24). More than three-quarters of the studies (142, 76.76%) reported funding sources, with a statistically significant difference between studies that reported funding sources and those that did not (p = 0.04). Over half of the studies (108, 58.38%) had pre-registration, and there was a statistically significant difference between protocol pre-registered and non-pre-registered (p < 0.001). The median impact factor was 4.60 (2.90–5.70). The median number of primary studies was 27.00 (14.00–54.00), and the median number of authors was 7.00 (5.00–10.00).

A total of 185 NMAs were assessed and scored across the 32 items of the PRISMA-NMA statement. The median reporting quality score was 29.50 (IQR 27.50–30.50). Seventeen NMAs reported all checklist items, while the lowest score was 18.5. Based on the total scores, 67 NMAs (36.22%) were rated as high quality, 63 (34.05%) as moderate quality, and 55 (29.73%) as low quality. A statistically significant difference was observed among these reporting quality categories (p < 0.0001).

Regarding the compliance rate for each item, the items with 100% compliance were item 1 (“Title”) and item 26 (“Conclusion of Evidence”), indicating that all included literatures mentioned “network meta-analysis” in their titles, summarized the results in the discussion section, and clarified the implications for future research. Items with high compliance rates included: item 3 (“Rationale”, 99.73%), item 25 (“Limitations”, 99.73%), item 21 (“Results of NMA”, 99.19%), item 27 (“Funding”, 98.92%), item 24 (“Conclusions of main findings”, 98.92%), item 18 (“Characteristics of studies”, 98.92%), item 4 (“PICOS”, 98.92%), item 9 (“Study selection methods”, 98.38%), item 17 (“Results of study selection”, 98.11%), item 12 (“Risk of bias within studies”, 96.76%), item 14 (“Analysis methods”, 96.76%), item 2 (“Abstract”, 96.22%), item 13 (“Summary measures”, 95.68%), item 7 (“Information sources”, 94.59%), item 10 (“Data collection process”, 94.05%), item 20 (“Results of individual studies”, 91.62%), and item 15 (“Risk of bias across studies”, 90.81%).

In contrast, the reporting of methodology and results specific to network meta-analysis was notably poorer. Lower adherence was found for the following items: item S3 (“Presentation of network structure”, 85.68%), item S1 (“Geometry of the network”, 81.35%), item S2 (“Assessment of inconsistency”, 79.73%), item S4 (“Summary of network geometry”, 68.92%), and item S5 (“Exploration for inconsistency”, 58.65%). Specific details are shown on Figure 3, Figure 4 and Supplementary Material 5.

The methodological quality of the 185 included studies was rigorously assessed by the AMSTAR 2. Only 1 study (0.54%) was rated “High”, 72 studies (38.92%) “Low”, and 112 studies (60.54%) “Critically low”; no study was rated “Moderate”. Among the 184 NMAs that did not fully satisfy the critical items, 180 (97.83%) did not report item 7 (“Consideration of risk of bias when interpreting results”). More than half of the studies (76.09%) only partially report item 4 (“Adequacy of the literature search”). Additionally, less than half of the NMAs failed to report the following critical items: item 2 (“Protocol registered before commencement”, 41.85%), item 9 (“Risk of bias from individual studies used appropriately”, 12.50%), item 11 (“Appropriate methods for statistical combination”, 19.02%), item 13 (“Consideration of risk of bias when interpreting results”, 20.65%), and item 15 (“Assessment of presence and likely impact of publication bias”, 32.07%). Compliance with individual AMSTAR 2 items varied substantially. Item 1: “PICO” was fully reported by all studies (100%). Whereas item 3: “explanation of study design selection” was not reported by any study (0%). Similarly, Item 7 “list of excluded studies” was adequately addressed in only 2.70% of the studies. Several items were satisfactorily reported “Yes” in over 80% of the studies: Item 6: “two reviewers perform data extraction” (89.73%), Item 16: “conflict of interest” (88.11%), Item 9: “risk of bias” (87.03%), Item 12: “Impact of risk of bias” (85.41%), Item 11: “statistical methods” (81.08%), and Item 5: “two reviewers perform study selection” (80.54%). Reporting for Item 2 (protocol registration) was nearly evenly split between “Yes” (58.38%) and “No” (41.62%). Item 4: “study search strategy”, Item 8: “Detailed PICOS of the Included Studies”, and Item 9 were special, with the proportions of “partial Yes” reports being 75.68%, 41.08%, and 0.54% respectively. Specific details are shown on Figure 5, Figure 6 and Supplementary Material 6.

Descriptive comparisons revealed significant heterogeneity in methodological quality in different subgroups. No statistically significant difference was observed among the different drug interventions (p = 0.59). “Comparisons of multiple antidiabetic drugs” constituted the highest proportion in both subgroups: specifically, in the critically low group (72, 64.29%) and the low group (49, 68.06%). Studies focusing on Insulin (8, 4.32%) and DPP-4is (6, 3.24%) accounted for a lower proportion in both groups. A statistically significant difference was observed across outcome categories (p = 0.04). The critically low group was predominantly comprised of studies focusing on glycemic (62, 55.36%), followed by cardiovascular (21, 18.75%). The low group also included a higher proportion of studies with glycemic (25, 34.72%), followed by safety outcomes (13, 18.06%). A statistically significant difference was noted among the analytical approaches. Bayesian methods were more frequently employed in the critically low group (71, 63.39%), whereas frequentist methods were more common in the low group (29, 40.28%). Approximately 20% of studies in both groups did not explicitly specify their analytical framework. These subgroup patterns should be interpreted as descriptive signals rather than inferential comparisons (Supplementary Material 7).

Given the imbalanced distribution of quality ratings, which could compromise model stability, the dependent variable was dichotomized for regression analysis. In the univariate analysis, several factors were significantly associated with higher methodological quality: studies published after the release of AMSTAR 2 (OR = 3.397, 95% CI = [1.39, 8.28], p = 0.007), higher journal impact factors (OR = 1.063, 95% CI = [1.01, 1.12], p = 0.017), a larger number of included studies (OR = 1.005, 95% CI = [1.00, 1.01], p = 0.012), protocol publication (compared with non-registration) (OR = 149.889, 95% CI = [19.75, 1137.43], p < 0.001), and higher JCR categories (OR = 4.636, 95% CI = [1.52, 14.10], p = 0.007). Conversely, studies from developed countries demonstrated lower quality (OR = 0.385, 95% CI = [0.20, 0.73], p = 0.003). No significant association was observed between the number of authors and study quality (OR = 1.063, 95% CI = [0.99, 1.14], p = 0.088). In the multivariate analysis adjusted for potential confounders, only three factors maintained statistical significance: protocol publication (OR = 399.130, 95% CI = [17.92, 8891.69], p < 0.001), and economic development level (OR = 0.380, 95% CI = [0.20, 0.73], p = 0.003). The following factors were no longer significantly associated with study quality in the adjusted model: journal impact factors (p = 0.008), publication year (p = 0.818), JCR category (p = 0.847), number of included studies (p = 0.258), and number of authors (p = 0.110). Specific details are shown on Table 2, Table 3, and Figure 7.

The multivariable regression analysis revealed an extremely elevated odds ratio for the variable “protocol registration status” (OR = 399.13). Statistical diagnostics indicated that this phenomenon was not caused by multicollinearity. Upon further investigation of the data characteristics, we found that among the included studies, those with a registered protocol were almost exclusively of higher quality. This distribution led to the extreme odds ratio and a wide confidence interval (17.92 – 8891.69). Nonetheless, the association remained statistically significant (p < 0.001), suggesting a strong correlation between protocol registration status and the methodological quality of the literature.

This study represents the first comprehensive assessment of the methodological quality of network meta-analyses (NMAs) that compares the efficacy of antidiabetic drugs for T2DM. Our analysis of 185 SCI-indexed NMAs revealed several key findings. Geographically, the majority of included studies originated from Asia and developing countries. While the annual number of NMAs on T2DM pharmacological treatment has shown a general increasing trend. Methodologically, over half of the NMAs employed Bayesian approaches for data synthesis. The majority of studies focused on Comparison of multiple antidiabetic drugs. Among studies comparing drugs within the same class, head-to-head comparisons of GLP-1 RAs were more commonly selected, while comparisons of DPP-4is were less frequent. Most studies primarily assessed glycemic, followed by cardiovascular and safety outcomes. Nearly all included studies were randomized controlled trials (RCTs), with very few non-randomized studies. However, none of the studies provided a rationale for their choice of study design, which may influence methodological quality. Overall, the general reporting quality of network meta-analyses was satisfactory, but notable deficiencies were observed in the reporting of methodology and results specific to NMA, primarily in the form of omissions or partial reporting. The methodological quality of most NMAs was rated low, with significant shortcomings in reporting critical items. Almost no studies reported on critical item 7, and only about half adequately addressed critical item 2. Therefore, researchers are advised to adopt more standardized practices in both reporting and methodological study design.

This study identified several methodological shortcomings in the network meta-analyses (NMAs) that compares the efficacy of antidiabetic drugs for T2DM published in SCI-indexed journals. These include a lack of protocol registration, failure to justify the selection of study design, and omission of searches in grey literature. In the absence of a registered protocol (29, 34), researchers might selectively include studies or outcomes based on preliminary results that favor their hypotheses or engage in excessive data dredging to achieve statistically significant findings. This practice can increase the likelihood of false-positive results while concealing unfavorable outcomes, thereby distorting conclusions and potentially misleading clinical decision-making, in addition to wasting research resources. Failure to provide a list of excluded studies with justifications compromises the reproducibility and transparency of the research. More critically, it may introduce selection bias, whereby researchers deliberately exclude undesirable results, skewing the final conclusions in a favorable direction. Furthermore, most studies did not provide a rationale for exclusively including randomized controlled trials (RCTs). While RCTs offer distinct advantages for evaluating drug efficacy, observational studies are often better suited for detecting long-term or rare adverse events. The exclusion of observational studies may lead to an underestimation or omission of such long-term safety outcomes, which are often difficult to identify in RCTs. The omission of grey literature is another significant concern. Grey literature may contain unpublished secondary outcomes, subgroup analyses, adverse event data, or important null results (35). Its absence can lead to a systematic overestimation of treatment benefits, obscure publication bias, and result in the omission of critical safety information.

In summary, methodologically flawed NMAs may lead clinicians and guideline developers to overinterpret imprecise or erroneous comparative findings. Therefore, enhancing the methodological rigor and reporting standards of network meta-analyses is crucial to ensure that treatment comparisons are not only statistically valid but also clinically reliable.

Several key methodological limitations were identified in the current landscape of network meta-analyses (NMAs). First, the majority of NMAs received low methodological quality ratings, primarily due to incomplete reporting of critical items. Notably, the processes of literature search, study selection, and the provision of a list of excluded studies with justifications were frequently omitted. Similarly, there was a marked deficiency in adhering to reporting guidelines specific to NMA. These shortcomings directly compromise the reliability of the evidence and may introduce bias into clinical decisions and guideline recommendations. Second, outcome selection remains heavily focused on glycemic, with relatively insufficient attention paid to cardiovascular, kidney, safety outcomes and other outcomes such as quality of life. This narrow focus may result in drug recommendation evidence that fails to fully capture a treatment’s risk profile, thereby increasing uncertainty in guideline development. Furthermore, most studies did not provide a rationale for exclusively including randomized controlled trials (RCTs), a critical component of methodological quality assessment. This omission may limit the generalizability of evidence to real-world clinical settings. Finally, nearly half of the reviewed studies lacked a preregistered protocol, which can lead to wasteful duplication of research efforts, increase the risk of data dredging, and directly undermine the study’s credibility.

Based on these findings, several practical strategies may enhance the methodological rigor of future NMAs in T2DM. First, routine protocol preregistration should be strongly encouraged to improve transparency and reduce selective reporting. Second, authors should explicitly justify their study design inclusion criteria and publish a complete list of excluded studies alongside the reasons for their exclusion. Third, reports should more clearly document the transitivity assumption, the assessment of inconsistency, and how the risk of bias from primary studies is considered within the network. Finally, guideline developers and journal editors could promote the use of structured methodological checklists to encourage more consistent reporting standards in NMAs. Implementing these approaches would help elevate the methodological and reporting quality of NMAs, thereby better serving evidence-based clinical decision-making.

In the univariate regression analysis, economic development level, journal impact factor, and protocol registration status demonstrated significant associations with methodological quality; These associations remained statistically significant after full adjustment for confounding factors in multivariate analysis. Notably, protocol registration exhibited a substantially stronger association with quality than other independent variables, and further analysis confirmed this relationship was independent of multicollinearity. This finding underscores the critical role of protocol registration in enhancing the methodological rigor and reporting quality of research. Publication year, JCR category, and number of included studies were associated with literature quality in the univariate regression analysis, but this association disappeared after full adjustment for confounding factors. This discrepancy may be due to the fact that in the univariate analysis, the association of publication year, JCR category, and number of included studies with quality essentially stemmed from high impact factors—rather than the independent variables themselves. Therefore, after adjusting for the confounding variable of impact factor, the correlation between these independent variables and literature quality vanished. The number of authors demonstrated no significant association with methodological quality in either univariate or multivariate analyses, suggesting that research team size does not substantially influence methodological quality in this domain.

Our analysis revealed a consistent annual increase in the number of network meta-analyses comparing the efficacy of antidiabetic drugs for type 2 diabetes mellitus. This trend likely reflects both growing public health concern regarding type 2 diabetes and the increasingly widespread adoption of network meta-analysis methodology. In addition, the number of studies from developing countries included in this study was greater than that from developed countries. Combined with the regression analysis result that a lower level of economic development may be associated with a higher study quality rating, this phenomenon may be influenced by factors such as regional variations in journal selection policies, editorial standards, preregistration practices, or literature inclusion criteria.

There are a large number of network meta-analysis studies on the efficacy comparison of antidiabetic drugs for type 2 diabetes, but no thorough evaluation of the methodological quality of these studies has been performed to date. This study represents the first comprehensive methodological assessment of network meta-analyses that compares the efficacy of antidiabetic drugs for type 2 diabetes mellitus. The investigation employed rigorous methodology with transparent procedures, yielding reliable findings that provide valuable reference for future NMAs in this field. Furthermore, our regression analysis identified specific factors influencing methodological quality, offering evidence-based guidance for quality improvement initiatives.

However, this study still has some limitations. First, the AMSTAR 2 tool utilized in our assessment was not specifically designed for NMA methodology, potentially limiting its ability to evaluate certain NMA-specific methodological characteristics. Consequently, our results may not fully capture the comprehensive methodology quality landscape of the included NMAs. Although multicollinearity diagnostics indicated no severe collinearity among variables, the presence of extremely large effect estimates with wide confidence intervals for certain predictors suggests potential bias or error. Second, in order to ensure the applicability of the research results, this study only included studies from SCI-indexed journals. This may lead to the omission of some studies with regional characteristics that have not been included in SCI journals, thus limiting the application range of the research conclusions to a certain extent. Finally, the scarcity of high-quality studies in our sample necessitates cautious interpretation when extrapolating conclusions to high-quality NMAs.

Future research should address these limitations by developing an NMA-specific methodological assessment tool to enhance evaluation precision. The inclusion of non-SCI journals and regionally representative studies would significantly strengthen the comprehensiveness of future assessments. Additionally, while we adjusted for common confounding factors, residual confounding remains possible; further investigation should explore additional determinants of methodological quality to inform more targeted improvement strategies.