Efficacy and safety of anti-prediabetic drugs in patients with prediabetes: a Bayesian network meta-analysis.

Background

Prediabetes refers to the transitional stage from normal glucose metabolism to diabetes, characterized by abnormal blood glucose levels, and is a precursor to type 2 diabetes mellitus (T2DM) [1]. The International Diabetes Federation (IDF) guidelines reported that, as of 2024, approximately 1.12 billion people globally were in the prediabetes stage, and this number will reach 1.49 billion by 2050 [2, 3]. Without intervention, pre-diabetic populations are highly likely to progress to T2DM [4]. Prediabetes significantly increases the risk of cardiovascular, cerebrovascular diseases, and cancer, etc., which threatens human health.

Effective interventions of prediabetes can significantly reduce the likelihood of progression to T2DM [5]. The interventions include lifestyle interventions (LFS) and pharmacological treatments [6]. Anti-prediabetic drugs include metformin [7], alpha-glucosidase inhibitors (e.g., acarbose) [8], sodium-glucose cotransporter 2 inhibitors (SGLT-2is, e.g., dapagliflozin), glucagon-like peptide-1 receptor agonists (GLP-1RAs, e.g., semaglutide) [8], glucose-dependent insulinotropic polypeptide (GIP)/GLP-1 dual receptor agonists (GIP/GLP-1RAs, e.g., tirzepatide), thiazolidinediones (TZDs, e.g., pioglitazone) [9], dipeptidyl peptidase-4 inhibitors (DPP-4is, e.g., sitagliptin) [10], orlistat [11] and vitamin D[12]. Although LFS is beneficial for most pre-diabetic populations [5], for those high-risk populations progressing to T2DM (e.g., severe obesity or blood glucose levels nearing the diagnostic threshold), LFS alone yield minimal and unsustainable effects. Consequently, medications are often favored due to a superior risk–benefit profile over LFS alone. However, the systematic, evidence-based evaluation of the efficacy and safety across these interventions is lacking.

As a recommended evidence-based analysis method, Bayesian network meta-analysis (BNMA) can offer a higher accuracy and more flexible modeling to provide high-quality evidence for clinical practice and decision-making. Thus, this study evaluated and compared the differences in efficacy and safety among various pre-diabetic interventions using BNMA, which provided evidence-based support for precision intervention in prediabetes, clinical decision-making, and the rational use of different anti-prediabetic drugs.

Methods

PubMed, Web of Science, Embase, Cochrane Library, and ClinicalTrials.gov databases were searched systematically for this study from November 1996 to March 2025. Medical Subject Headings (MeSH) and keywords such as “prediabetes,” “reverse,” and “delay” are used while searching. The search strategy, including all strings, is shown in the online Additional file 1: Table S1. The protocol was registered in the Prospective Register of Systematic Reviews (CRD42025636991). And this BNMA adheres to the preferred reporting items for systematic reviews and meta-analyses (PRISMA) extension statement for NMA (Additional file 1: Table S2) [13].

The assessment of studies’ eligibility was performed with the participants, interventions, comparison, outcomes, and study design (PICOS) criteria. Eligible randomized controlled trials (RCT) that met the following criteria were included: (1) Patients aged 18 ~ 75 years old with prediabetes (or high risk of diabetes), with hemoglobin A1c (HbA1c) 5.7–6.4%, or fasting plasma glucose (FPG) 5.6 mmol/L to 6.9 mmol/L (IFG), or with 2 h-postprandial plasma glucose (2 h-PPG) during 75-g oral glucose tolerance test (OGTT) 7.8 mmol/L to 11.0 mmol/L (IGT) [1], or IGT plus IFG. There are no gender or racial restrictions. (2) At least one of the 9 classes of potential anti-prediabetic drugs (metformin, GLP-1RAs, SGLT-2is, DPP-4is, GIP/GLP-1RAs, thiazolidinedione, orlistat, and vitamin D) was used as intervention. (3) Placebo or other appropriate intervention was used as a comparator in the control group. (4) Efficacy outcomes include reduction from baseline in HbA1c, FPG, weight loss (WL), and body mass index (BMI), and safety outcomes include incidence of total adverse events (AEs) and serious adverse events (SAEs).

Studies with those following features were excluded: (1) patients already diagnosed with diabetes mellitus (DM), aged < 18 or > 75, with certain comorbidities or complications; (2) incomplete data; (3) reviews, case reports, conferences, animal studies, phase Ⅰ studies, and non-English articles; (4) data from follow-up was excluded due to its high heterogeneity.

The retrieved items from database were imported into EndNote (version 21.4.0). After removing duplicates of research results, 2 investigators (YW1 and ZW) independently screened studies by title and abstract. The articles selected from the first stage were further reviewed in the full text. Only those that met all inclusion criteria were regarded as eligible. Any disagreement was resolved through discussion among reviewers and a third reviewer (AT).

Two investigators (YW1 and ZW) independently extracted data from all eligible studies. The data includes the following: (1) study characteristics (title of published article or trial name, registration number, first author, study design, phase of the trial, and study sites); (2) the population (sample size, age, gender, race, and baseline); (3) intervention (drug’s name and dose); (4) efficacy outcomes (WL, BMI, FPG, HbA1c, total cholesterol (TC), total triglyceride (TG), high-density lipoprotein (HDL), and low-density lipoprotein (LDL)); (5) safety outcomes (incidence of AEs, SAEs, renal and urinary disorders (RUD), and gastrointestinal disorders (GIAE)). Disagreement was discussed among the investigators and a third investigator (AT).

Each study’s risk of bias was assessed using the Cochrane Risk of Bias Tool (version 2.0) [14]. It consists of five domains, including the randomization process, deviations from the intended interventions, missing outcome data, measurement of outcomes, and selection of the reported result. Results of the assessment were shown by “high risk,” “concerned,” or “low risk” icons. Also, the CINeMA was used to grading the evidence of works in network meta-analysis [15, 16]. In addition, funnel plots were used to assess the presence of small study effects for each study species, and the tool used was Stata (version 17.0).

Eight efficacy outcomes are represented in the form of mean (standard deviation (SD)), and 4 safety outcomes are represented in the form of incidence (%). Investigators converted those reported in standard error (SE) or CI into SD according to the Cochrane Handbook [17]. The network plots and BNMA were created and performed by R (version 4.4.1) with its package BUGSnet (version 1.1.2). Mean difference (MD) and odds ratio (OR) values were used to calculate continuous and dichotomous variables, respectively. When constructing random-effects and fixed-effects models using the package, the parameters we selected are as follows: n.adapt = 5000, n.burnin = 3, n.iter = 20,000. Other settings were adjusted according to the data type (continuous or binary). To facilitate the assessment of clinical significance, the minimal clinically important difference (MCID) was employed as a threshold for comparison with the effect sizes derived from the analysis of continuous variables. Generally, a WL of 3% is considered clinically meaningful [18, 19], which is equivalent to a reduction of 2.7 kg based on the baseline calculation. Previous studies have confirmed a significant linear correlation between HbA1c and FPG [20]. Given the current absence of a consensus on the MCID for HbA1c in the prediabetes population, we have opted for an HbA1c MCID of 0.5%. This decision is informed by previous research findings and relevant guidelines [21, 22], while also taking into account the ease of interpretation of study outcomes.

Each outcome was measured by a random-effect consistency model, and the Markov chains with 20,000 simulations. R software with package BUGSnet was used to create the surface under the cumulative ranking (SUCRA) table, the Rankogram, and the league table. Interventions with a higher SUCRA value would be considered as more optimal treatments. And the forest plots were created using the R package meta (version 8.2–1).

As for the heterogeneity test, R software was used with the package BUSnet, gemtc (version 1.0–2), and JAGS (version 4.3.1). The heterogeneity of each study was assessed with I2. An I2 ≥ 50% or a p < 0.05 was considered a high between-study heterogeneity [23]. The difference of deviance information criterion (DIC) between the consistency and inconsistency models was used to assess the global consistency of the study, which was considered to show inconsistency with a difference > 5 [24, 25]. For those studies that showed potential inconsistency, the node-splitting model was used to detect. Finally, a sensitive analysis was conducted to assess the stability of results. Placebo was used as the comparison. Studies with high risk or heterogeneity, and studies lasting over 56 weeks were excluded for sensitive analysis, respectively.

Results

From the initial 6663 search results, 3516 duplicate records were first removed. Subsequently, through a process of screening titles and abstracts, followed by full-text review, 55 clinical trials that met our inclusion and exclusion criteria were ultimately included, and the characteristics of these studies were demonstrated in Additional file 1: Table S3 [26–80]. Figure 1 illustrates the entire process of literature screening for this systematic review. Finally, a total of 16,610 patients were included in this study, who received interventions across nine classes of drugs, including metformin, GLP-1RAs, SGLT-2is, DPP-4is, GIP/GLP-1RA, thiazolidinediones, orlistart, and vitamin D. Figure 2 presents the network diagrams for the eight efficacy outcomes and four safety outcomes, respectively.Fig. 1PRISMA 2020 flow diagram. RCT, randomized controlled trial. PICOS, the participants, interventions, comparators, outcomes and study design criteriaFig. 2Network plots comparing included interventions. Each node represents a clinical trial, with the size of the node indicating the sample size of the trial. The width of lines represents the amount of studies comparing each pair of treatments. A body mass index (BMI), B weight loss (WL), C fasting plasma glucose (FPG), D hemoglobin A1c (HbA1c), E total cholesterol (TC), F total triglyceride (TG), G high-density lipoprotein (HDL), H low-density lipoprotein (LDL), I adverse events (AEs), J serious adverse events (SAEs), K gastrointestinal disorders (GIAE), L renal and urinary disorders (RUD). Dap, dapagliflozin; Exe, exenatide; Aca, acarbose; Cof, cofrogliptin; Sit, sitagliptin; Sax, saxagliptin; Vid, vildagliptin; Ro, rosiglitazone; Pio, pioglitazone; Sem, semaglutide; Lir, liraglutide; Orl, orlistat; Tir, tirzepatide; Tro, troglitazone; Vit, vitamin D; Met, metformin; Lin, linagliptin; LFS, lifestyle intervention; PLA, placebo

PRISMA 2020 flow diagram. RCT, randomized controlled trial. PICOS, the participants, interventions, comparators, outcomes and study design criteria

Network plots comparing included interventions. Each node represents a clinical trial, with the size of the node indicating the sample size of the trial. The width of lines represents the amount of studies comparing each pair of treatments. A body mass index (BMI), B weight loss (WL), C fasting plasma glucose (FPG), D hemoglobin A1c (HbA1c), E total cholesterol (TC), F total triglyceride (TG), G high-density lipoprotein (HDL), H low-density lipoprotein (LDL), I adverse events (AEs), J serious adverse events (SAEs), K gastrointestinal disorders (GIAE), L renal and urinary disorders (RUD). Dap, dapagliflozin; Exe, exenatide; Aca, acarbose; Cof, cofrogliptin; Sit, sitagliptin; Sax, saxagliptin; Vid, vildagliptin; Ro, rosiglitazone; Pio, pioglitazone; Sem, semaglutide; Lir, liraglutide; Orl, orlistat; Tir, tirzepatide; Tro, troglitazone; Vit, vitamin D; Met, metformin; Lin, linagliptin; LFS, lifestyle intervention; PLA, placebo

Table S3 in the additional file 1 provides an overview of the 55 trials, including demographic information, interventions, and baseline characteristics. Among all studies with available data, the mean age of patients (SD) was 50.60 (6.17) years, with 8,802 (52.99%) female participants. Baseline BW, BMI, FPG, and HbA1c were 89.11 (15.00) kg, 30.81 (4.48) kg/m2, 106.40 (11.41) mg/dL, and 5.88 (0.73)%, respectively.

The overall results of the quality assessment are shown in Fig. 3. No studies were considered to have a high risk of bias in terms of selective reporting of results and deviations from intended interventions. In the other three domains, one study (1.8%) was rated as having a high risk of bias, respectively. Overall, four studies (7.3%) were classified as having a high risk of bias, 30 studies (54.5%) had a low risk of bias, and the remaining 21 studies (38.2%) were considered to have an unclear risk of bias. Among the four high-risk studies, two had potential biases in data measurement, one had incomplete data, and one had potential bias in the randomization process.Fig. 3Assessment of risk of bias. Created with Cochrane Risk of Bias Tool (version 2.0)

Assessment of risk of bias. Created with Cochrane Risk of Bias Tool (version 2.0)

Moreover, the results of grading the evidence by CINeMA were reported in Additional file 1: Table S4 and Fig. S1. It indicated that comparisons of DPP-4is or GLP-1RAs versus PLA achieved “high” or “moderate” confidence for selected outcomes, furnishing robust evidence for clinical decision-making. Remaining contrasts were downgraded to “low” or “very low” confidence, chiefly because of limited sample sizes, between-study heterogeneity and potential bias accumulated while pooling effect estimates. Thus, these findings should be interpreted cautiously.

A consistency random-effects model was employed for data analysis in this study. The results of SUCRA plot are presented in Fig. 4, the league table is displayed in Additional file 2: Table S1, and the plots of Rankogram is demonstrated in Additional file 1: Fig. S2. For each outcome analysis, studies with missing data were excluded in advance.Fig. 4Plots of SUCRA. The graph shows the cumulative probability of each intervention ranking. Each curve represents for an intervention. Interventions with higher SUCRA value show better efficacy or safety. A body mass index (BMI), B weight loss (WL), C fasting plasma glucose (FPG), D hemoglobin A1c (HbA1c), E total cholesterol (TC), F total triglyceride (TG), G high-density lipoprotein (HDL), H low-density lipoprotein (LDL), I adverse events (AEs), J serious adverse events (SAEs), K gastrointestinal disorders (GIAE), L renal and urinary disorders (RUD). SUCRA, surface under the cumulative ranking curve; Dap, dapagliflozin; Exe, exenatide; Aca, acarbose; Cof, cofrogliptin; Sit, sitagliptin; Sax, saxagliptin; Vid, vildagliptin; Ro, rosiglitazone; Pio, pioglitazone; Sem, semaglutide; Lir, liraglutide; Orl, orlistat; Tir, tirzepatide; Tro, troglitazone; Vit, vitamin D; Met, metformin; Lin, linagliptin; LFS, lifestyle intervention; PLA, placebo

Plots of SUCRA. The graph shows the cumulative probability of each intervention ranking. Each curve represents for an intervention. Interventions with higher SUCRA value show better efficacy or safety. A body mass index (BMI), B weight loss (WL), C fasting plasma glucose (FPG), D hemoglobin A1c (HbA1c), E total cholesterol (TC), F total triglyceride (TG), G high-density lipoprotein (HDL), H low-density lipoprotein (LDL), I adverse events (AEs), J serious adverse events (SAEs), K gastrointestinal disorders (GIAE), L renal and urinary disorders (RUD). SUCRA, surface under the cumulative ranking curve; Dap, dapagliflozin; Exe, exenatide; Aca, acarbose; Cof, cofrogliptin; Sit, sitagliptin; Sax, saxagliptin; Vid, vildagliptin; Ro, rosiglitazone; Pio, pioglitazone; Sem, semaglutide; Lir, liraglutide; Orl, orlistat; Tir, tirzepatide; Tro, troglitazone; Vit, vitamin D; Met, metformin; Lin, linagliptin; LFS, lifestyle intervention; PLA, placebo

A total of 33 studies involving 24 interventions were included in the BNMA of HbA1c. The pairwise comparisons of interventions presented in the league table (Additional file 2: Table S1) revealed differences in their effects. Compared with placebo, 14 interventions were found to significantly reduce HbA1c levels, with MD ranging from -0.01% to -0.93%. For example, 1.2 mg of liraglutide, 500 mg of metformin, 5 mg of linagliptin, and 2.4 mg of semaglutide reduced HbA1c by − 0.94% (− 1.34, − 0.54), − 0.75% (− 1.34, − 0.14), − 0.75% (− 1.47, − 0.01), and − 0.39% (− 0.55, − 0.25), respectively. Nineteen interventions showed no significant difference in HbA1c reduction compared with placebo, e.g., 2.5 mg of linagliptin plus 850 mg of metformin, 40 µg of vitamin D, and 80 µg of vitamin D, with MD values of 0.16% (− 0.15, 0.46), 0.10% (− 0.18, 0.38), and 0.10% (− 0.18, 0.38), respectively. The SUCRA values indicated that, 1.2 mg of liraglutide (97.41%) demonstrated the optimal HbA1c reduction, followed by 500 mg of metformin (91.65%), 5 mg of linagliptin (89.80%), 2.4 mg of semaglutide (84.06%), and 45 mg of pioglitazone (77.78%) (Additional file 1: Table S5).

A total of 45 studies involving 32 interventions were included in the BNMA of FPG. Studies 14 and 44 were excluded from the FPG analysis due to their high heterogeneity. As shown in Additional file 2: Table S1, most interventions reduced FPG (MD ranging from − 26.42 to − 0.15 mg/dL) compared with placebo. Fourteen interventions demonstrated a significant reduction in FPG, including 30 mg of pioglitazone, 15 mg of tirzepatide, 2.4 mg of semaglutide, 10 mg of tirzepatide, and 1.8 mg of liraglutide (MD (95%CI) − 26.42 (− 36.99, − 16.73), − 9.58 (− 12.00, − 7.15), − 9.13 (− 10.97, − 7.57), − 8.71 (− 11.20, − 6.22), and − 9.01 (− 12.93, − 5.11), respectively). However, several interventions did not show a significant reduction in FPG, such as 10 mg of dapagliflozin plus 2 mg of exenatide, 2.5 mg of linagliptin plus 850 mg of metformin, and 75 µg of vitamin D (MD (95%CI) − 1.13 (− 6.39, 4.19), 3.61 (− 0.68, 7.07), and 0.08 (− 4.88, 4.71), respectively). The SUCRA values for FPG reduction revealed that the most optimal intervention was 30 mg of pioglitazone (99.98%), followed by 15 mg of tirzepatide (90.60%), 2.4 mg of semaglutide (88.21%), 10 mg of tirzepatide (84.69%), and 1.8 mg of liraglutide (84.56%) (Additional file 1: Table S5).

The BNMA of BW involved 31 interventions across 29 studies. As shown in Additional file 2: Table S1, only 2.4 mg of semaglutide and 3 mg of liraglutide demonstrated a significant reduction in BW compared with placebo (MD (95%CI) − 13.89 (− 19.31, − 8.64) and − 5.99 (− 10.44, − 1.71)). Other interventions, such as 10 mg of dapagliflozin plus 2 mg of exenatide, 2000 mg of metformin, and 5 mg of linagliptin, showed a reduction in BW but did not reach statistical significance. The SUCRA values indicated that 2.4 mg of semaglutide (97.47%) exhibited the most optimal efficacy in WL, followed by 3 mg of liraglutide (82.25%), 10 mg of dapagliflozin plus 2 mg of exenatide (72.49%), 2000 mg of metformin (69.95%), and 1500 mg of metformin (69.45%). 850 mg of metformin (33.66%), 45 mg of pioglitazone (34.39%), and 1.8 mg of liraglutide (34.51%) showed the worst efficacy in WL (Additional file 1: Table S5).

A total of 35 studies involving 31 interventions were included in the BNMA of BMI. The league table (Additional file 2: Table S1) showed that 3 interventions significantly reduced BMI compared with placebo, including 15 mg of tirzepatide, 10 mg of tirzepatide, and 2.4 mg of semaglutide (MD (95% CI) − 4.50 (− 8.45, − 0.51), − 4.20 (− 8.17, − 0.21), and − 3.54 (− 5.82, − 1.31), respectively). The remaining 28 interventions did not show a significant reduction in BMI. The SUCRA values for BMI reduction indicated that the most effective interventions were 15 mg of tirzepatide (97.47%), followed by 10 mg of tirzepatide (82.25%), 2.4 mg of semaglutide (86.53%), 5 mg of tirzepatide (80.72%), and 1.2 mg of liraglutide (78.69%) (Additional file 1: Table S5).

In the BNMA of lipid profiles, three measures were included, namely TC, HDL, and LDL. These analyses involved 24, 22, and 17 studies, covering 23, 21, and 19 interventions, respectively. According to pairwise comparisons (Additional file 2: Table S1), 15 mg of tirzepatide, 10 mg of tirzepatide, and 5 mg of tirzepatide significantly reduced total cholesterol (MD 95% CI − 17.34 (− 27.76, − 7.10), − 13.71 (− 24.26, − 3.53), and − 12.74 (− 23.20, − 2.37), respectively). The SUCRA values for TC reduction showed that 15 mg of tirzepatide (89.31%) had the optimal efficacy in lowering lipid, followed by 10 mg (80.61%) and 5 mg (78.00%) of tirzepatide, 30 mg of pioglitazone (71.42%), and 2.5 mg of linagliptin plus 850 mg of metformin (71.42%) (Additional file 1: Table S5). However, no intervention showed a significant difference compared with placebo in altering HDL and LDL levels, e.g., 5 mg of linagliptin and 15 mg of tirzepatide (MD 95% CI of − 0.69 (− 2.75, 1.37) and − 0.50 (− 1.80, 0.78), respectively) in the LDL analysis, and 4 mg of rosiglitazone and 5 mg of linagliptin (MD 95% CI of 0.08 (− 0.16, 0.35) and 0.04 (− 0.50, 0.55), respectively) in the HDL analysis.

A total of 21 studies involving 21 interventions were included in the BNMA of triglycerides. According to the pairwise comparison results (Additional file 2: Table S1), 30 mg of pioglitazone demonstrated a significant reduction in triglycerides compared with placebo (MD (95% CI) − 6.16 (− 11.73, − 1.03)). However, the remaining interventions did not show a significant difference, e.g., 5 mg of saxagliptin, 5 mg of tirzepatide, and 1500 mg of metformin (MD (95% CI) − 2.18 (− 6.45, 1.74), − 1.69 (− 4.74, 1.26), and − 1.37 (− 5.60, 2.59), respectively). According to the SUCRA values, the most effective interventions for reducing triglycerides were 30 mg of pioglitazone (97.54%), followed by 5 mg of saxagliptin (85.75%), 5 mg of tirzepatide (80.34%), 10 mg of tirzepatide (77.41%), and 1500 mg of metformin (74.41%) (Additional file 1: Table S5).

The BNMA of AEs included 21 interventions, with the network diagram presented in Fig. 2I. Compared with placebo, no intervention showed a significant difference in the incidence of AEs, e.g., 4 mg of rosiglitazone and 500 µg of vitamin D (OR (95% CI) 0.06 (0.00, 2.20) and 0.00 (0.00, 8.64e + 10), respectively) (Additional file 2: Table S1). The SUCRA values indicated that 2.5 mg of linagliptin plus 850 mg of metformin (91.21%) had the lowest incidence of AEs, followed by 850 mg of metformin (71.06%) and 10 mg of dapagliflozin (65.61%). In contrast, 4 mg of rosiglitazone (22.59%), 40 µg of vitamin D (22.88%), and 100 mg of sitagliptin (26.87%) had the highest incidence of AEs (Additional file 1: Table S5).

A total of 24 studies involving 18 interventions were included in the BNMA of SAEs. Pairwise comparison results (Additional file 2: Table S1) showed that, compared with placebo, 100 mg of sitagliptin had a higher incidence of SAEs (OR (95% CI) 1.70e + 07 (17.63, 2.53e + 18)), while the remaining interventions did not show significant differences. According to the SUCRA values, 100 mg of sitagliptin showed the lowest incidence of SAEs (96.18%), followed by 1.8 mg of liraglutide (85.76%) and 45 mg of pioglitazone (59.80%). By contrast, 500 µg of vitamin D (24.29%), 10 mg of cofrogliptin (25.12%) and 75 µg of vitamin D (34.44%) had a higher incidence of SAEs (Additional file 1: Table S5).

A total of 8 studies involving 9 interventions were included in the BNMA of GIAE. Pairwise comparison results showed that, compared with placebo, only 45 mg of pioglitazone decreased the incidence of GIAE, other interventions showed no significant result, e.g., 10 mg of dapaglifrozin plus 2 mg of exenatide and 50 mg of vildagliptin (OR (95% CI), 2.64 (0.03, 237.46) and 2.44 (0.02, 523.22)) (Additional file 2: Table S1). The SUCRA values indicated that 5 mg of tirzepatide (24.23%) had the highest incidence of GIAE, followed by 500 μg of vitamin D (39.62%), 3 mg of liraglutide (39.62%), 10 mg of tirzepatide (40.76%) and 15 mg of tirzepatide (47.82%) (Additional file 1: Table S5).

To ensure the formation of the network, data from Study 5 were temporarily excluded. Therefore, a total of 8 interventions were included in the analysis of RUD. According to the league table results (Additional file 2: Table S1), 10 mg of dapagliflozin had a lower incidence of RUD (OR (95% CI) 0.00 (0.00, 0.16)), with the SUCRA value of 91.76%. However, other interventions did not show significant differences.

As shown in Additional file 1: Fig. S3, consistency checks for efficacy outcomes were conducted. For each outcome, the difference in DIC between the consistency and inconsistency models was less than 5, indicating high global consistency among the included studies for these outcomes. However, due to the unique nature of event counts and limited sample sizes, DIC checks could not be performed for safety outcomes. Additionally, the node-splitting method was employed to assess local inconsistencies. For LDL, SAEs, GIAE, and RUD, suitable comparison groups could not be identified within the network for inconsistency analysis. For BMI, WL, HbA1c, TC, TG, HDL, and AEs, the p-values obtained from the node-splitting method were all greater than 0.05 (Additional file 1: Table S6), indicating good local consistency. However, for FPG, local inconsistencies were observed in some comparisons, such as 10 mg of dapagliflozin versus 850 mg of metformin (p = 0.02) and 10 mg of dapagliflozin versus placebo (p = 0.03). The results of the node-splitting method are presented in Additional file 1: Table S6. The results of I2 test of heterogeneity were shown in Additional file 1: Table S7. Some comparisons exhibited notable heterogeneity, e.g., 100 mg of acarbose versus placebo, 10 mg of dapagliflozin versus 850 mg of metformin, and lifestyle intervention versus 850 mg of metformin.

According to the results of the heterogeneity analysis, two studies (Study 14 and 44) that contributed significantly to heterogeneity in the FPG outcome analysis were excluded. Subsequently, sensitivity analyses for 12 outcomes were conducted, and the results were shown in Additional file 1: Table S8.

First, studies with a high risk of bias (Study 7, 8, 18, and 48) were removed. The number of interventions included in each outcome was as follows: 31 for BMI, 29 for WL, 32 for FPG, 24 for HbA1c, 23 for TC, 21 for TG, 21 for HDL, 19 for LDL, 21 for AEs, 18 for SAEs, 9 for GIAE, and 8 for RUD. Apart from 1.8 mg of liraglutide showing a higher risk of SAEs, most data remained unchanged after excluding studies with a high risk of bias.

Additionally, considering that differences in the study duration could affect model robustness, studies with a duration exceeding 56 weeks (Study 2, 3, 7, 21, 22, and 49) were excluded. The results indicated that, although some interventions, such as 5 mg of tirzepatide in BMI and TC, 10 mg of dapagliflozin in FPG, and 1.8 mg of liraglutide in inducing SAEs, were slightly affected, the vast majority of data remained unchanged and had little effect on the conclusions. Overall, the results of the two sensitivity analyses were largely consistent with the main analysis, indicating good model robustness.

Funnel plots for publication bias across the included studies for each outcome were shown in Fig. 5. Due to limited sample sizes, publication bias analysis could not be conducted for GIAE and RUD. Most funnel plots showed symmetry, suggesting the absence of small-study effects. However, some publication bias was still observed in outcomes such as FPG.Fig. 5Funnel plots of efficacy outcomes. The dashed lines represent the 95% confidence intervals. A body mass index (BMI), B weight loss (WL), C fasting plasma glucose (FPG), D hemoglobin A1c (HbA1c), E total cholesterol (TC), F total triglyceride (TG), G high-density lipoprotein (HDL), H low-density lipoprotein (LDL), I adverse events (AEs), J serious adverse events (SAEs)

Funnel plots of efficacy outcomes. The dashed lines represent the 95% confidence intervals. A body mass index (BMI), B weight loss (WL), C fasting plasma glucose (FPG), D hemoglobin A1c (HbA1c), E total cholesterol (TC), F total triglyceride (TG), G high-density lipoprotein (HDL), H low-density lipoprotein (LDL), I adverse events (AEs), J serious adverse events (SAEs)

Discussion

The present results indicated that, compared with placebo, some antidiabetic drugs significantly reduce HbA1c, FPG, BW, and BMI in the prediabetic patients, demonstrating a notable efficacy in reducing the risk of prediabetes progressing to T2DM. Moreover, they did not significantly increase the incidence of AEs, SAEs, GIAE, or RUD, suggesting acceptable safety. In short, the GLP-1RAs and GIP/GLP-1RAs demonstrated the most favorable combined efficacy in weight reduction and glycemic control. In terms of safety, these drugs may be associated with an increased incidence of GIAEs compared with placebo. Additionally, TZDs had shown relatively better performance in lowering blood glucose levels. However, previous researches [81, 82] indicated that TZDs carry a significant risk of inducing AEs such as heart failure and fractures. Therefore, this study did not recommend TZDs as a first-choice for prediabetes. SGLT-2is, DPP-4is, and orlistat showed relatively poor anti-prediabetic efficacy compared with GLP-1RAs, GIP/GLP-1RAs, and TZDs in this study. Moreover, despite that vitamin D had a high safety profile, its contribution to glycemic control was minimal.

Semaglutide, liraglutide, and tirzepatide demonstrated potent anti-prediabetic efficacy in all the included interventions. Of them, 1.2 mg of liraglutide showed the optimal efficacy in HbA1c reduction and exceeding the MCID threshold of 0.5%. Fifteen milligrams of tirzepatide showed the optimal efficacy in reducing BMI and FPG. Notably, 2.4 mg of semaglutide stood out in WL, with an effect size (MD (95% CI) − 13.59 (− 17.30, − 9.91)) exceeding the MCID threshold of 2.7 kg.

Additionally, tirzepatide demonstrated superior lipid-lowering efficacy compared with semaglutide (Additional file 1: Table S5). Given that most patients with prediabetes are overweight or obese, GLP-1RAs and GIP/GLP-1RAs (e.g., semaglutide, liraglutide, and tirzepatide) were preferable since they offer significant WL and lipid-lowering benefits alongside glycemic control. The Standards of Care in Diabetes 2025 published by ADA [18] indicated that BW management was very important for patients with overweight or obesity, and a 3–7% BW reduction from baseline could lower the risk of diabetes and related diseases. However, compared with other interventions, tirzepatide exhibited a higher risk of GIAEs and other AEs, and should be monitored carefully. In addition, in recent years, evidence has emerged that GLP-1RAs have a cardiovascular and renal function protection effect in T2DM patients [83].

TZDs, a class of insulin sensitizers, reduced blood glucose by enhancing peripheral tissue sensitivity to insulin and improving insulin resistance. The current results showed that 30 mg of pioglitazone outperformed other drugs in reducing FPG, and had a superior efficacy in reducing triglyceride and cholesterol levels compared with 4 mg of rosiglitazone. Dutta et al. [84] found that TZDs had triglyceride reduction benefits, which were consistent with our results. This suggested that TZDs not only reduced the risk of diabetes onset but may offer cardiovascular benefits. However, the BNMA results indicated that pioglitazone and rosiglitazone may lead to weight gain, and rosiglitazone had a higher probability of AEs. Given the risk–benefit profile of interventions for prediabetes, TZDs were not recommended as the first-line option in the treatment of prediabetes.

Metformin, a well-established insulin sensitizer, has been shown to prevent diabetes onset in previous studies [85, 86]. Our study indicated that metformin significantly reduced HbA1c but did not demonstrate a marked advantage in other outcomes. Echouffo-Tcheugui et al. [2] found that, while metformin delays prediabetes, its benefits were less pronounced than those of LFS. Although few AE reports were collected for metformin in the present study, several RCTs had highlighted its high risk of GIAE [85]. Additionally, acarbose did not show significant effects on glycemic control, WL, or triglyceride reduction.

The BNMA results indicated that some DPP-4is, particularly 5 mg of linagliptin, perform well in WL and glycemic control. For SGLT-2i dapagliflozin, it did not show significant efficacy in WL. However, 10 mg of dapagliflozin plus 2 mg exenatide demonstrated a notable advantage in protecting renal function, which was consistent with a previous study [87]. Meanwhile, many studies suggested SGLT-2is played an important role in reducing kidney damage when treating diabetes [88]. However, dapagliflozin did not demonstrate a significant renal protective effect in the present study. This discrepancy was possibly attributed to the limited number of studies and sample sizes for the intervention of dapagliflozin in our analysis or potential publication bias in the included studies, warranting further investigation.

Recent studies [89, 90] suggested that vitamin D may have potential antidiabetic effects. However, the BNMA results indicated that vitamin D supplementation at various doses did not significantly reduce FPG, HbA1c, BMI, or BW. Some researchers also suggested that vitamin D deficiency was common in obese individuals [91] and may affect insulin sensitivity. However, the number of studies on vitamin D intervention for diabetes progression remains limited, with most results lacking statistical significance and evidence support [90]. Consequently, there is still a need for more robust evidence to substantiate the role of vitamin D in the treatment of prediabetes.

Additionally, according to the research conducted by Torgerson et al. [11], the WL medication orlistat had been shown to reduce the risk of prediabetes converting to T2DM. However, the BNMA results indicated that orlistat did not significantly reduce FPG and HbA1c in the prediabetic patients compared with other agents.

LFS is an important strategy for the treatment of prediabetes [92]. However, the BNMA results suggested that LFS showed no significant WL, glycemic control, or lipid-lowering effects compared with other included interventions, implying limited efficacy in treating prediabetes. From a purely efficacy perspective, it is reasonable to posit that drug interventions tend to exhibit greater efficacy compared to LFS. That is, while LFS is effective for many individuals with prediabetes as presented by the guidelines, for those who are severely obese or have blood glucose levels nearing the diagnostic threshold for T2DM, pharmacological interventions such as GLP-1RAs, metformin or TZDs may be more advantageous after a careful risk–benefit analysis [5, 93].

Additionally, from a pathophysiological perspective: Individuals with IFG primarily exhibit hepatic insulin resistance with normal muscle insulin sensitivity. The combination of hepatic insulin resistance and insulin secretion deficiency leads to elevated fasting blood glucose levels. In contrast, IGT patients generally have normal hepatic insulin sensitivity, but muscle insulin resistance can range from moderate to severe [94]. Considering the impact of age on islet function, some studies advocate for personalized pharmacological interventions for prediabetes [95]. For instance, LFS may be more effective for elderly IFG patients with lower insulin sensitivity. Patients with both IFG and IGT, who have lower β-cell function, could benefit from therapeutic strategies aimed at enhancing insulin secretion, such as DPP-4is. Younger IGT patients, despite having hepatic insulin resistance, tend to have relatively preserved insulin secretion capacity and may be treated with metformin and TZDs.

Overall, for the prediabetic patients at high risk of progressing to T2DM, drug interventions can be considered to delay weight gain and glycemic elevation. For obese individuals, GLP-1RAs and GIP/GLP-1RAs (e.g., semaglutide, tirzepatide, and liraglutide) not only directly control blood glucose levels but also promote WL, which reduces the risk of prediabetes progressing to diabetes and its complications. TZDs (e.g., pioglitazone and rosiglitazone) demonstrate potent efficacy in treating prediabetes, potentially offering cardiovascular benefits. Given that prediabetes is a high-risk factor for cardiovascular disease, TZDs may have significant potential in cardiovascular protection, as highlighted by the ADA [5]. In addition to drug interventions, LFS is also a valuable anti-prediabetic strategy. Although its objective data may not be as pronounced as some medications, its long-term, stable benefits, which do not dissipate with discontinuation, make it a widely recommended approach.

This study still has some limitations. First, due to the limited number of clinical trials related to prediabetes, the number of RCTs included for each class of drugs was restricted, and the span of study initiation years was extensive. Secondly, when analyzing efficacy and safety outcomes, some studies had small sample sizes and exhibited publication bias or high heterogeneity. As can be observed from the forest plots (Additional file 1: Fig. S4) and the results of heterogeneity and inconsistency detection (Additional file 1: Tables S6 and S7), certain studies exhibit heterogeneity with respect to outcome measures such as FPG, HbA1c, and AEs. The reasons for this heterogeneity may include differences in the types of medications used, methods of outcome assessment, and the ethnic composition of the study populations. Therefore, sensitivity analyses were conducted, excluding high-risk or long-term studies, and the conclusions remained largely consistent. Studies with excessive heterogeneity that could not be resolved were excluded. However, differences in the countries and regions where the studies were conducted, population ethnicity, baseline characteristics, and placebo use could still impact the results. Particularly for safety outcomes, there was a higher incidence of missing data, and the total number of studies included for outcomes such as GIAE and RUD was less than 10. Also, there were several interventions with very limited sample sizes. This could potentially lead to small-study effects that highly bias the results. Thirdly, the dosing methods and frequencies for each drug were not uniform, and the study durations varied significantly, ranging from 12 to 120 weeks. These factors could potentially increase the heterogeneity of the studies. But in the sensitivity analysis, it was demonstrated that duration had minimal impact on the robustness of the model. However, in order to cover as many anti-prediabetic drugs as possible, some clinical trial results that exhibited slight heterogeneity were retained. Given the diverse range of drugs included in this study, it was believed that the observed heterogeneity is unavoidable and acceptable, but some results need to be interpreted with caution. For instance, 2.4 mg of semaglutide SC achieved a high SUCRA value in the GIAE analysis, which is inconsistent with the known AE incidence of semaglutide. This discrepancy may stem from the impact of Study 2’s findings and the limited number of studies included, as well as the potential impact of zero-event data on transmissibility hypotheses (Additional file 1: Fig. S4K also suggests this). Additionally, 1.2 mg of liraglutide rests on a single trial (Study 4). The limited evidence base and modest risk-of-bias concerns may overstate its efficacy in prediabetes, so the findings should be interpreted cautiously. Future research should incorporate additional efficacy and safety data on prediabetes to enhance the precision of the analysis. And discrepancies in the analysis of indicators assessing the same physiological condition, such as BMI and BW, TC and TG, may stem from: BW not accounting for factors like height as BMI does; heterogeneity among studies due to variations in research years, timeframes, and analytical methodologies; and a limited number of studies including certain interventions. Moreover, a noticeable divergence was observed between the forest-plot estimates and the SUCRA rankings for safety outcomes. This discordance is attributable to the intrinsic methodological differences: SUCRA integrates the entire network of direct and indirect evidence, whereas the forest plot isolates pairwise contrasts and is therefore more vulnerable to extreme values and statistical heterogeneity. Precisely because of this focus, it remains instrumental in pinpointing potential sources of inconsistency.

Also, LFS are among the most fundamental and significant interventions for both prediabetes and diabetes. However, due to the limited duration of clinical trials included in this study (up to 120 weeks), it is challenging to observe the long-term benefits of LFS, potentially underestimating its role in delaying T2DM. To compare the effects of lifestyle intervention and pharmacological interventions on reversing prediabetes, future research should include more studies with longer durations. Furthermore, due to data limitations, we were unable to assess the cardiovascular event benefits of pharmacological interventions in the prediabetes population, which is currently believed to be highly associated with the intrinsic effects of the medications.

Moreover, most of the included RCTs were placebo-controlled rather than head-to-head comparisons, resulting in a lack of direct comparative evidence. Therefore, there is an urgent need for larger-scale, higher-quality head-to-head RCTs to make the conclusions more accurate and robust.

Conclusion

In summary, GLP-1RAs, GIP/GLP-1RAs and TZDs exhibit favorable anti-prediabetic efficacy and acceptable safety profiles. 2.4 mg of semaglutide SC, 15 mg of tirzepatide, and 1.2 mg of liraglutide were the best option among the included interventions, considering favorable BMI and glycemic control.