Risk of Stomach Cancer with DPP-4 Inhibitors, GLP-1 Receptor Agonists, and SGLT2 Inhibitors: Comparison Using Clinical Trial Data

The PRISMA flow diagram is presented in Figure 1. After duplicate removal and exclusion of 317 records during screening (Table S4), 52 RCTs met eligibility criteria and were included in the final evidence network (Table S5) [17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68]. Across these trials, 171,165 participants were enrolled (mean age 63.6 years; 36.9% women), with a mean follow-up of 141.8 weeks, providing substantial person-time for the identification of incident gastric neoplasm events.

The treatment network included:

Although certain DPP4is (omarigliptin and teneligliptin), GLP1RAs (albiglutide, efpeglenatide, and orforglipron), triple GIP/GLP-1/glucagon receptor agonists (e.g., retatrutide), mitochondrial bioenergetics modulators (e.g., imeglimin), glucokinase activators (e.g., dorzagliatin), amylin analogues (e.g., petrelintide), and anti-CD3 monoclonal antibodies (e.g., teplizumab) were prespecified in the search scope, eligible RCTs did not report gastric tumor outcomes in a manner permitting quantitative synthesis.

In the primary class-level NMA, GLP1RAs (event/subjects = 0.09%) were associated with a significantly lower risk of gastric neoplasm compared with controls (event/subjects = 0.13%) (RR = 0.51, 95%CIs = 0.28 to 0.92). In contrast, DPP4is (event/subjects = 0.21%) were associated with a significantly higher risk of gastric tumor (benign or malignant) than controls (RR = 1.77, 95%CIs = 1.09 to 2.85). DualRA (event/subjects = 0.16%), SGLT2is (event/subjects = 0.14%), and SGLT2i+DPP4is (event/subjects = 0.28%) did not have a statistically significant difference in comparison with controls (Figure 2A, Figure 3A and Figure S3). Additionally, GLP1RAs were associated with lower gastric neoplasm risk than SGLT2is (RR = 0.49, 95%CIs = 0.24 to 0.98) and DPP4is (RR = 0.29, 95%CIs = 0.14 to 0.61) (Table 1).

When the network was examined at the regimen level, no individual regimen showed a statistically significant difference in gastric neoplasms risk compared with control (Figure 2B, Figure 3B, and Table 2).

.

When analyses were restricted to RCTs enrolling participants with type 2 diabetes mellitus, findings were consistent with the overall network: GLP1RAs remained associated with a significantly lower risk of gastric tumor versus controls (RR = 0.51, 95%CIs = 0.28 to 0.92), whereas DPP4is remained associated with a significantly higher risk versus controls (RR = 1.76, 95%CIs = 1.09 to 2.84) (Figure 2C and Figure 3C, and Table 3).

On the other hand, the subgroup analysis of obesity or chronic renal insufficiency could not be conducted due to the lack of a sufficient dataset.

Among trials with study duration ≥ 52 weeks, GLP1RAs were again associated with lower gastric tumor risk compared with controls (RR = 0.52, 95%CIs = 0.28 to 0.95), while DPP4is were associated with higher risk compared with controls (RR = 2.05, 95%CIs = 1.19 to 3.55) (Figures S1A and S2A, and Table S6A).

In contrast, among trials with study duration < 52 weeks, no regimen demonstrated a statistically significant difference versus controls (Figures S1B and S2B and Table S6B).

In trials enrolling older populations (mean age ≥ 60 years), DPP4is were associated with a significantly higher gastric tumor risk than controls (RR = 2.08, 95%CIs = 1.15 to 3.77), whereas GLP1RAs were associated with a significantly lower gastric tumor risk than DPP4is (RR = 0.24, 95%CIs = 0.10 to 0.61) (Figures S1C and S2C, and Table S6C).

In trials enrolling younger populations (mean age < 60 years), no regimen demonstrated a statistically significant difference versus controls (Figures S1D and S2D, and Table S6D).

In the secondary outcome, no regimen demonstrated a significantly different risk of Helicobacter pylori infection versus controls (Figures S1E and S2E, and Table S6E).

For acceptability, lower drop-out rates were observed for SGLT2is plus DPP4is, DPP4is, and SGLT2is compared to controls (). Figures S1F and S2F, and Table S6F

When the network was examined at the dose level, no individual regimen at any dosage showed a statistically significant difference in gastric neoplasms risk compared with controls, which might have resulted from the widening confidence intervals related to reduced sample sizes (and S2G). Figures S1G

Comparison-adjusted funnel plots did not suggest marked asymmetry (Figure S4), and Egger’s tests were not statistically significant (Figure S5), indicating limited evidence of small-study effects or publication bias. SUCRA rankings are provided in Table S7. Between-study heterogeneity (τ²) was generally modest (Table S8). Inconsistency assessments using node-splitting, loop-specific, and design-by-treatment interaction approaches did not reveal substantial disagreement between direct and indirect evidence (Table S9A,B), supporting network coherence. Bayesian sensitivity analyses were aligned with frequentist findings: GLP1RAs were associated with lower gastric tumor risk versus controls, whereas DPP4is were associated with higher risk versus controls. By better accommodating zero-event trials, Bayesian models reinforced robustness under sparse-event conditions (Figure S6).

Among the 52 included RCTs, 32/52 (61.5%) were judged low risk of bias, 16/52 (30.8%) raised some concerns, and 4/52 (7.7%) were judged high risk using RoB 2.0 (). Using GRADE adapted for NMA, certainty was rated as moderate to high for most comparisons involving gastric tumor outcomes (). Figure S7 Table S10

This NMA has several strengths. First, by synthesizing evidence from 52 RCTs and 171,165 participants, we addressed a clinically consequential safety question that individual trials are seldom powered to resolve. Second, restricting inclusion to adults without baseline gastric tumors and focusing on incident events improves interpretability for primary prevention and reduces bias introduced by pre-existing disease. Third, prespecified subgroup analyses (diabetes mellitus status, study duration, and age) enhance clinical utility because real-world treatment decisions are often shaped by patient risk profiles and anticipated exposure duration. Fourth, methodological rigor was reinforced through protocol registration, PRISMA-NMA reporting, RoB 2.0 assessment, GRADE certainty ratings, and complementary frequentist and Bayesian modeling—an approach that strengthens transparency, reproducibility, and robustness, especially under rare-event conditions. Finally, the class-level estimates (GLP1RA: RR = 0.51, 95%CIs = 0.28 to 0.92; DPP4i: RR = 1.77, 95%CIs = 1.09 to 2.85) provide clinically interpretable signals that justify both clinical awareness and further confirmatory research.

Several limitations deserve attention. Evidence for newer or emerging agents (e.g., omarigliptin, teneligliptin, orforglipron, retatrutide, imeglimin, dorzagliatin, and teplizumab) was insufficient for inclusion, limiting generalizability to those regimens. Many included trials were optimized for metabolic and cardiovascular endpoints; systematic gastric tumor screening, centralized adjudication, intense surveillance, and standardized ascertainment were uncommon. This issue would be important because several important risk factors, such as smoking prevalence, obesity severity, gastroesophageal reflux, and metabolic comorbidity burden, would be tightly linked to gastric cancer risk. Although the statistical inconsistency test and heterogeneity test revealed insignificance, those clinical variables might still pose potential confounding effects on the main result of our NMA. Further, the sparse events of gastric tumors across the recruited RCTs would pose a potential risk of possible influence of a few large, long trials on DPP4i estimates. Accordingly, outcome misclassification and between-trial variability in event capture are possible, particularly where gastric tumors were recorded as adverse-event codes rather than prespecified trial endpoints. Besides, the mean follow-up of 141.8 weeks (around 2.7 years) might be insufficiently long for carcinogenesis inference. Although we arranged subgroup analysis with a cut-off point of 52 weeks of exposure time, the “52 weeks” were still insufficiently long regarding tumor formation. In addition, although we explored heterogeneity, inconsistency, and small-study effects, residual uncertainty is unavoidable when event rates are low, and statistically significant signals should be interpreted with appropriate clinical restraint. Finally, the indistinguishability of benign/malignant tumors might limit the clinical interpretation of the result of the current NMA. Despite our efforts to classify the benign versus malignant outcomes, the number of RCTs clearly reporting gastric benign/malignant tumor outcomes was too sparse to conduct network analysis.

We conducted a prespecified, hypothesis-driven network meta-analysis focused on incident gastric tumor outcomes potentially associated with newer glucose-lowering therapies, following Cochrane guidance for evaluating harms in randomized trials [77]. Reporting adhered to the PRISMA extension for network meta-analyses (PRISMA-NMA; Table S1) [78]. The protocol was registered prospectively in PROSPERO (CRD420261282728) and approved by the Institutional Review Board of Tri-Service General Hospital, National Defense Medical Center, Taipei, Taiwan (TSGHIRB E202516007).

We implemented a comprehensive search across eight sources—PubMed, Embase, ClinicalKey, Cochrane CENTRAL, ProQuest, ScienceDirect, Web of Science, and ClinicalTrials.gov—from inception to 10 January 2026 (). Search strategies combined controlled vocabulary and free-text terms for each drug class and individual agents, with terms capturing gastric tumor outcomes and relevant adverse-event categories. Two reviewers (PT Tseng and YW Chen) independently screened titles and abstracts and then assessed full texts for eligibility. Disagreements were resolved through discussion, with third-reviewer arbitration when needed. We also hand-searched reference lists of relevant systematic reviews and meta-analyses to identify additional eligible trials. No restrictions were applied by language, publication status, or geographic setting. Table S2

Eligibility criteria were prespecified using a PICOS framework.

Regimens of interest included DPP4is, GLP1RAs, SGLT2is, dual GIP/GLP-1 receptor agonists (e.g., tirzepatide), triple GIP/GLP-1/glucagon receptor agonists (e.g., retatrutide), mitochondrial bioenergetics modulators (e.g., imeglimin), glucokinase activators (e.g., dorzagliatin), amylin analogues (e.g., petrelintide), and anti-CD3 monoclonal antibodies (e.g., teplizumab). Trials were eligible if they (a) enrolled adults without documented baseline gastric tumors, (b) evaluated at least one prespecified agent, (c) systematically collected and reported adverse events including gastric tumors, and (d) maintained structured safety surveillance throughout follow-up [79]. We excluded (a) non-randomized studies; (b) RCTs restricted to participants with pre-existing gastric tumors; (c) trials lacking an appropriate comparator group; (d) RCTs that did not report any gastric tumor outcomes; (e) pediatric-only trials; (f) animal or preclinical investigations; and (g) trials with clearly compromised randomization procedures or major baseline imbalances in core characteristics (e.g., age, sex, comorbidities, or concomitant therapies).

Two reviewers independently assessed each included RCT using the Cochrane Risk of Bias 2.0 tool [80]. We evaluated bias arising from the randomization process, deviations from intended interventions, missing outcome data, outcome measurement, and selective reporting. Each domain—and the overall trial—was classified as low risk, some concerns, or high risk. Discrepancies were resolved by consensus, with third-reviewer input when necessary.

The primary endpoint was incident gastric tumors, defined as the first occurrence of any gastric tumor event reported during trial follow-up. We acknowledge that histopathologic terminology, diagnostic thresholds for early lesions (e.g., high-grade dysplasia vs intramucosal carcinoma) [81], and coding/reporting practices for tumors with malignant potential (e.g., GIST [82,83] and neuroendocrine tumors [84,85]) have evolved across countries and decades. Because many eligible trials reported gastric events only as ‘gastric tumor/neoplasm’ without consistent adjudication or benign/malignant labeling, we did not impose restrictions on diagnostic criteria and analyzed overall gastric neoplasms as a composite endpoint to reduce differential misclassification across the network. When statistically meaningful associations were observed at the class or regimen levels, we performed dose-stratified analyses based on protocol-defined dosage categories (Table S3).

As a secondary outcome, we investigated the risk of Helicobacter pylori infection, which was one of the most important risk factors for gastric tumors. Further, for a patient-relevant measure of treatment acceptability, we assessed all-cause discontinuation (drop-out rate). This choice was prespecified and aligned with our prior large-scale NMA work in related therapeutic domains [4,12,13]. When statistically meaningful signals emerged for the primary or secondary outcomes, we conducted prespecified subgroup analyses to test consistency in groups most directly exposed in routine care: (a) RCTs enrolling participants with type 2 diabetes mellitus (the principal indication for these agents), obesity, or renal insufficiency, (b) trials with study duration stratified as ≥52 weeks versus <52 weeks, and (c) trials stratified by mean participant age as ≥60 years versus <60 years. We chose 60 years as the cut-off point of age based on the previous study [86], in which the authors reported a significantly worse survival rate of gastric tumor in subjects as ≥60 years.

Study selection proceeded in two stages: title/abstract screening followed by full-text confirmation. Two authors (PT Tseng and YW Chen) independently extracted data using a standardized form capturing trial design, follow-up duration, sample size, setting, participant characteristics, eligibility criteria, treatment regimens, comparator types, target outcomes, and drop-out rates. When outcome reporting was incomplete or unclear, we contacted corresponding authors up to two times. Data handling followed the Cochrane Handbook and related methodological standards to maximize transparency, auditability, and reproducibility [87].

For dichotomous outcomes, we calculated risk ratios (RRs) with 95% confidence intervals (95%CIs). Primary analyses used a random-effects NMA framework implemented with the “network” suite in STATA 16.0 (StataCorp, College Station, TX, USA) [88]. These contrast-based models integrate direct and indirect comparisons across the evidence network to estimate relative effects among all included regimens [89]. To summarize comparative rankings, we calculated the surface under the cumulative ranking curve (SUCRA), where higher values reflect a greater probability of being among the most favorable options [90]. Between-study heterogeneity was summarized using τ². We assessed network consistency using complementary approaches, including loop-specific methods, node-splitting, and design-by-treatment interaction models [91]. Certainty of evidence was evaluated using GRADE adapted for NMA, incorporating considerations of study limitations, inconsistency, indirectness, imprecision, and publication bias [92]. Small-study effects and publication bias were explored using comparison-adjusted funnel plots and Egger’s regression tests.

Because gastric tumor events are uncommon in RCTs, sparse-event bias and zero-event cells are plausible. Therefore, we also conducted Bayesian NMA using the “multinma” package in R as a sensitivity analysis [93]. This approach is well-suited to rare outcomes, can reduce reliance on arbitrary continuity corrections in zero-event settings, and offers an additional robustness check under alternative modeling assumptions.

These data were derived from the following resources available in the public domain: https://clinicaltrials.gov/↗ (access date: 10 January2026).