We conducted an initial search of electronic databases through PubMed and Embase on 11 November 2022. A subsequent search update was performed on 20 March 2024, utilizing the identical databases for consistency. Our search employed targeted keywords pertinent to “metformin,” “colorectal cancer,” and “colorectal adenoma,” with the specific methodologies detailed in Supplementary Table S2. Titles and abstracts were screened to eliminate irrelevant studies, followed by a full-text review to evaluate eligible articles. The reference lists of each primary eligible study, along with previous systematic reviews, were scanned to identify additional relevant studies. Each stage of study selection, data extraction, and study quality assessment was independently conducted by the same two investigators (Mengdan, Shen and Shan, Lu). Any discrepancies in the above process finally reached a consensus through a negotiation conference with a third reviewer (Li-Ting, Sheng).

Studies were included if they met the following criteria: (1) observational studies (cohort, case–control or nested case–control) or randomized controlled trials; (2) studies that exposure to any dose or duration of metformin with no use of metformin or other anti-diabetic drugs for comparison; (3) studies that listed occurrence of incident colorectal adenoma and colorectal cancer as an outcome of interest; (4) studies that reported the risk point estimate as an adjusted odds ratio (OR), hazard ratio (HR), or relative risk (RR) with a 95% confidence interval (CI); (5) articles written or translated into the English language. Studies were excluded if they (1) enrolled participants with a prior history of colorectal adenomas; (2) did not report specific results or contained duplicate or identical data; or (3) were case reports, systematic reviews, meta-analyses, editorials, commentaries, guidelines, conference abstracts, or letters. For the included studies, prior adenoma status was generally based on the absence of a documented diagnosis of colorectal adenomas or colorectal cancer.

We designed standardized forms to extract data on the first author, year of the publication, country of study location, study type, total number of participants, comparison and demographic features such as sex, age and body mass index (BMI). Adjusted effect estimates with 95% CI as well as covariates that were adjusted in the multivariable analysis were recorded.

The Newcastle–Ottawa Scale (NOS) was utilized to evaluate the methodological quality of observational studies. It consists of three dimensions (group selection, groups comparability and exposure), eight items and a maximum score of 9. Scores of 7–9, 4–6 and 0–3 were categorized as high, moderate and low quality, respectively.

All statistical analyses were performed using Stata/SE 16.0 software and p ≤ 0.05 was considered as a statistically significant result. The association between metformin therapy and colorectal neoplasms was assessed by consolidating the point estimate from each study using the generic inverse variance method of DerSimonian and Laird. Results stratified by gender were regarded as two distinct studies. Due to the relatively infrequent outcome, we assumed that the reported RRs were approximately ORs to simplify the results and ensure the comprehensiveness of the analysis and maximization of the statistical power (Greenland, 1987). Q test and I² statistics were employed to evaluate the extent of heterogeneity among the studies. An I² value of 0%–25% denotes heterogeneity might not be important, 26%–50% denotes low heterogeneity, 51%–75% denotes moderate heterogeneity, and more than 75% denotes high heterogeneity (Higgins et al., 2003). A random-effects model was employed if significant heterogeneity (I² statistic >50% or Q test <0.1) was observed (DerSimonian and Laird, 2015). Subgroup analysis was conducted by categorizing the studies based on outcome (cancer or adenoma), region (America, Europe, or Asia), mean age (<60 years, 60–70 years, or ≥70 years), male percentage (≥50%, <50%) and mean BMI (24–28 kg/m², ≥28 kg/m²). Publication bias was assessed by funnel plots as well as Egger’s test. Sensitivity analyses were performed through restricting the analytical cohort to studies meeting the following prespecified criteria: (1) cohort study design; (2) high methodological quality (NOS score ≥7); (3) comparator groups utilizing other antidiabetic medications (4) study populations exclusively comprising individuals with diabetes mellitus.

Further assessment of dose-response relationship between metformin use and CRC risk was conducted under the generalized least squares trend (GLST) estimation and restricted cubic spline random-effects model with four knots at the 5%, 35%, 65% and 95% percentile (Orsini et al., 2006; Orsini and Greenland, 2011). To verify the model, a p value for non-linear relationship was computed by testing the null hypothesis that the coefficient of the second spline equaled zero. If model verification indicated significance (p < 0.05), a non-linear model was adopted; otherwise, a linear model was used. Reported doses were consistently converted to the unit milligram/day. Since some studies did not report the number of cases in each category, we estimated these data based on the number of total cases and the reported effect size (Bekkering et al., 2008). When exposures were reported as a range, we took the midpoint. For open-ended categories, we set the dose to be 20% higher or lower than the nearest breakpoint, which proved to produce a better fit (Berlin et al., 1993; Filippini et al., 2022).

During the initial search, we identified 2021 results from Embase and PubMed and then removed 311 duplicated records and 1647 records based on titles and abstracts screening. After full-text reading, 34 articles were further excluded because of inappropriate outcomes, leaving 29 articles for data extraction. Four additional articles were retrieved for full text review after screening previous meta-analyses. As a result, 33 articles comprising 35 independent studies were initially retrieved, with two gender-stratified analyses by Lee et al. and Onitilo et al. identified as distinct datasets (Lee et al., 2011; Onitilo et al., 2014). The update searches were conducted in 20 March 2024 and resulted in the identification of 2 additional articles. Ultimately, 35 articles including 37 studies with 1,416,085 participants were integrated into this systematic review and meta-analysis comprising a total of 26 cohort studies, 5 case-control studies, and 6 nested case-control studies (Lee et al., 2011; Rosato et al., 2016; Seo et al., 2022; Tseng, 2017; Libby et al., 2009; Tsilidis et al., 2014; Onitilo et al., 2014; Bodmer et al., 2012; Bradley et al., 2018; Cardel et al., 2014; Chang et al., 2018; de Jong et al., 2017; Farmer et al., 2019; Krigel et al., 2022; Lee et al., 2021; Lin et al., 2015; Lopez et al., 2022; Murff et al., 2018; Rennert et al., 2020; Sehdev et al., 2015; Shin et al., 2020; Tseng, 2012; You et al., 2020; Zhang et al., 2022; Wang et al., 2013; Smiechowski et al., 2013; Yang et al., 2004; Oliveria et al., 2008; Ruiter et al., 2012; Oh et al., 2023; Chalhoub et al., 2023; Chung et al., 2008; Kanadiya et al., 2013; Cho et al., 2014; Kim et al., 2015). The search strategy and process for study selection are illustrated in Figure 1. Among these 37 studies, three only report effect sizes by doses of metformin instead of whether or not accepting metformin therapy, so these three studies were only analyzed in the dose-response meta-analysis (Bodmer et al., 2012; Chang et al., 2018; Tseng, 2012). The key characteristics of these studies are summarized in Table 1 and detailed characteristics for each study are available in Supplementary Table S3.

The average score of the NOS was 7.4 (ranging from 5 to 9). The methodological quality of eligible observational studies was moderate to high. Detailed scoring sheets are available in Supplementary Tables S4, S5.

34 studies were included in overall meta-analysis. Considering the inherent time-to-event information encapsulated in HRs, the meta-analysis of HR and OR with 95% CI was separately conducted.

The overall meta-analysis was shown in Figure 2. A lower risk of colorectal neoplasm in metformin users compared to non-users was observed with the pooled HR of 0.79 (95% CI, 0.73–0.85) under the random-effects model (I ² = 70.1%, P for heterogeneity <0.0001, 18 studies) Using the same model, the pooled OR obtained similar results with the value of 0.80 (95% CI, 0.74–0.87) and showed relatively high heterogeneity (I ² = 55.2%, P for heterogeneity <0.0001, 16 studies). Both analyses showed statistical significance.

Figure 3 present the subgroup analysis results. For HR analyses, metformin use was consistently associated with reduced colorectal neoplasm risk across most subgroups. Significant protective effects were observed in Asian populations (HR = 0.69, 95% CI, 0.58–0.81), younger than 60 (HR = 0.68, 95% CI, 0.54–0.86), male-predominant cohorts (HR = 0.75, 95% CI, 0.67–0.85) and the >24 to <28 kg/m² BMI subgroup (HR = 0.74, 95% CI, 0.62–0.87). No significant associations were found in American cohorts (HR = 0.90, 95% CI, 0.79–1.03) or female-predominant groups (HR = 0.89, 95% CI, 0.78–1.01).

For OR analyses, metformin demonstrated a significant protective effect against both colorectal cancer (OR = 0.82, 95% CI, 0.75–0.90) and adenoma (OR = 0.69, 95% CI, 0.58–0.83). Protective effects were pronounced in Asian (OR = 0.80, 95% CI, 0.67–0.95) and American populations (OR = 0.77, 95% CI, 0.69–0.86), but non-significant in Europe (OR = 0.85, 95% CI, 0.71–1.02). All age subgroups showed similar risk reductions. Male-predominant cohorts exhibited robust effects (OR = 0.79, 95% CI, 0.73–0.85), whereas male <50% subgroups had inconclusive results (OR = 0.76, 95% CI, 0.45–1.31). No significant associations emerged in BMI subgroups. Detailed subgroup analyses plots were shown in Supplementary Figures S1–S10.

In the analysis of publication bias for studies reporting HR, the funnel plot showed asymmetry, suggesting potential publication bias, which was corroborated by Egger’s test (P = 0.009) (Figure 4A). The trim-and-fill method revealed no evidence of missing studies (Supplementary Figure S11), lending support to the robustness of our findings; however, given the limited number of studies, the possibility of publication bias cannot be entirely excluded. For studies reporting OR, the funnel plot appeared approximately symmetric, with Egger’s test showing no evidence of publication bias (P = 0.312) (Figure 4B).

In sensitivity analyses, the results consistently demonstrated statistically significant inverse associations between metformin use and colorectal neoplasia risk (Table 2; Supplementary Figures S12–S17). Notably, all stratified estimates retained strong statistical significance, with HR ranging from 0.74 (95% CI, 0.65–0.83) to 0.79 (95% CI, 0.73–0.86) and OR spanning 0.78 (95% CI, 0.74–0.82) to 0.88 (95% CI, 0.83–0.94), indicating methodologically stable associations. In prespecified sensitivity analyses, all HRs originated from cohort designs and ORs uniformly employed other antidiabetic comparators, prespecified sensitivity analyses for these parameters were methodologically redundant and therefore excluded.

Five studies reported OR estimates for intensity of metformin use and eleven for metformin duration. Detailed dose-response effect data of these studies were listed in Supplementary Table S6. The dose-response analysis revealed that there was an overall nonlinear relationship among metformin intake and the risk of CRC (P for non-linearity = 0.0006) (Figure 5). Specifically, the OR exhibited its steepest decline (OR = 0.83, 95% CI, 0.76–0.91) across the 0–331 mg/day dose range, defining the dose of maximum benefit. Beyond this threshold, the OR plateaued at its nadir (0.8, 95% CI, 0.74–0.87) until 670 mg/day, marking a critical inflection point beyond which the protective effect gradually diminished with increasing doses. A sensitivity analysis excluding studies with extreme doses showed that the non-linear association and the estimated optimal dose remained largely unchanged (Supplementary Figure S18). In the duration-response analysis, a linear relationship was observed between metformin treatment duration and colorectal cancer risk (P for non-linearity = 0.67), with study inclusion restricted to treatment durations ≤60 months. Specifically, each incremental year of metformin use was associated with a 3% reduction in CRC risk (OR = 0.97, 95% CI, 0.95–0.99), reflecting a modest yet statistically significant protective effect (Figure 6).