The pooled analysis combined estimates from 12 independent studies drawn from 15 source populations. The overall pooled relative risk for metformin use and incident breast cancer was reported as 0.89 (95% CI 0.74–1.07). This overall estimate came with substantial heterogeneity (I-squared = 90.5%). Short, direct interpretation: across all pooled studies there was no definitive protective effect by conventional thresholds. The authors then separated studies by whether they explicitly addressed time-related biases—an important source of distortion in pharmacoepidemiology. In the nine studies that used a new-user design, time-varying exposure coding, or target trial emulation to reduce immortal time or time-window bias, the pooled RR was 1.00 (95% CI 0.94–1.08) and heterogeneity fell to 0%. The subgroup of studies that did not address these biases, or where bias control was unclear, produced a pooled RR of 0.69 (95% CI 0.56–0.84). The between-group test gave Q = 12.78 with p = 0.0004, indicating a statistically detectable difference between bias-controlled and uncontrolled studies; reclassification of one unclear study removed that between-group significance (Q = 1.15, p = 0.28). The study team therefore observed that apparent protective signals in pooled literature are driven largely by studies vulnerable to time-related biases.
The investigators searched PubMed up to 1 November 2026 and scanned reference lists from prior meta-analyses. They considered cohort and case-control studies reporting breast cancer-specific effect estimates for metformin use and incident breast cancer. One report per source population was selected for pooling to avoid double counting. The meta-analysis used a DerSimonian-Laird random-effects model. Studies were classified on whether they addressed time-related biases using explicit design features: new-user design, time-varying exposure modelling, or emulation of a target trial. The authors reported extracting all source PMIDs/DOIs, and provided a full extraction table, the PubMed search record, risk-of-bias assessments, excluded studies list, and Python analysis code as Supplementary Data. The study was registered on PROSPERO (CRD420261301586); the manuscript reports deviations from the protocol in Methods.
The authors applied the same eligibility and bias-assessment framework to imeglimin, a glucose-lowering agent structurally related to metformin and approved in Japan in 2021. The systematic search returned 218 PubMed records and 10 registered studies, but no clinical study reported breast cancer or any cancer outcome. No clinical-data meta-analysis was therefore feasible. The available evidence for imeglimin consisted of preclinical reports: one in vitro breast cancer cell study, one molecular docking study, and cancer models using concentrations at least about 40-fold higher than clinical plasma levels. The authors emphasized that structural similarity to metformin alone does not permit inference of a preventive effect against breast cancer. The report also noted that exposure time since imeglimin's approval is brief, which further limits the ability to detect incident cancer events in routine clinical data.
Heterogeneity across all pooled studies was high (I-squared = 90.5%), reflecting variation in design, exposure measurement, and bias control. When analyses were restricted to studies that explicitly handled time-related biases, heterogeneity collapsed (I-squared = 0%). That contrast signals that methodological differences, not consistent biological effects, drove much of the between-study variance. The authors ran subgroup comparisons that produced a statistically significant difference between bias-addressing versus non-addressing studies (Q = 12.78, p = 0.0004); when one study with unclear bias handling was reclassified the subgroup difference lost significance (Q = 1.15, p = 0.28). The paper provides transparency on risk-of-bias assessments and includes sensitivity tables in Supplementary Data, allowing readers to inspect which specific studies influenced pooled estimates.
The authors list key limitations anchored in available evidence. First, the pooled literature included studies with heterogeneous designs and variable control for immortal time and other time-window biases; many earlier observational signals appear contingent on those design flaws. Second, no clinical data exist to evaluate imeglimin and cancer outcomes, and preclinical models used supra-physiologic concentrations. Third, the authors did not access individual participant data; only aggregate published estimates were used. The report calculated that a comparative new-user, active-comparator cohort would need approximately 631 breast cancer events to detect a hazard ratio of 0.80, a concrete sample-size benchmark for future observational work. The authors recommended an adequately powered new-user, active-comparator cohort study as the next step. That recommendation is narrow and operational: design should avoid immortal-time bias, use time-varying exposure definitions or target-trial emulation, and ensure sufficient follow-up and event accrual before attempting inference about cancer prevention potential. These are the specific, documented unanswered questions and near-term research needs reported by the study team.