Metformin side effects: what 354,745 FDA adverse event reports show

Causality screen of the 40 strongest signals

Strongest disproportionality signals

Flagged terms (PRR ≥ 2, χ² ≥ 4, n ≥ 10), ranked by PRR; infinite PRRs (no comparator reports) excluded. The EBGM column is a second, shrinkage-based estimate (MGPS) with its 5th percentile EB05 and the BCPNN IC025; a pair flagged by PRR alone is marked PRR-only and should be read as fragile. The strata column recomputes the PRR on serious reports only (and, for the 300 most-reported drugs, adjusted for age band), and flags pairs whose reports are unusually skewed by sex or come mostly from consumers. Each row also shows the PRR against the other drugs of the same FDA class, which removes most confounding by the condition being treated, and a calibrated p-value: on validation, known non-causes average an ROR of about 2, so significance is judged against that empirical null (✓ = p < 0.05 after calibration). 1,113 further terms are in the interactive explorer. Reaction links open the term's cross-drug page.

Clinical concept groups

MedDRA splits one clinical problem across several terms (e.g. "neutropenia", "neutrophil count decreased", "agranulocytosis"), which dilutes term-level signals. These rows count reports with any term in a concept group (an exact union, not a sum), compared with all other drugs. Groups are hand-built and listed in the methods.

Trial evidence (ClinicalTrials.gov)

Most frequently reported reactions

By raw co-report count, regardless of disproportionality. Common terms here are often the illness itself or generic outcomes.

Unlabeled candidate signals

These flagged terms were not found in the FDA label text and do not match an artifact pattern. They are the pairs that would merit a structured assessment (temporality, dose-response, biological plausibility, consistency across databases, trial or cohort evidence). The screen has not done that; string-matching misses label wording, and a strong PRR can still be stimulated reporting, co-medication, or the disease itself.

• Neurologic neglect syndrome — n=289, PRR 62.7
• Pulmonary vasculitis — n=226, PRR 43.8
• Carotid artery thrombosis — n=296, PRR 31.0
• Hyperlactacidaemia — n=804, PRR 25.6
• Adjustment disorder — n=301, PRR 17.5
• Hypoglycaemic coma — n=508, PRR 12.4
• Gastrointestinal hypomotility — n=384, PRR 12.1
• Neuritis — n=227, PRR 10.1
• Distributive shock — n=305, PRR 9.3
• Impaired gastric emptying — n=1,751, PRR 9.3

✓ A hand-written Bradford Hill causality assessment exists for this drug: Metformin ↔ Lactic acidosis.

Want to assess one of these properly? The runbook describes the Bradford Hill / trial-evidence workflow used on the assessment page, and handoff_top1000_signals.csv lists the strongest unassessed pairs across all drugs.

Common questions

Sources & methods

• FDA label via openFDA (set ID 7e41818c-60e9-4bcf-9586-7bb8d33d5e89, effective 20260615; matched on openfda.generic_name.exact:"METFORMIN" AND openfda.product_type:"HUMAN PRESCRIPTION DRUG")
• PubChem CID 4091 description (ChEBI)
• Bayesian shrinkage measures computed offline over all 1,317,300 cached pairs: BCPNN information component (IC, with Norén closed-form IC025) and MGPS empirical-Bayes geometric mean (EBGM, EB05) with a two-component gamma prior fitted by maximum marginal likelihood (α₁=0.57, β₁=0.02, α₂=1.42, β₂=0.70, w=0.17); unstratified
• FAERS pair statistics from the OpenPV build of 2026-09-14 (openFDA count= aggregation, field patient.drug.openfda.generic_name.exact, 1,153 reaction terms, min 3 co-reports); PRR/ROR with 95% CI, chi-square with Yates correction
• Serious-only and stratified counts (serious:1, patient.patientsex, primarysource.qualification) from openFDA count queries; serious-only PRR uses the drug serious report total and the term global serious total (N_serious=11,882,979); strata shown only when n ≥ 20
• Seriousness and sex counts from openFDA count=serious, patient.patientsex, seriousnessdeath
• ClinicalTrials.gov API v2 (query METFORMIN, results-bearing studies sorted by enrollment)

Page generated 2026-10-14. Label and PubChem text are retrieved automatically and may lag the current label; always check the current prescribing information. OpenPV is a beta research tool of the Open Source Medicine Foundation.