Metformin Linked to Lower 5-Year Intermediate AMD Incidence in Diabetes
A population-based 5-year case-control analysis nested within a diabetic retinopathy screening program linked to general-practice prescribing records found that metformin exposure was associated with lower odds of incident intermediate AMD graded directly from images.
In adjusted models, the observational signal for metformin clustered around OR 0.64–0.66 for incident intermediate AMD (p≈0.01–0.02). This is not practice-changing, but it highlights how imaging-graded endpoints within screening workflows can support pragmatic prevention trials while underscoring the limits of causal inference in observational data.
The key point is what the endpoint represents: the primary signal reflects 5-year incidence of intermediate AMD among participants considered at risk at baseline, not necessarily progression within an already-intermediate subgroup unless separately defined by the study’s “population at risk” framework. The team tested robustness across age/sex-adjusted, multivariable complete-record, and multivariable multiple-imputation models, with the inverse association remaining directionally consistent rather than emerging from a single analytic choice. The late-AMD estimate, by contrast, attenuated and became non-significant after basic adjustment, aligning with limited event counts and wider uncertainty. Overall, the statistics support a reproducible association for incident intermediate AMD, while evidence for late AMD remains inconclusive.
How exposure was defined sets the interpretive ceiling. Metformin use was based on prescribing records (not confirmed ingestion), requiring a metformin prescription at baseline and at another point during the 5-year follow-up window. Screening photographs were linked through the Individualised Screening for Diabetic Retinopathy (ISDR) dataset and connected to general-practice prescription data, making this a real-world signal derived from routine care rather than trial-grade dispensing or adherence monitoring. That pragmatism comes with predictable gaps: no dose, incomplete characterization of pre-baseline duration, limited insight into adherence, and vulnerability to time-varying confounding as diabetes severity and comorbidity management change over time. Any prospective protocol built from this signal would need to operationalize exposure with far greater granularity.
The outcome, in contrast, was tightly phenotype-driven. AMD severity was graded on color fundus photographs using a modified AREDS level 1–4 scheme, with each participant classified by the worse eye. Incidence was constructed using a level-specific “population at risk” approach that excluded participants already at or beyond the relevant level at baseline, reducing the common problem of blending prevalent disease into incident endpoints. Compared with claims-based case finding, photographic grading anchors the endpoint to observable lesions rather than coding practices or care-seeking behavior. Clinically, that makes intermediate AMD incidence measurable within screening pathways without relying on diagnostic codes as a proxy for stage.
Measurement reliability and confounding strategy matter as much as the headline OR. Grading reproducibility was supported by QA re-grading in a subset, with kappa ≈0.77–0.79—consistent with good agreement for AMD severity assignment in this setting. Confounder control followed a prespecified plan: age, sex, HbA1c, and diabetes duration were retained a priori; additional covariates entered if their univariate association met p<0.20; and a collinearity rule (for example, excluding one of two variables when correlation reached r≥0.8) was used to stabilize multivariable models. These steps strengthen internal consistency, but they do not convert an adjusted association into a treatment effect, particularly when unmeasured determinants of AMD risk may differ between prescribing groups.
Missingness was addressed with both complete-record analysis and multivariate imputation by chained equations (MICE), and the intermediate-AMD association persisted under imputation rather than reversing. Residual bias remains plausible through baseline imbalances (including differences in age and baseline AMD prevalence by prescribing group), limitations in smoking measurement and modeling, unmeasured factors such as supplement use, reliance on color photography without OCT phenotyping, restriction to a diabetes-only cohort that may not generalize beyond that population, and low numbers of late-AMD events that constrain precision. Taken as a whole, the evidence is hypothesis-generating and best positioned to inform a prospective trial—one that recruits efficiently via screening programs, prespecifies imaging-graded intermediate AMD incidence over a pragmatic 5-year horizon, and explicitly captures dose, duration, and adherence.