Answers are grounded in this paper and cite their sources.
Purpose: To assess how associations between retinal findings and retinal age-gap (RAG) readouts vary with age handling, refitting uncertainty, label release, reference definition, and representation. Methods: We analyzed 9545 Brazilian Multilabel Ophthalmological Dataset (BRSET) 1.0.1 records from 5227 patients using patient-grouped nested Ridge readouts trained on finding-negative controls with fold-local age correction. Finding contrasts used same-fold, different-patient, age-matched references. Fixed out-of-fold (OOF) inference used 2000 patient resamples with dynamic rematching; a 1000-replicate patient full-refit bootstrap repeated-fold construction, tuning, fitting, correction, and matching. Results: Seven of 12 associations met the fixed-OOF Benjamini–Hochberg false discovery rate (FDR) criterion, including diabetic retinopathy (DR), 4.06 years (95% confidence interval [CI], 3.38–4.80), and macular edema, 4.30 years (95% CI, 3.37–5.17). Five of 12 full-refit intervals excluded zero. DR remained positive across all representations and in multilabel-adjusted and isolated-finding analyses, whereas macular edema remained positive across representations but attenuated to near zero after simultaneous multilabel adjustment. In the harmonized representation analysis, age-related macular degeneration (AMD) was 1.45 years (95% CI, 0.27–2.82) with the historical RETFound ViT-L/16 representation, −2.99 years (95% CI, −4.81 to −1.43) with DINOv3 ViT-S/16, and −4.88 years (95% CI, −6.43 to −2.95) with the PhysioNet-released ConvNeXt-Tiny embedding. The AMD full-refit interval included zero. Conclusions: Retinal-finding associations with RAG varied across representations in magnitude and direction; inferential support changed with model-refitting uncertainty. RAG is a model-conditional measurement of fundus appearance rather than a direct measure of biological aging. Translational Relevance: Retrospective findings from frozen representations and coarse labels have limited generalizability and require prospective validation before clinical use.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: