Three-parameter online nomogram predicts diabetic retinopathy risk in community-dwelling type 2 diabetes patients.

This retrospective cohort study assessed a three-parameter online nomogram designed to predict diabetic retinopathy risk among community-dwelling individuals with type 2 diabetes. The model incorporated diabetes duration, HbA1c, and high BMI (≥24 kg/m²) as input variables. Data were collected from 45 community health centers in Shenzhen and a center in Nanjing, comprising a primary cohort of 1,215 participants and an external validation cohort of 329 participants. The study aimed to provide a simple, readily implementable tool for diabetic retinopathy risk stratification in primary care, offering a cost-effective solution to bridge the screening gap.

The primary outcome measured discrimination using AUC-ROC values. In the training set, the AUC-ROC was 0.77 (95% CI: 0.73–0.81). Internal validation yielded an AUC-ROC of 0.79 (95% CI: 0.73–0.85), while the external validation set achieved an AUC-ROC of 0.81 (95% CI: 0.75–0.87). Calibration was assessed via the Hosmer–Lemeshow test, which was non-significant (P > 0.05), indicating good model fit. Brier scores remained below 0.15 across analyses, and net benefit was positive at threshold probabilities of 10–95%. The ratio of predicted to observed diabetic retinopathy cases was 1:1 at thresholds above 40%.

Safety and tolerability data were not reported, as adverse events, serious adverse events, discontinuations, and tolerability metrics were not applicable to this diagnostic prediction model. Key limitations include the observational nature of the retrospective design, which precludes causal inference regarding the nomogram's performance in diverse populations beyond the studied cohorts. Additionally, the external validation cohort size of 329 participants may limit generalizability to other geographic regions or healthcare systems. The study did not report funding sources or conflicts of interest.

This tool offers a practical approach for risk stratification in primary care settings where retinal imaging may be unavailable. However, clinicians should interpret the AUC-ROC values cautiously, recognizing that performance metrics derived from retrospective data may not fully replicate real-world clinical utility. The model's reliance on specific thresholds, such as BMI ≥24 kg/m², suggests it may be less applicable to populations with different anthropometric distributions. Further prospective studies are needed to confirm its utility across broader demographics before integration into standard care pathways.