A deep learning–derived retinal biomarker improved discrimination of diabetic foot-related amputation risk when added to conventional clinical variables in patients with type 2 diabetes, according to a retrospective observational study published in Frontiers in Endocrinology.

The study included 392 patients with type 2 diabetes treated at a university hospital in South Korea between 2018 and 2024. Among them, 79 patients had undergone diabetic foot-related amputation and 313 had not. Participants were randomly divided into training (70%) and validation (30%) cohorts. 

Researchers evaluated a retinal biomarker generated by the artificial intelligence software Dr. Noon CVD, which was originally developed to estimate coronary artery calcium probability from fundus photographs. 

Four logistic regression models were developed to assess the incremental value of the retinal biomarker: Model 1 included baseline clinical variables (age, sex, diabetes duration, and glycated hemoglobin [HbA1c]); Model 2 added the retinal biomarker to the baseline model; Model 3 added diabetic retinopathy to the baseline model; and Model 4 combined baseline variables, HbA1c, diabetic retinopathy, and the retinal biomarker. 

Compared with the baseline model (Model 1), adding the retinal biomarker (Model 2) increased the area under the receiver operating characteristic curve (AUROC) by 0.146 from 0.606 to 0.754. When diabetic retinopathy was already included (Models 3 and 4), the biomarker increased the AUROC by 0.081 from 0.710 to 0.791. Reclassification analyses also favored inclusion of the biomarker, with improvements in both continuous net reclassification index (0.629) and integrated discrimination improvement (0.062). 

The sensitivity analysis using diabetic foot risk scores derived from the external retinal dataset showed similar findings. Incorporating the retinal biomarker into diabetic foot risk models improved discrimination regardless of whether diabetic retinopathy was included.  

The investigators also evaluated clinically relevant operating thresholds. Assuming a 27% amputation prevalence among patients with diabetic foot, the prespecified rule-out threshold yielded a sensitivity of 88%, and a negative predictive value of 93%. At the rule-in threshold, specificity was 90%, and the positive predictive value was 56%. 

The researchers noted that predictive value was substantially lower when applying the model to the broader diabetes population because of the much lower prevalence of amputation. 

Feature importance analyses showed that diabetic retinopathy, the retinal biomarker, and glycated hemoglobin were the strongest contributors to the full model, whereas age, sex, and diabetes duration contributed relatively little after adjustment. The retinal biomarker also remained independently associated with amputation after accounting for diabetic retinopathy, suggesting it captured additional systemic vascular risk signals beyond retinal microvascular disease complications. 

The researchers noted several study limitations. The retrospective, case-control–like design demonstrated association rather than prospective prediction, and external validation relied on a diabetic foot dataset rather than an independent amputation cohort. The external dataset also differed from the primary cohort in ethnicity, imaging equipment, and available clinical variables. In addition, the study lacked information on neuropathy, peripheral artery disease, body mass index, and lipid profiles. The researchers added that larger prospective multicenter studies with independent amputation outcomes are needed before clinical implementation.  

"These findings suggest the potential utility of retinal imaging for vascular risk stratification in patients with diabetes, particularly among individuals with a high baseline risk of amputation," the authors wrote.  

Junseok Park, MD, Dongjin Nam, MD, and Sahil Thakur reported receiving honoraria from Mediwhale Inc, the company that developed Dr. Noon CVD. The other authors declared no commercial or financial conflicts of interest.