Syed · Cardiovascular diabetology 2025 · prognostic cohort study · n=6127

Deep-learning prediction of cardiovascular outcomes from routine retinal images in individuals with type 2 diabetes.

Cited 37 times in the scientific literature.

Level 3 - non-randomized controlled study

Cohort study developing and validating a prognostic deep-learning model

PubMed 39748380 · doi:10.1186/s12933-024-02564-w · record verified 2026-08-27

What was done

Researchers evaluated whether a deep-learning model (EfficientNet-B2) could predict 10-year cardiovascular disease (CVD) outcomes from routine diabetic retinal screening photographs. The cohort comprised 6,127 individuals with type 2 diabetes and no prior history of myocardial infarction or stroke, split into training (70%), validation (10%), and testing (20%) sets. Retinal predictions were evaluated alongside the Pooled Cohort Equation (PCE) 10-year risk score and a coronary heart disease polygenic risk score (PRS). The primary outcome was time to first major adverse cardiovascular event (MACE: CV death, myocardial infarction, or stroke).

What was found

In the test set (n = 1,241; 288 MACE events; mean PCE 10-year risk 35%): - Retinal-predicted risk correlated strongly with the PCE score (r = 0.66) but not the PRS (r = 0.05). - Higher retinal risk was associated with MACE (unadjusted HR 1.05 per 1% increase, 95% CI 1.04–1.06, p < 0.001; adjusted for PCE and PRS: HR 1.03, 95% CI 1.02–1.04, p < 0.001). - The retinal AI model achieved an AUC of 0.697, identical to the PCE score (AUC 0.697). - Combining retinal risk with PCE and PRS improved discrimination to AUC 0.728. - An increase in retinal-predicted risk within 3 years was associated with higher subsequent MACE rate.

Why it matters

Routine diabetic retinal screening images can estimate 10-year cardiovascular risk on par with standard clinical risk scores, offering potential opportunistic CVD screening and modest incremental prognostic value when added to clinical and genetic risk factors.

Limits

Overall predictive discrimination was modest (AUC ~0.70). The analysis was conducted within a single cohort split into subsets without external validation described in the abstract. The study is restricted to individuals with type 2 diabetes without baseline myocardial infarction or stroke.

Cited by