Does lens opacity matter? The effect of cataract on deep learning based cardiovascular disease risk scores from...

To investigate the effect of cataracts on a deep learning (DL) model for cardiovascular disease (CVD) risk prediction.

This retrospective, dual-cohort study analyzed fundus images at baseline, 1, and 6-months post-cataract surgery from a longitudinal cohort (patients who underwent cataract surgery at Hanyang University Guri Hospital [HUGH]) and a cross-sectional replication cohort (Singapore Epidemiology of Eye Diseases [SEED] study). CVD risk scores were generated using an artificial intelligence (AI) software (Dr.Noon CVD). Longitudinal changes were evaluated using a generalized estimating equation (GEE) model. The association between CVD scores and cataract surgery was also assessed in the SEED study using multivariable linear regression. Subgroup analyses were performed based on diabetic retinopathy (DR) status.

In the longitudinal cohort, Dr. Noon CVD scores significantly increased from baseline after cataract surgery at 1 month (β = +2.14; 95% CI, 1.28-3.01; P < 0.001) and 6 months (β = +1.69; 95% CI, 0.82-2.56; P < 0.001). A significant interaction with DR was observed, showing substantially larger score increases at 1 month (interaction β = +3.02; 95% CI, 1.20-4.85; P = 0.001) and 6 months (interaction β = +4.78; 95% CI, 2.99-6.56, P < 0.001). These findings were partially replicated in the SEED cohort, where pseudophakia was significantly associated with a higher CVD score compared to phakic eyes (β = +2.60; 95% CI, 1.71-3.49, P < 0.001); however, the interaction between cataract surgery and DR status was not formally significant in SEED (P = 0.451).

Cataract-induced media opacity is associated with significant attenuation of DL-derived CVD risk scores, particularly in eyes with DR. This score shift following improvement in optical clarity highlights the critical need to account for lens status when interpreting fundus-based AI prediction models.