Deep Learning-Derived Retinal Age Detects Cognitive Impairment

To investigate the association between deep learning-derived retinal age and cognitive function and to evaluate whether retinal age outperforms chronological age as a screening tool for cognitive impairment.

Cross-sectional analysis of the data from multicenter Artificial Intelligence-Ready and Exploratory Atlas for Diabetes Insights (AI-READI).

We included 1049 participants (≥40 years) (mean [standard deviation] chronological age, 60.3 [11.2] years) enrolled in the AI-READI cohort, a multicenter, cross-sectional study.

Retinal age was estimated from color fundus photographs using a pretrained deep learning model. Cognitive function was assessed using the Montreal Cognitive Assessment (MoCA), with cognitive impairment defined as a MoCA score <26. Generalized linear models were used to compare MoCA scores across retinal age quartiles. Poisson regression and logistic regression were used to evaluate the association between retinal age and cognitive impairment, adjusting for chronological age and cardiometabolic risk factors.

Retinal age, MoCA scores, and cognitive impairment.

Retinal age correlated more strongly with MoCA total score than chronological age (R, -0.47 vs. -0.21). Retinal age was associated with cognitive impairment in a dose-response manner with adjusted risk ratio (RR) of 2.81, 6.03, and 10.13 for the second, third, and fourth quartiles in comparison to the first quartile of retinal age (all P < 0.001), while chronological age had a RR of 1.37, 1.42, and 1.65 for the second, third, and fourth quartiles compared with the first quartile (all P < 0.01). Retinal age had higher area under the curves (AUCs) than chronological age in detecting cognitive impairment (0.81 vs. 0.59 without covariates, and 0.83 vs. 0.66 with covariates). The combination of retinal age and chronological age yielded an AUC of 0.87. In the analysis stratified by diabetic severity, retinal age maintained a similar AUC (0.89 with covariates) for detecting cognitive impairment in nondiabetic healthy controls.

Deep learning-derived retinal age is a biological biomarker that significantly outperforms chronological age in its association with cognitive function and has better performance in detecting cognitive impairment. This scalable, image-based biomarker has the potential for opportunistic screening within existing clinical workflows, facilitating earlier detection of cognitive decline.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.