Identifying the genetic and non-genetic factors associated with accelerated eye aging by using deep learning to predict age from fundus and optical coherence tomography images

BackgroundThe rate at which different portions of the eye ages can be measured using eye fungus and optical coherence tomography (OCT) images; however, their genetic and environmental contributors have been elusive. MethodsWe built an eye age predictor by training convolutional neural networks to predict age from 175,000 eye fundus and OCT images from participants of the UK Biobank cohort, capturing two different dimensions of eye (retinal, macula, fovea) aging. We performed a genome-wide association study (GWAS) and high-throughput epidemiology to identify novel genetic and environmental variables associated with the new age predictor, finding variables associated with accelerated eye aging. FindingsFundus-based and OCT-based eye aging capture different dimensions of eye aging, whose combination predicted chronological age with an R2 and mean absolute error of 83.6{+/-}0.6%/2.62{+/-}0.05 years. In comparison, the fundus-based and OCT-based predictor alone predicted age with R2 of 76.6{+/-}1.3% vs. 70.8{+/-}1.2% respectively. Accelerated eye fundus- and OCT-measured accelerated aging has a significant genetic component, with heritability (total contribution of GWAS variants) of 26 and 23% respectively. For eye fundus measured aging, we report novel variants in the FAM150B gene (ALKAL2, or ALK ligand 2) (p<1x10-150); for OCT-measured eye aging, we found variants in genes such as CFH (complement factor H), COL4A4 (type 4 collagen), and RLBP (retinaldehyde binding protein 1, all p<1x10-20). Eye accelerated aging is also associated with behaviors and socioeconomic status, such as sleep deprivation and lower income. ConclusionsOur new deep-learning-based digital readouts, the best eye aging predictor to date, suggest a biological basis of eye aging. These new data can be harnessed for scalable genetic and epidemiological dissection and discovery of aging specific to different components of the eye and their relationship with different diseases of aging. FundingNational Institutes of Health, National Science Foundation, MassCATS, Sanofi. Funders had no role in the project. Research in context Evidence before this studyWe performed a search on NCBI PubMed and Google Scholar searching for the terms, "eye aging", "optical coherence tomography" (OCT), "fundus", and/or "deep learning". We found others have shown feasibility of predicting chronological age from eye image modalities, finding five publications that demonstrated chronological age may be predicted from images inside and outside of the eye, with mean absolute errors ranging from 2.3-5.82 years. Added value of this studyOur new eye age predictor combines both OCT and fundus images to assemble the most accurate fundus/OCT age predictor to date (mean absolute error of 2.62 years). Second, we have identified new genetic loci (e.g., in FAM150B) and epidemiological associations with eye accelerated age, highlighting the biological and environmental correlates of eye age, elusive in other investigations and made scalable by deep learning.