The macula, crucial for vision, requires costly and complicated optical coherence tomography (OCT) for measurement of its thickness. However, large scale datasets such as the Canadian Longitudinal Study on Aging (CLSA) have only simple 2D fundus images available. Here we aimed to use Artificial Intelligence (AI) to estimate macular thickness from fundus images, leading to large sample sizes for GWAS of the trait.
We used 2D fundus images and OCT derived retinal thickness measures in the UK Biobank (UKBB) to train an ML model to predict average macular thickness. We applied a diverse range of ML techniques. We used pre-trained EfficientNet models for feature extraction, and fine-tuned them to predict macular thickness in UKBB. Both Bayesian and deterministic approaches were applied for the prediction tasks. We applied the model to fundus images from the CLSA. We conducted a GWAS using REGENIE on directly measured values in UKBB and predicted values in CLSA. We assessed the correlation between effect sizes of the significant loci for both measured and predicted values, and across the entire genome, using LD score regression. Subsequently, we jointly analysed GWAS results from UKBB and CLSA to enhance the power of gene discovery.
Among ML methods, the Bayesian approach demonstrated superior performance. The LD score analysis revealed a genetic correlation of approximately 0.77 (P-value: 2.03e-50) between UKBB and CLSA datasets. Additionally, the correlation between GWAS Beta values for genome-wide significant SNPs of UKBB (P-value < 10e-8) and CLSA was strong (R: 0.80, CI: 0.71 to 0.86). The GWAS results revealed over 100 genomic risk loci.
Our Innovative AI approach suggests a potential cost-saving and convenient alternative for measuring challenging phenotypes such as macular thickness. This improvement can enhance the power of gene and pathway discovery, paving the way for development of novel treatment targets and improving risk prediction.