The Retinal-Glycemic Link: Using Retinal Imaging to Explore Glycemic Variability
Retinal vasculature provides insight into systemic body health. In this work, we characterize and illustrate the potential for early detection of type 2 diabetes mellitus (T2DM) using optical coherence tomography angiography (OCTA) images. Using the AI-READI cohort (N=1062), we extract vascular metrics from OCTA images and correlate them with diabetic state (based on hemoglobin A1 c, HbA 1 c) as well as continuous glucose monitoring (CGM) variables. Our results show that OCTA features of foveal avascular zone area, roundness of vessels, and degree of branching are most highly-correlated with HbA 1 c and CGM variables. We also explore the ability of multimodal deep learning (DL) models trained on OCTA metrics and additional retinal imaging to classify high vs. low glycemic variability, which has been shown to be an early biomarker of glycemic dysregulation and insulin resistance. We find that combining fundus images with OCTA metrics moderately improves DL classification performance compared to unimodal inputs. These findings highlight the potential of multimodal retinal imaging as a widely deployable tool for early metabolic screening.
