Causal-RetiGraph: Cross-Cohort Retinal Support and Same-Subject Pathway Analysis for Diabetic Retinopathy

Modern AI models for diabetic retinopathy (DR) often achieve high classification accuracy but fail to explain the underlying connections between local retinal lesions, vascular structure, and systemic disease pathways. This study asks how we can bridge the gap between image-derived retinal phenotypes and systemic epidemiological data to better understand the microvascular pathways of diabetes.

The authors introduce Causal-RetiGraph, a two-fold biomedical informatics framework. The first fold (retinal-image) processes fundus images to construct an interpretable phenotype, X1234, which fuses four evidence streams: vessel maps, lesion evidence, image embeddings, and AutoMorph biomarkers. This fold uses a spatial branch to model lesion-vessel geometry and a Jacobian branch to capture embedding-biomarker sensitivity. The second fold (NHANES) uses an independent cohort of diabetic participants to anchor systemic exposures (e.g., HbA1c, blood pressure) and outcomes. The framework uses X1234 to prioritize systemic pathways and a same-subject retinal mediator family, R*, to perform participant-level pathway summaries.

The retinal graph phenotype X1234 achieved high performance in DR grading (0.9711 AUROC). The study identified that glycaemic, renal, and haemodynamic axes consistently dominate both the retinal-image and systemic NHANES folds. Specifically, HbA1c, urine albumin, and pulse pressure emerged as the strongest systemic anchors for DR. Participant-level analysis confirmed that glycaemic-renal and glycaemic-haemodynamic pathways are the most significant mediator-style signals, whereas inflammatory pathways showed weaker associations. These results suggest that retinal graph phenotypes can serve as a structured intermediate layer between local retinal injury and systemic microvascular stress.

This framework moves beyond black-box classification by providing a structured, interpretable link between ocular imaging and systemic pathophysiology. By separating image-derived support from same-subject mediation, the authors provide a rigorous way to generate and prioritize biomedical hypotheses without overstating causal claims. This approach helps clinicians and researchers understand the eye as a window into systemic health, potentially informing future targeted interventions for diabetic microvascular complications.