AI-informed retinal biomarkers predict 10-year risk of onset of multiple hematological malignancies

Leukemia, myeloma, and lymphoma remain hard to catch early: their clinical presentation is heterogeneous, symptoms are often subtle, and together they account for roughly 10% of all malignancies with mortality rates that exceed more common cancers like prostate and breast cancer. Diagnostic delay matters enormously — emergency diagnoses have a 3-year survival rate of just 40%, versus 77% for cases caught through primary care. Chronic inflammation, a key driver of these cancers, is known to induce compensatory microvascular remodeling, and a growing body of evidence links hematological malignancies to early ocular changes: about 40% of newly diagnosed multiple myeloma patients already show pre-existing ocular alterations, and retinal vessel density changes have been documented at diagnosis and even in leukemia remission.

RetHemo asks whether the retina — imaged non-invasively and cheaply via routine fundus photography — already carries a measurable signature of a hematological cancer that hasn't been diagnosed yet. Using 1,237 UK Biobank patients with pre-diagnosis fundus scans, this study builds an explainable retinal vasculature risk score and shows it can flag elevated leukemia and myeloma risk up to a decade before clinical diagnosis, with a biological basis corroborated by serum proteomics.

Retinal fundus scans are automatically segmented for vessels, arteries, and veins using a pre-trained CNN. A separate ResNet50 locates the fovea to define concentric parafoveal zones, and an ensemble of pre-trained CNNs segments the optic disk and cup to define zones capturing larger vessels near the disk. Within vessel, artery, and vein masks — separately for the parafoveal and optic-disk zones — curvature, tortuosity, and branching-angle features are extracted and aggregated into summary statistics.

The top three oculomics features (selected on the training split only) feed a Cox proportional hazards model trained on each cancer's training set (D1_Leuk, D1_Myel, D1_Lymp), producing three disease-specific risk scores — RetHemo_Leuk, RetHemo_Myel, RetHemo_Lymp — each validated on its own untouched holdout set. To reduce reliance on any single feature-selection choice, an independent unsupervised hierarchical clustering analysis on the full vasculature feature set is used to identify retinal phenotype clusters, tested for differential cancer risk via multivariate and univariate log-rank tests. Finally, RetHemo predictions are correlated with serum protein expression (Olink Proximity Extension Assay) to probe the biological basis of the retinal signal.

Cox models built from clinical variables alone (RetHemo_clinical: age, sex, hypertension, diabetes) or blood cell counts alone (RetHemo_Blood: RBC, lymphocyte, monocyte, neutrophil, eosinophil, basophil) failed to significantly stratify risk in any of the three cohorts (all p > 0.18). RetHemo's retinal vasculature signal outperforms both, suggesting the retina captures prognostic information not already present in standard clinical risk factors or blood counts at the time of imaging.

A gradient-boosted classifier trained purely to distinguish future cases from controls (rather than model time-to-event) also carried prognostic signal for leukemia — high-risk patients by classifier probability had a significantly elevated risk of developing leukemia (HR 3.16, 95% CI: 1.42–7.06, p = 0.029) — though equivalent classifiers for myeloma and lymphoma did not reach significance, reinforcing that the Cox-model approach with selected oculomics features is the more robust formulation.

This is a single-site, predominantly white cohort from the UK Biobank; multi-institutional and multi-ethnicity validation is underway. Marrow involvement in lymphoma is less widespread and more localized to lymph nodes than in leukemia or myeloma, which may explain the weaker prognostic signal observed for RetHemo_Lymp in the supervised (though not the unsupervised clustering) analysis. The authors note this work complements existing evidence for retinal vasculature biomarkers in stroke, cardiovascular disease, and chronic kidney disease, extending oculomics into cardio-oncological risk assessment.