Skip to content

AI Identifies Links Between Retinal Vascular Patterns and Risk of Cardiometabolic, Ocular Diseases

Key findings

  • In this study, convolutional neural networks quantified retinal microvascular indices on 97,895 retinal fundus photographs from 54,813 UK Biobank participants and the results were used to characterize phenome-wide clinical associations
  • Low vascular density and low fractal dimension (FD, an indicator of vascular branching complexity) were each strongly associated with higher risk of death among individuals who had either hypertension or type 2 diabetes mellitus
  • Those vascular indices were also significantly associated with numerous systemic conditions, including hypertension, hypertensive heart disease, diabetes, congestive heart failure, renal failure, sleep apnea and iron deficiency anemia
  • Associations were also identified with ocular conditions including diabetic retinopathy, ophthalmic manifestations of type 1 and type 2 diabetes, myopia, age-related macular degeneration, retinal detachments with defects, primary open-angle glaucoma and cataracts
  • The results of a Mendelian randomization analysis suggested low retinal density is a causal risk factor for retinal detachment, a finding that may prove useful for patient monitoring and adjusting treatment

For decades, it's been appreciated that retinal fundus photographs allow noninvasive assessment of the microvasculature. Newer imaging methods are able to discern geometrical and branching patterns that have been consistently linked with stroke, coronary artery disease and renal disease.

Massachusetts General Hospital researchers have demonstrated the potential for artificial intelligence (AI) to quantify retinal microvasculature indices from fundus images. The measures obtained show the potential to serve as biomarkers of cardiometabolic disease severity and predict the risk of multiple ocular conditions.

Seyedeh Maryam Zekavat, BS, MD-PhD student at Yale School of Medicine, the Broad Institute and Massachusetts General Hospital, Vineet K. Raghu, PhD, research faculty at the Cardiovascular Imaging Research Center, Pradeep Natarajan, MD, MMSc, director of the Cardiovascular Disease Prevention Center and investigator in the Cardiovascular Research Center, and colleagues report the findings in Circulation.


The team trained convolutional neural networks (CNN) to evaluate 97,895 retinal fundus images from 54,813 UK Biobank participants. Through exposure to tens of thousands of examples, a CNN learns to extract patterns from medical imaging data.

The neural networks calculated retinal vascular density and fractal dimension (FD), a measure of vascular branching complexity. The researchers investigated associations between these indices and mortality, 1,866 systemic and ocular conditions (median 10-year follow-up), and 88 quantitative traits, adjusting for age, sex, ancestry and smoking status.


Low vascular density or low FD (two standard deviations below the mean) was most strongly associated with incident mortality among individuals with prevalent hypertension or type 2 diabetes mellitus:

  • Low vascular density—HR for mortality, 1.83 (P=2.1x10–5) among those with hypertension or diabetes vs. 1.30 (P=0.072) among those without either condition
  • Low FD—HR, 1.83 (P=2.01x10–7) vs. 1.29 (P=0.03)

Systemic Conditions and Traits

Among cardiometabolic conditions, the strongest association with the vascular indices was for new-onset hypertensive heart disease (HR with low vascular density, 2.07; HR with low FD, 1.84).

Low vascular density and low FD were also significantly associated with a higher risk of the development of numerous other conditions, including:

  • Hypertension, type 2 diabetes, diabetic retinopathy, ophthalmic manifestations of type 1 and type 2 diabetes, congestive heart failure, renal failure, sleep apnea or iron-deficiency anemia
  • Markers of metabolic disease including elevated body mass index, body fat percentage and HbA1c
  • Impaired performance on pulmonary function tests

Ocular Conditions

Low vascular density and FD were also tied to increased future risk of multiple ocular conditions including myopia, age-related macular degeneration, retinal detachments with defects, primary open-angle glaucoma, and cataracts.

Genome-wide Association Study

Genome-wide association study of retinal vascular fractal dimension and density identified seven and three significant loci respectively. These regions were enriched for pathways linked to angiogenesis (VEGF, angiopoietin), and inflammation (interleukin, cytokine signaling).

Mendelian Randomization

Because of the strong associations of low vascular density and low FD with hypertension and diabetes, the team conducted Mendelian randomization analyses, which use genetic variants to distinguish correlation from causation in observational data.

The researchers calculated polygenic risk scores using gene variants for blood pressure and diabetes that came from a genome-wide association study external to the UK Biobank. Individuals with a genetically higher risk of hypertension and diabetes had lower retinal vascular density and FD, concordant with observed epidemiological associations.

In the converse analysis, the team calculated polygenic risk scores for vascular density and FD. Genetically lower vascular density was independently associated with an elevated risk of myopia, retinal detachments and defects (independent of myopia), and skin cancer (independent of ancestry and self-reported skin color, sun exposure and sun sensitivity).

Clinical Outlook

In the future, the results of AI-based examination of retinal fundus photos could be integrated with electronic health records, biomarkers and genetic data to inform risk prediction and risk modification. In particular, low retinal density seems to be a novel, causal risk factor for retinal detachment that could be used for patient monitoring. Additionally, genetic contributors to microvascular indices may provide insights towards therapeutic targets for microvascular disease across organ systems.

increased risk of mortality in people who had low retinal vascular density and/or low (<2 standard deviations below the mean) fractal dimension among those with existing type 2 diabetes or hypertension

increased risk of new-onset hypertensive heart disease per standard deviation lower retinal vascular density or retinal vascular fractal dimension

increased risk of new-onset retinal manifestations of type 1 diabetes per standard deviation lower retinal vascular density

increased risk of retinal detachments with defects, and 1.6x increased risk of primary open-angle glaucoma per standard deviation lower retinal vascular density

Learn more about the Cardiovascular Disease Prevention Center

Refer a patient to the Corrigan Minehan Heart Center


Variants in chromosome Xq23 are linked to reduced total cholesterol, LDL cholesterol and triglycerides, and reduced risk of coronary heart disease and type 2 diabetes, according to whole-genome data analyzed by Pradeep Natarajan, MD, MMSc, and an international team of colleagues.


By reviewing genomic data, cardiologists at Massachusetts General Hospital confirmed that somatic age-related mutations in myeloid-lineage blood cells, termed clonal hematopoiesis of indeterminate potential, is associated with an increased risk of stroke and found the risk is highest for hemorrhagic stroke, especially subarachnoid hemorrhage.