Purchase this article with an account.
Yiu Lun Wong, Marco Yu, Crystal Chong, Tien Yin Wong, Ching Yu Cheng, Carol Yim-lui Cheung; Deep-learning measurement of retinal vessel caliber predicts incident myocardial infarction. Invest. Ophthalmol. Vis. Sci. 2022;63(7):177 – F0024.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Retinal arteriolar and venular calibers have been proposed to be surrogate measures of systemic microvascular health. We evaluated the associations of deep-learning based retinal vessel caliber measurements with incidence of fatal and non-fatal myocardial infarction (MI) using data from the UK Biobank.
Prevalent and incident MI was defined with an algorithm developed by UK Biobank, combining information from self-reported medical conditions, operations and medications, hospital admissions and death registries before and after the code date of baseline assessment respectively. Fully automated measurements of central retinal artery and vein equivalents (CRAE/CRVE) from retinal fundus photographs in the area 0.5-2.0 disc diameters away from the optic disc were estimated with a recently validated deep-learning system (SIVA-DLS). Hazard ratios (HR) and area-under-the-curve (AUC) were calculated for evaluating the association and improvement in discriminative performance between CRAE/CRVE with incident MI.
After excluding individuals with prevalent MI and ungradable retinal photographs, 34,841 subjects at risk of incident MI were included in the final analysis. Baseline assessment was conducted from 2009 to 2013, with subjects at risk of MI followed up to March 2020. A total of 375 subjects developed incident MI. Persons with narrower CRAE were more likely to have incident MI (HR 1.67 [95% CI: 1.19-2.36], comparing 1st to 4th quartile, and HR 0.86 [95% CI: 0.75-0.98], per SD increase in CRAE), adjusting for age and gender. When further adjusted for total cholesterol, high-density lipoprotein cholesterol, systolic blood pressure, and other risk factors, the association between CRAE and incident MI persisted (HR 1.64 [95% CI: 1.16-2.32], 1st vs. 4th quartile). CRVE was associated when adjusted for age and gender (HR 1.51 [95% CI: 1.06-2.13]) but not in multivariable models. AUC increased significantly from 0.718 to 0.724 (adjusted for age and gender, p = 0.030) when CRAE/CRVE is included in the models.
Narrowed retinal arteriolar caliber, reflecting systemic microvascular damage, was associated with incident MI. Deep-learning based retinal vessel caliber measurements from retinal photographs may be a useful biomarker of cardiovascular disease in clinical and epidemiological studies.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only