Fundus images from the left eyes of 440 subjects aged 50-59 years enrolled in the Northern Ireland Cohort of Longitudinal Ageing were analyzed. Subjects were categorized as normotensive or hypertensive, according to thresholds on systolic/diastolic blood pressure measurement (140/90 mm Hg) averaged over two sitting measurements in a clinical setting. A fully automatic system to analyze each image used conventional and deep neural network machine learning techniques to locate retinal landmarks and detect, classify and measure retinal vessels. From this data, a measure of the arteriolar-venular ratio (AVR) in the peripheral retina was calculated. Semi-automatic analysis was also performed. Results are presented in Table 1. Subjects had mean age of 54.6 ± 2.9 years; 56.1% (247 of 440) females, with 34.3% (151 of 440) subjects categorized as hypertensive. Narrower arterioles and smaller AVR were observed in subjects with hypertension. This was also observed in fully-automated analysis, however 4% (17 of 440) subjects failed to be processed by the system. In fully-automated analysis the area under a receiver operator characteristic curve of AVR for hypertensive status was 0.69 (95% CI, 0.63 to 0.74). Table 1 - Results for semi-automated and automated analysis of retinal vessel parameters. *p<0.005 Automated measurement of AVR in ultra-widefield fundus imaging was associated with hypertension. With further development, such as evaluation against diagnosis of hypertension obtained from ambulatory blood pressure monitoring clinics, this system could become a test for undiagnosed hypertension in people attending routine eye health check-ups.
