Artificial intelligence-driven analysis of retinal fundus images for hypertension subtyping and risk stratification

Abstract Non-invasive and quick methods are necessary for high blood pressure screening and classification and risk assessment procedures. The research developed and validated an artificial intelligence model which analyzes retinal images to identify blood pressure subtypes and forecast potential health complications for the Uzbek population. The research involved 600 patients with hypertension and 400 healthy participants who took part in the study. The researchers used a non-mydriatic camera to capture retinal images which they analyzed using a Residual Network model that contained 50 layers. The study included 500 patients who underwent 24-hour blood pressure monitoring and a 5-year follow-up assessment which documented complications for 300 patients. The model estimated systolic blood pressure with a mean absolute error which reached 8.4 mmHg. The researchers achieved 84% accuracy in classifying blood pressure subgroups which included isolated systolic and isolated diastolic and systole-diastolic blood pressure subgroups. The model successfully identified nocturnal blood pressure patterns which included deeper and non-dipper patterns with 78% accuracy. The AUC of 0.85 predicted 5-year cardiovascular risk which exceeded the Framingham model AUC of 0.78. The model performed better than two experienced ophthalmologists (accuracy of 86% vs. 75%), and the time to analyze each image was 0.2 seconds compared to 120 seconds. The model performed equally well for all ethnic groups and all geographical areas of Uzbekistan. The research demonstrates that AI-powered retinal image analysis serves as a fast and precise noninvasive technique for blood pressure assessment and risk evaluation of complications and group identification.