Skip to main content

  • Measuring Vascular Health via AI Retinal Vasculometry

    By Lynda Seminara
    Selected by Prem S. Subramanian, MD, PhD
    Add to My Bookmarks

    Journal Highlights

    British Journal of Ophthalmology
    2022;106(12):1722-1729

    Download PDF

    Rudnicka et al. developed and validated risk models based on artificial intelligence (AI) analysis of retinal vessel images and found that AI-enabled ret­inal vasculometry (RV) is an effective noninvasive method to predict circula­tory mortality, stroke, and heart attack. Its performance was on par with that of current risk-scoring systems. Future formal assessment with experimen­tal evidence will help determine the clinical utility, said the authors, who emphasized that the advantages of RV, including the low cost, could make it attractive for community screening programs.

    For this work, retinal vessel analysis was performed on images from records of patients listed in the U.K. Biobank and from participants of the European Prospective Investigation into Cancer–Norfolk (EPIC-Norfolk). Data extracted from images included retinal arteriolar and venular width, area, and tortuosity. Predictive models were developed from the Biobank dataset, using multivari­able Cox proportional hazards regres­sion for circulatory mortality, incident stroke, and myocardial infarction (MI), which were validated in the EPIC-Nor­folk dataset. The authors compared the performance of a simple model based on RV, age, smoking status, and medical history with that of the Framingham risk score system for incident stroke and incident MI. They also looked at the utility of adding RV to the Fram­ingham system.

    Altogether, they developed prognos­tic models from 65,144 U.K. Biobank participants (mean age, 56.8 years; me­dian follow-up, 7.7 years), which they validated by data for 5,862 EPIC-Norfolk participants (mean age, 67.6 years; median follow-up, 9.1 years). To determine model performance, they applied optimism-adjusted calibration, C statistics, and R2 statistics. The pri­mary outcome measure was circulatory mortality as defined by ICD-10 codes.

    The prediction models for circulato­ry mortality had optimism-adjusted C statistics ranging from .75 to .77 and R2 statistics ranging from .33 to .44. For stroke and MI, adding RV to the Fram­ingham score system did not improve the model’s performance. However, the simple RV model performed at least as well as the Framingham system.

    The authors emphasize that their system is focused on the retinal tree and offers detailed quantification of vasculometry, making it suitable for large populations. With further research, this simple low-cost method could become a key adjunct or stand-alone assessment tool.

    The original article can be found here.