Multimodal deep learning of fundus abnormalities and traditional risk factors for cardiovascular risk prediction

AI Generated Summary*

Fundus photographs fed through a DenseNet-169 branch were combined with seven clinical risk factors in a fully connected network to flag existing coronary or cerebrovascular disease. Development used 3518 images from a Korean hospital, with 2954 images for internal testing and 11,298 UK Biobank images (613 cases) externally. Adding photographs lifted the AUROC marginally internally, from 0.766 to 0.781 (p = 0.047), and from 0.849 to 0.872 in the UK Biobank, where the best configuration reached 0.905 and the Pooled Cohort Equation reached 0.677. Restricted to confidently classified patients, a non-invasive variant matched the full model near 0.9, and its positive predictions were associated with a hazard ratio of 6.28 for incident events among those initially free of disease.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.