Scholarlink

<p>Cardiovascular-kidney-metabolic (CKM) syndrome requires integrated risk assessment, but traditional evaluation depends on numerous clinical variables and specialized examinations, limiting scalability. We developed a retina-enhanced multimodal deep learning framework integrating fundus photographs with eight routine clinical variables, incorporating two novel modules without manual segmentation: a multiscale vessel-background decoupling module that separates vascular and background features, and a clinical-guided residual rectification module that extracts retina-specific information complementary to clinical variables. Evaluated on 3619 participants for advanced CKM (ACKM), chronic kidney disease (CKD), and cardiovascular disease (CVD), our full model significantly outperformed all traditional clinical models, achieving AUC improvements from the best-performing clinical models (0.758 for ACKM, 0.692 for CKD, and 0.581 for CVD) to 0.812, 0.791, and 0.743, respectively. Ablation studies confirmed independent contributions of both modules, and feature perturbation tests verified reliance on genuine retinal spatial structures. This noninvasive, scalable framework offers a simplified alternative to conventional CKM evaluation, with strong potential for deployment in primary care and resource-limited settings.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Retina-enhanced multimodal deep learning for assessment of cardiovascular-kidney metabolic syndrome related outcomes

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Di Cheng,
  • Dan Liu,
  • Keqing Dong,
  • Yuren Zhang,
  • Sichao Cheng,
  • Yuqian Bao,
  • Huating Li,
  • Zhen Liang,
  • Nan Jiang,
  • Weiping Jia

摘要

Cardiovascular-kidney-metabolic (CKM) syndrome requires integrated risk assessment, but traditional evaluation depends on numerous clinical variables and specialized examinations, limiting scalability. We developed a retina-enhanced multimodal deep learning framework integrating fundus photographs with eight routine clinical variables, incorporating two novel modules without manual segmentation: a multiscale vessel-background decoupling module that separates vascular and background features, and a clinical-guided residual rectification module that extracts retina-specific information complementary to clinical variables. Evaluated on 3619 participants for advanced CKM (ACKM), chronic kidney disease (CKD), and cardiovascular disease (CVD), our full model significantly outperformed all traditional clinical models, achieving AUC improvements from the best-performing clinical models (0.758 for ACKM, 0.692 for CKD, and 0.581 for CVD) to 0.812, 0.791, and 0.743, respectively. Ablation studies confirmed independent contributions of both modules, and feature perturbation tests verified reliance on genuine retinal spatial structures. This noninvasive, scalable framework offers a simplified alternative to conventional CKM evaluation, with strong potential for deployment in primary care and resource-limited settings.