Scholarlink

Background <p>Cardiovascular diseases (CVDs) remain the leading cause of global mortality, yet scalable, non-invasive screening tools are limited. We systematically evaluate the ability of artificial-intelligence-derived retinal biomarkers to simultaneously quantify multiple CVD phenotypes with clinically meaningful accuracy across diverse populations.</p> Methods <p>Following PRISMA-ScR and Arksey &amp; O’Malley, we searched PubMed, Scopus, Web of Science, and Embase through April 2025, screening 643 records. Twenty-one studies met inclusion; we extracted and synthesized data on algorithm class, imaging modality, external-validation strategies, and fairness metrics.</p> Results <p>From 643 screened records, 21 studies published between 2018 and 2025 were included; six (29%) were multi-national, with China contributing the largest single-country share (24%). Deep Learning (DL) convolutional neural networks (CNNs) dominated algorithmic approaches (57%), followed by transformer-based or hybrid models (38%). Retinal artificial intelligence (Retinal-AI) achieved area under the receiver operating characteristic curve (AUROC) values of 0.89–0.90 for 10-year atherosclerotic cardiovascular disease (ASCVD) risk estimation, 0.97–0.99 for prevalent coronary artery disease (CAD) detection, and 0.64–0.74 for prediction of incident major adverse cardiac events. Multimodal fusion of fundus imaging with basic clinical variables further improved AUROCs by 0.035–0.12 and increased net reclassification by 12–18% across endpoints. Fewer than 10% of studies shared code or datasets, and two-thirds lacked external validation.</p> Conclusion <p>Retinal-AI provides a single-image, multi-risk screening platform suitable for prospective, multi-ethnic trials. Our review establishes a benchmarked evidence base and FAIR-compliant reporting framework to accelerate regulatory qualification and clinical adoption.</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

Through the eye to the heart: a scoping review of artificial intelligence in retinal imaging for cardiovascular disease assessment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月26日

查看全文
  • Hesamaddin Kamalzadeh,
  • Farid Khorrami,
  • Asma Ahmadi,
  • Seyed Reza Mirlohi,
  • Mobina Vatankhah,
  • Niloofar Choobin

摘要

Background

Cardiovascular diseases (CVDs) remain the leading cause of global mortality, yet scalable, non-invasive screening tools are limited. We systematically evaluate the ability of artificial-intelligence-derived retinal biomarkers to simultaneously quantify multiple CVD phenotypes with clinically meaningful accuracy across diverse populations.

Methods

Following PRISMA-ScR and Arksey & O’Malley, we searched PubMed, Scopus, Web of Science, and Embase through April 2025, screening 643 records. Twenty-one studies met inclusion; we extracted and synthesized data on algorithm class, imaging modality, external-validation strategies, and fairness metrics.

Results

From 643 screened records, 21 studies published between 2018 and 2025 were included; six (29%) were multi-national, with China contributing the largest single-country share (24%). Deep Learning (DL) convolutional neural networks (CNNs) dominated algorithmic approaches (57%), followed by transformer-based or hybrid models (38%). Retinal artificial intelligence (Retinal-AI) achieved area under the receiver operating characteristic curve (AUROC) values of 0.89–0.90 for 10-year atherosclerotic cardiovascular disease (ASCVD) risk estimation, 0.97–0.99 for prevalent coronary artery disease (CAD) detection, and 0.64–0.74 for prediction of incident major adverse cardiac events. Multimodal fusion of fundus imaging with basic clinical variables further improved AUROCs by 0.035–0.12 and increased net reclassification by 12–18% across endpoints. Fewer than 10% of studies shared code or datasets, and two-thirds lacked external validation.

Conclusion

Retinal-AI provides a single-image, multi-risk screening platform suitable for prospective, multi-ethnic trials. Our review establishes a benchmarked evidence base and FAIR-compliant reporting framework to accelerate regulatory qualification and clinical adoption.