Scholarlink

Background <p>Coronary artery disease (CAD), remains a leading global cause of mortality. Invasive coronary angiography (CAG), the diagnostic gold standard, is unsuitable for large-scale screening due to its cost and procedural risks. Retinal microvascular changes, reflecting systemic vascular pathology via the retino-coronary axis, offer a promising noninvasive alternative. Optical coherence tomography angiography (OCTA) enables high-resolution retinal imaging but lacks robust computational frameworks for accurate CAD prediction.</p> Objective <p>To develop and validate an artificial intelligence (AI)-driven CAD prediction model by integrating OCTA-derived retinal features with electronic health record (EHR) data.</p> Methods <p>This retrospective cohort study included 542 patients undergoing both OCTA and invasive coronary angiography between July 2022 and October 2023. A Transformer-based multi-scale style learning algorithm was developed to extract features from 3 × 3 mm<sup>2</sup> macular OCTA images, achieving joint segmentation of the foveal avascular zone (FAZ) and retinal vessel (RV) by leveraging task-specific characteristics, while simultaneously quantifying 98 retinal vascular imaging biomarkers. Relevant predictors were selected via LASSO regression and modeled using restricted cubic splines. An XGBoost classifier was trained on the combined OCTA-EHR feature set and compare it with four baseline deep learning network models. Model performance was evaluated using AUC, sensitivity, specificity, calibration, and decision curve analysis.</p> Results <p>The multimodal model demonstrated superior discriminative power (AUC = 0.850; sensitivity = 0.806; specificity = 0.667), significantly outperforming OCTA-only models (AUC = 0.512; sensitivity = 0.581; specificity = 0.417). Key predictors included Glycosylated Hemoglobin (Ghb), Hypersensitive C-Reactive Protein (hs-CRP), and OCTA-derived vascular density metrics (e.g., vessel length density in temporal-internal macular sectors).</p> Conclusion <p>Integration of retinal OCTA biomarkers with EHR data enables accurate, noninvasive CAD prediction. This approach validates the retina-coronary axis and establishes a scalable screening paradigm for subclinical atherosclerosis.</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

Prediction of coronary artery disease using retinal optical coherence tomography angiography imaging and electronic health records: a multimodal machine learning approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Ziyi Wu,
  • Lili Ji,
  • Wenbin Tang,
  • Shijia Zhou,
  • Qifeng Yan,
  • Xiaomin Chen,
  • Jinjin Zhu,
  • Jianjun Ma,
  • Yitian Zhao,
  • Wenming He

摘要

Background

Coronary artery disease (CAD), remains a leading global cause of mortality. Invasive coronary angiography (CAG), the diagnostic gold standard, is unsuitable for large-scale screening due to its cost and procedural risks. Retinal microvascular changes, reflecting systemic vascular pathology via the retino-coronary axis, offer a promising noninvasive alternative. Optical coherence tomography angiography (OCTA) enables high-resolution retinal imaging but lacks robust computational frameworks for accurate CAD prediction.

Objective

To develop and validate an artificial intelligence (AI)-driven CAD prediction model by integrating OCTA-derived retinal features with electronic health record (EHR) data.

Methods

This retrospective cohort study included 542 patients undergoing both OCTA and invasive coronary angiography between July 2022 and October 2023. A Transformer-based multi-scale style learning algorithm was developed to extract features from 3 × 3 mm2 macular OCTA images, achieving joint segmentation of the foveal avascular zone (FAZ) and retinal vessel (RV) by leveraging task-specific characteristics, while simultaneously quantifying 98 retinal vascular imaging biomarkers. Relevant predictors were selected via LASSO regression and modeled using restricted cubic splines. An XGBoost classifier was trained on the combined OCTA-EHR feature set and compare it with four baseline deep learning network models. Model performance was evaluated using AUC, sensitivity, specificity, calibration, and decision curve analysis.

Results

The multimodal model demonstrated superior discriminative power (AUC = 0.850; sensitivity = 0.806; specificity = 0.667), significantly outperforming OCTA-only models (AUC = 0.512; sensitivity = 0.581; specificity = 0.417). Key predictors included Glycosylated Hemoglobin (Ghb), Hypersensitive C-Reactive Protein (hs-CRP), and OCTA-derived vascular density metrics (e.g., vessel length density in temporal-internal macular sectors).

Conclusion

Integration of retinal OCTA biomarkers with EHR data enables accurate, noninvasive CAD prediction. This approach validates the retina-coronary axis and establishes a scalable screening paradigm for subclinical atherosclerosis.