Scholarlink

Background <p>Cardiovascular–kidney–metabolic (CKM) syndrome refers to the co-occurrence of obesity, diabetes, chronic kidney disease (CKD), and cardiovascular disease. However, it is underdiagnosed due to silent clinical nature of the early stages of its components and subsequent siloed medical care. Electrocardiography (ECG) is an inexpensive and widely available diagnostic tool but its utility in automated detection of CKM syndrome has not been previously explored.</p> Objective <p>To develop and evaluate deep learning models for predicting CKM syndrome using scanned limb and augmented limb leads ECGs images in people with diabetes.</p> Methods <p>Clinical data of adults with type 1 or type 2 diabetes enrolled in the prospective Silesia Diabetes-Heart Project were analyzed. CKM syndrome was defined by the presence of either CKD [estimated glomerular filtration rate (eGFR) &lt; 60&#xa0;mL/min/1.73&#xa0;m<sup>2</sup> and/or urine albumin to creatinine ratio (UACR) ≥ 30&#xa0;mg/g) or established CVD. High-resolution scanned ECG tracings were processed into lead-specific inputs to train ResNet-50-based convolutional neural network (CNN) models, including single- and dual-channel variants. Class imbalance was addressed using resampling strategies, and model performance was evaluated using standard classification metrics, including area under the receiver operating characteristic curve (AUROC), with confidence intervals estimated by bootstrap resampling.</p> Results <p>Among 2779 participants, 492 (17.7%) met criteria for CKM syndrome. The best-performing individual model was a dual-channel ResNet-50 with soft voting ensemble, achieving an AUROC of 0.8199 (95% CI 0.7549–0.8795), F1-score of 0.7213 (95% CI 0.6404–0.7957), accuracy of 0.7385, and balanced precision and recall. Ensemble models consistently outperformed individual architectures, particularly in handling class imbalance and improving generalization.</p> Conclusion <p>Deep learning applied to scanned ECG image data predicts CKM syndrome in individuals with diabetes with reasonable accuracy. This approach holds promise as a low-cost, scalable risk stratification tool and which could augment clinical decision-making in settings particularly with limited access to advanced diagnostics.</p> <p><i>Trial registration</i> The study is registered at ClinicalTrials.gov (NCT05626413).</p> Graphical abstract <p></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

Deep learning analysis of ECGs detects Cardiovascular–Kidney–Metabolic syndrome burden in people with diabetes: a report from the Silesia Diabetes-Heart Project

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月29日

查看全文
  • Oliwia Janota-Sosińska,
  • Qinkai Yu,
  • Krzysztof Irlik,
  • Hanna Kwiendacz,
  • Aleksandra Włosowicz-Momot,
  • Patrycja Pabis,
  • Wiktoria Wójcik,
  • Anna Olejarz,
  • Julia Piaśnik,
  • Uazman Alam,
  • Yalin Zheng,
  • Janusz Gumprecht,
  • Gregory Y. H. Lip,
  • Katarzyna Nabrdalik

摘要

Background

Cardiovascular–kidney–metabolic (CKM) syndrome refers to the co-occurrence of obesity, diabetes, chronic kidney disease (CKD), and cardiovascular disease. However, it is underdiagnosed due to silent clinical nature of the early stages of its components and subsequent siloed medical care. Electrocardiography (ECG) is an inexpensive and widely available diagnostic tool but its utility in automated detection of CKM syndrome has not been previously explored.

Objective

To develop and evaluate deep learning models for predicting CKM syndrome using scanned limb and augmented limb leads ECGs images in people with diabetes.

Methods

Clinical data of adults with type 1 or type 2 diabetes enrolled in the prospective Silesia Diabetes-Heart Project were analyzed. CKM syndrome was defined by the presence of either CKD [estimated glomerular filtration rate (eGFR) < 60 mL/min/1.73 m2 and/or urine albumin to creatinine ratio (UACR) ≥ 30 mg/g) or established CVD. High-resolution scanned ECG tracings were processed into lead-specific inputs to train ResNet-50-based convolutional neural network (CNN) models, including single- and dual-channel variants. Class imbalance was addressed using resampling strategies, and model performance was evaluated using standard classification metrics, including area under the receiver operating characteristic curve (AUROC), with confidence intervals estimated by bootstrap resampling.

Results

Among 2779 participants, 492 (17.7%) met criteria for CKM syndrome. The best-performing individual model was a dual-channel ResNet-50 with soft voting ensemble, achieving an AUROC of 0.8199 (95% CI 0.7549–0.8795), F1-score of 0.7213 (95% CI 0.6404–0.7957), accuracy of 0.7385, and balanced precision and recall. Ensemble models consistently outperformed individual architectures, particularly in handling class imbalance and improving generalization.

Conclusion

Deep learning applied to scanned ECG image data predicts CKM syndrome in individuals with diabetes with reasonable accuracy. This approach holds promise as a low-cost, scalable risk stratification tool and which could augment clinical decision-making in settings particularly with limited access to advanced diagnostics.

Trial registration The study is registered at ClinicalTrials.gov (NCT05626413).

Graphical abstract