Pairing cardiac magnetic resonance with electrocardiograms (ECGs) improved artificial intelligence (AI) assessments of cardiovascular diseases

1. Ding and colleagues developed a deep-learning model that learns relationships between ECGs and paired cardiac magnetic resonance images.

2. The model outperformed models that exclusively analyze ECGs for cardiovascular disease detection.

Evidence Rating Level: 2 (Good)

Study Rundown: Cardiac magnetic resonance (CMR) imaging provides a comprehensive evaluation of cardiac morphology and function, but its high cost impedes widespread adoption. Meanwhile, ECGs are low-cost and widely used but are limited in assessing cardiac structure and function. Ding and colleagues developed CardioNets, a deep-learning model that learns relationships between paired ECGs and CMR images to improve cardiac assessment using ECG input alone. The model also generates synthetic CMR images from ECG readings. Researchers evaluated the cardiac measurement prediction, cardiovascular disease detection, image synthesis, and physician performance of CardioNets. The study found that CardioNets improved prediction of 82 cardiac measurements compared with the strongest ECG-only baseline, with average R² values of 0.310 versus 0.269. For cardiomyopathy detection, the area under the receiver operating characteristic curve (AUROC) was 0.890 versus 0.867 for the best ECG baseline. Additionally, synthesized CMR images by CardioNets achieved a structural similarity index of 0.205 compared to 0.122 for the current gold standard. This study demonstrated that CardioNets has the potential to improve ECG translation to CMR and cardiovascular disease detection.

Click here to read the study in NEJM AI

Relevant Reading: ECG-Image-Kit: a synthetic image generation toolbox to facilitate deep learning-based electrocardiogram digitization

In-Depth [retrospective cohort]: 159,819 samples from the UK Biobank, MIMIC-IV-ECG, and two Chinese hospitals were used for model training and validation. Measurement prediction used 28,542 participants with complete data on 82 cardiac indices. Diagnostic labels were retrieved from hospital records or discharge diagnoses. The main outcomes were R² for measurement prediction, AUROC for disease detection, and structural similarity index for synthesized images. CardioNets’ measurement prediction achieved R² = 0.310 (95% confidence interval [CI], 0.304-0.315), versus 0.269 (95% CI, 0.265-0.272) for the best ECG baseline. Its AUROC for predicting cardiomyopathy in the UK Biobank dataset was 0.890 (95% CI, 0.836-0.944) versus 0.867 (95% CI, 0.812-0.921) for the best ECG baseline. Similarly, CardioNets’ AUROC for predicting pulmonary hypertension in the MIMIC database was superior to the best ECG baseline (0.879 versus 0.853). In terms of structural similarity of synthesized images, CardioNets consistently outperformed baseline models, achieving a similarity index of 0.205 (95% CI, 0.202-0.208) versus 0.122 (95% CI, 0.120-0.125) for the best-performing baseline model. This study was limited by limited discrimination of specific cardiomyopathy subtypes and imperfect synthetic image fidelity. Nonetheless, this study provided promising evidence that a deep-learning model such as CardioNets has the potential to accurately translate ECGs into CMR images and improve cardiovascular disease detection.

Image: PD

©2026 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.