Cardiac MRI-informed electrocardiography enables biologically grounded cardiomyopathy detection

Zhengyao Ding, Kewen Wang, Yujian Hu, Ziyu Li, Yiheng Mao, Yue Wang, Haitao Li, Dengnan Zhang, Xuanhong Liu, Jing Wang, Yue Chen, Longbo Wang, Mengjia Chen, Chengchen Zhao, Xuesen Chu, Yubin Zhang, Mor Peleg, Judith Somekh, Mingxuan Yuan, Hui Yan, Yue Gu, Fan Yi, Zhilong Jia, Zhentao Tang, Xi Hu, Fei Wu, Hongkun Zhang, Ting Chen, Zhengxing HuangarXiv 2026 · arXiv

TL;DR

MyoECG is a cross-modal deep learning framework that "teaches" standard 12-lead ECGs to recognize the structural and functional signatures typically only visible on Cardiac MRI (CMR). By distilling CMR-derived phenotypes into ECG representations, the model achieves state-of-the-art cardiomyopathy detection and subtype classification using only low-cost, widely available electrical signals.

What Problem Does This Paper Address?

Cardiomyopathies—diseases of the heart muscle like dilated (DCM) or hypertrophic (HCM) cardiomyopathy—are leading causes of heart failure and sudden death, yet they often remain undiagnosed until a major clinical event occurs.

The clinical "gold standard" for diagnosis is Cardiac MRI (CMR), which provides exquisite detail on heart structure and function. However, CMR is expensive, time-consuming, and inaccessible for population-scale screening. In contrast, the 12-lead electrocardiogram (ECG) is ubiquitous and inexpensive but primarily reflects electrical activity. While AI has previously been used to detect heart disease from ECGs, these models often act as "black boxes" that rely on dataset-specific shortcuts rather than grounded biological features, limiting their clinical trust and generalizability.

Methods

The authors developed MyoECG using a teacher-student representation learning strategy to bridge the gap between electrical signals and physical structure.

  1. CMR Teacher Construction: A CMR encoder was trained to predict 27 quantitative phenotypes (e.g., ventricular volumes, wall thickness, ejection fraction) from four-chamber cine MRI sequences.
  2. Cross-Modal Alignment (Pretraining): Using paired ECG-CMR data, an ECG encoder (based on the ST-MEM transformer architecture) was trained to align its latent representation with the CMR teacher. This involved three loss functions:
  3. Semantic Distillation: Forcing ECG embeddings to be close to CMR embeddings in latent space.
  4. Phenotype Regression: Training the ECG encoder to directly predict the 27 CMR-derived phenotypes.
  5. Relational Alignment: Ensuring that if two patients have similar heart structures on MRI, their ECG embeddings are also similar.
  6. Supervised Fine-Tuning: The "CMR-informed" ECG encoder was then fine-tuned on clinical labels to detect cardiomyopathy and classify it into three subtypes: DCM, HCM, and restrictive cardiomyopathy (RCM).
Overview of the MyoECG framework for CMR-informed ECG representation learning

Datasets

The study utilized a massive multi-centre dataset comprising over 1.8 million ECGs:

  • Paired Training: 45,241 paired ECG-CMR examinations from the UK Biobank (UKB) and Fuwai Central China Cardiovascular Hospital (FCCCH).
  • External Validation: Six independent datasets including the Harvard–Emory ECG Database (HEEDB), MIMIC-IV-ECG, and several large Chinese hospital cohorts (SRRSH, FAHZU, SAHZU).
  • Reader Study: A dedicated dataset (FCCCH-RS) of 403 ECGs used to evaluate human-AI collaboration.

Key Insights and Ideas

  • Biologically Grounded "Vision": Unlike standard ECG-AI, MyoECG is explicitly "informed" by anatomy. Saliency maps (Figure 4b) show the model focuses on the QRS complex and T-wave—regions known to reflect ventricular depolarization and repolarization—rather than random noise.
  • The "Genetic Ablation" Proof: To prove the model isn't just finding shortcuts, the authors performed a GWAS on the MyoECG score. They found strong genetic correlations with cardiomyopathy. Crucially, when they adjusted the analysis for the 27 CMR traits, this correlation vanished. This confirms that the model’s predictive power comes specifically from the structural/functional information it "inherited" from the MRI teacher.
  • Superior Generalizability: By learning fundamental biological representations rather than just fitting labels, MyoECG maintained high performance across seven different datasets with vastly different disease prevalences (0.2% to 54.7%).
Feature separability and saliency analysis

Key Results

  • Detection Performance: MyoECG achieved AUROCs ranging from 0.800 to 0.974 for cardiomyopathy detection, consistently outperforming five state-of-the-art ECG foundation models (like ECGFounder and ST-MEM).
  • Subtype Classification: For distinguishing between DCM, HCM, and RCM, the model achieved a macro-AUROC of 0.941 in internal testing and remained robust in external cohorts.
  • Clinical Utility: In a reader study with six cardiologists, AI assistance significantly boosted diagnostic accuracy. Mean sensitivity increased by 0.109 and specificity by 0.056.
  • Biological Links: The MyoECG score was associated with 69 circulating proteins (including the heart stress marker NT-proBNP) and identified four independent genomic loci (including NFIA and KCNQ1) linked to heart disease.
Performance across different datasets and prevalence levels

Limitations and Future Work

  • Retrospective Nature: All evaluations were performed on historical data. Prospective clinical trials are needed to see if this actually improves patient outcomes in the real world.
  • Limited MRI Input: The CMR teacher only used "cine" (motion) sequences. It did not incorporate late gadolinium enhancement (LGE), which is used to detect heart scarring (fibrosis). Adding this could further improve detection of specific disease stages.
  • Low PPV in General Population: While MyoECG is much better than previous models, the Positive Predictive Value (PPV) remains low in very low-prevalence settings (like the general public). The authors suggest it is best used as a "triage" tool to prioritize patients for follow-up imaging rather than a definitive diagnostic test.

Summary updated

Abstract

Precise myocardial phenotyping at population scale is limited by the cost and accessibility of cardiac magnetic resonance (CMR), whereas electrocardiography (ECG) is widely available but reflects electrical rather than structural and functional information. Here we develop MyoECG, a cross-modal representation learning framework that distils myocardial structural and functional knowledge from CMR into standard 12-lead ECGs, enabling imaging-informed cardiomyopathy detection and subtype classification from ECGs alone. MyoECG was pretrained on 45,241 paired ECG-CMR examinations from UK Biobank and a tertiary cardiovascular centre, then developed and validated across more than 1.8 million ECGs from seven independent datasets. Across internal and external datasets spanning prevalences of 0.2-54.7%, MyoECG outperformed five pretrained ECG foundation models, with AUROCs of 0.800-0.974 for cardiomyopathy detection. In a multi-reader study of six cardiologists, MyoECG assistance improved mean sensitivity by 0.109 (95% CI, 0.041-0.175; P = 0.009) and specificity by 0.056 (95% CI, 0.004-0.108; P = 0.040). In 41,128 UK Biobank participants, the MyoECG score showed genetic correlations with cardiomyopathy (rg = 0.409) and dilated cardiomyopathy (rg = 0.371), genome-wide associations at four loci, including regions mapped to NFIA and KCNQ1, and associations with 69 circulating proteins related to myocardial stress, cell adhesion, vascular biology and metabolism. The cardiomyopathy correlation attenuated to non-significance after adjustment for 27 measured CMR traits (rg = 0.131; P = 0.143), indicating that the score captures CMR-derived information rather than dataset-specific shortcuts. MyoECG offers a scalable route to CMR-informed cardiomyopathy assessment and a strategy for transferring knowledge from resource-intensive imaging to widely deployable physiological signals.
BibTeX
@misc{ding2026cardiac,
  author = {Zhengyao Ding and Kewen Wang and Yujian Hu and Ziyu Li and Yiheng Mao and Yue Wang and Haitao Li and Dengnan Zhang and Xuanhong Liu and Yue Wang and Jing Wang and Yue Chen and Longbo Wang and Mengjia Chen and Chengchen Zhao and Xuesen Chu and Yubin Zhang and Mor Peleg and Judith Somekh and Mingxuan Yuan and Hui Yan and Yue Gu and Fan Yi and Zhilong Jia and Zhentao Tang and Xi Hu and Fei Wu and Hongkun Zhang and Ting Chen and Zhengxing Huang},
  title = {{Cardiac MRI-informed electrocardiography enables biologically grounded cardiomyopathy detection}},
  year = {2026},
  eprint = {2611.23925},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23925},
}
Figure