MyoECG transfers cardiac MRI knowledge into standard ECGs for cardiomyopathy detection
Researchers describe a cross-modal learning framework that distils cardiac magnetic resonance information into 12-lead ECGs, reporting improved cardiomyopathy detection across seven datasets.
In plain words
Cardiac MRI scans show the heart's structure and function in detail but are costly, while ECGs are cheap and everywhere but capture only electrical signals. The MyoECG framework was trained to embed MRI-derived knowledge of the heart muscle into ECG readings. The result is a tool that can flag cardiomyopathy, a disease of the heart muscle, from an ECG alone.
What was announced
The team pretrained MyoECG on 45,241 paired ECG-CMR examinations from UK Biobank and a tertiary cardiovascular centre, then validated it on more than 1.8 million ECGs from seven independent datasets. Reported AUROCs for cardiomyopathy detection ranged from 0.800 to 0.974 across datasets with disease prevalences of 0.2-54.7%, and the authors state it outperformed five pretrained ECG foundation models.
Supporting evidence goes beyond classification scores. In a reader study with six cardiologists, MyoECG assistance raised mean sensitivity by 0.109 and specificity by 0.056, both statistically significant. In 41,128 UK Biobank participants, the score showed genetic correlations with cardiomyopathy and dilated cardiomyopathy, genome-wide associations at four loci including NFIA and KCNQ1, and links to 69 circulating proteins tied to myocardial stress and metabolism. When the analysis adjusted for 27 measured CMR traits, the genetic correlation attenuated to non-significance, which the authors interpret as evidence the score captures CMR-derived information rather than dataset-specific shortcuts.
What it does not show
- Evidence is limited to the paper's abstract; full methods, cohort details and validation procedures could not be checked.
- All reported performance, reader-study and genetic findings are author-reported and have not been independently confirmed.
- The abstract does not describe regulatory status, clinical deployment or availability of the tool.
For security and IT teams
- Scope
- The source is the arXiv abstract only; it covers pretraining data, validation datasets, reported metrics and genetic analyses but not implementation or deployment details.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.