ECG and cardiac MRI foundation model aligned by contrastive learning, predicting atrial fibrillation and heart failure risk from ECG alone.
No providers recorded yet. Browse all providers
A 12-lead electrocardiogram is cheap, fast and recorded in almost every clinic, but it measures the heart's electrical activity only indirectly. Cardiac MRI shows chamber size, wall thickness and motion directly, which is exactly the structural remodelling that precedes atrial fibrillation and heart failure, yet it is expensive and rarely available. CARDIAC-FM (Contrastive Alignment for Risk Detection In Cardiovascular Foundation Model) uses the two together during training so that the ECG encoder learns to carry information the MRI can see, and then lets that ECG encoder be deployed on its own wherever MRI is absent.
The model was developed at the University of Washington by biostatisticians, computer scientists and the Cardiovascular Health Research Unit, with collaborators at UCSF and the University of North Carolina at Chapel Hill. It was posted to medRxiv in March 2026, with a revised version in September 2026, and has not yet been peer reviewed. It is unrelated to the Cardiac Sensing Foundation Model, which pairs ECG with photoplethysmography and clinical text rather than with imaging.
A new cardiac MRI encoder is pretrained from scratch by masked reconstruction of raw cine sequences, without labels, then aligned contrastively with an ECG encoder initialized from ECG-FM; a lightweight head is fine-tuned for each downstream task.
The MRI encoder is a convolution–transformer masked autoencoder that the authors describe as inspired by MultiMAE and by CineMA, extended from single-frame 2D encoding to spatiotemporal encoding with VideoMAE-style tube masking. Each of five view streams is tokenized by a view-specific convolutional stem into a 4×14×14 grid of 784 tokens; visible tokens from all streams pass through a ViT-base encoder (width 768, 12 layers, 12 heads), and a lightweight decoder reconstructs the masked patches. Per-view pooled embeddings are concatenated into a 2,304-dimensional study-level vector. The ECG encoder is ECG-FM's convolutional feature extractor followed by a BERT-Large transformer, fully fine-tuned during alignment.
UK Biobank provided 77,326 imaging visits from 71,144 participants; after quality control, 57,609 visits had paired, high-quality ECG and MRI, and contrastive training used the 34,596 paired samples in the training partition. On the held-out UK Biobank test set, the ECG-only model reached an AUROC of 0.769 for five-year incident atrial fibrillation and 0.817 for heart failure, rising to 0.816 and 0.830 with MRI. Models were compared against ECGFounder, ECG-FM, DeepECG-SL and DeepECG-SSL fine-tuned with the same protocol. Applied without any cohort-specific fine-tuning to the Cardiovascular Health Study and the Multi-Ethnic Study of Atherosclerosis, it outperformed three of the four comparators for both outcomes. Late fusion with clinical risk scores (CHARGE-AF for atrial fibrillation, PREVENT-HF for heart failure) raised heart failure AUROC from 0.751 to 0.792 in CHS and from 0.789 to 0.826 in MESA. ECG-predicted MRI phenotypes correlated with measured values at r = 0.45–0.79 in UK Biobank.
The main use is cardiovascular risk stratification in settings where only an ECG is routinely available: flagging patients at elevated five-year risk of atrial fibrillation or heart failure, or estimating left ventricular mass and chamber volumes where imaging is impractical. Released checkpoints fine-tuned for atrial fibrillation and heart failure can score new ECGs directly, and the aligned encoder can be fine-tuned on other outcomes; in CHS and MESA, fine-tuning on 10–20% of each cohort adapted it to myocardial infarction, ischaemic stroke, cardiovascular death and all-cause mortality.
CARDIAC-FM shows a practical pattern for clinical multimodal learning: use a deeply phenotyped but scarce modality during pretraining to enrich a cheap, scalable one, and keep a deployment path that does not require the scarce one. Validation in two US cohorts that differ from UK Biobank in age and ancestry strengthens the transfer claim. The authors are explicit about limits: absolute risk calibration shifted across cohorts and needed local recalibration, ECG-derived MRI phenotypes approximate rather than replace imaging, and prospective studies are still needed to show clinical benefit. Code is released under the MIT license and checkpoints are on Hugging Face, although the ECG-FM backbone must be obtained from its own release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.