← News

Models & toolsSources: arXiv · 18 Nov 2026

CardioLoom ECG model pretrained on 2.78 million report pairs

A retrospective study describes a report-guided electrocardiogram foundation model trained on millions of ECG-report pairs, with label-efficient diagnosis and external testing across nine cohorts.

In plain words

Researchers pretrained an AI model called CardioLoom to read electrocardiograms, using the text of routine ECG reports as training supervision instead of manually assigned labels. The authors report it matched or beat a comparable self-supervised method and needed labels from only a fraction of patients to reach strong diagnostic scores. Because the study is retrospective, these are research results, not a deployed clinical tool.

What was announced

The pretraining set covered 2,781,289 ECG-report pairs from 1,298,359 patients, while downstream development used a separate cohort of 721,317 ECGs from 233,798 patients. Evaluation spanned 89 targets with external testing in nine cohorts. On 45 internal ECG diagnoses, the authors report a macro area under the precision-recall curve of 0.624 for CardioLoom versus 0.615 for DeepECG-SSL; across 203 external evaluations the means were 0.425 and 0.424, a near tie.

The label-efficiency result stands out: with labels from just 10% of training patients, CardioLoom reached macro-AUPRC 0.559, exceeding fully supervised training with all labels at 0.549. For heart failure with reduced ejection fraction, AUPRCs were 0.690, 0.641 and 0.677 at Zhongshan, Shiyuan and BIDMC. Cardiac amyloidosis screening showed a 4.5% positive predictive value at a threshold targeting 99% specificity. The authors note transfer varied across echocardiographic phenotypes, and the low amyloidosis positive predictive value means most flagged cases would not have the condition.

What it does not show

  • The evidence is the paper's abstract only; methods, cohort details and validation of the reported figures cannot be checked from this excerpt.
  • All results are retrospective and author-reported; no independent confirmation or evidence of clinical deployment is provided.
  • External performance was close to the comparator (0.425 versus 0.424 mean AUPRC), and transfer across echocardiographic phenotypes was variable per the authors.

Sources

Source-support review completed 19 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 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.