Loading…
Twelve-lead ECG foundation model aligned to clinician-overread report text, adapting to new cardiac tasks from as few as ten labels.
No providers recorded yet. Browse all providers
A cardiologist's overread of a 12-lead electrocardiogram is a sentence, not a checkbox — "normal sinus rhythm, left axis deviation, non-specific T wave changes." A supervised ECG classifier trained for one diagnosis throws nearly all of that away and keeps a single bit per recording, which is why such models need large curated label sets for every new task. ECG-CLIP, built at the Scripps Research Translational Institute (Scripps Research) and published in The Lancet Digital Health on 1 September 2026, treats the overread itself as the supervisory signal.
Pretraining runs in two stages on one institutional corpus. The first stage is masked reconstruction: 75% of the ECG is hidden and the model learns to regenerate it, forcing the encoder to internalise ordinary cardiac physiology before it ever sees a diagnosis. The second stage keeps that reconstruction objective and adds a contrastive term that pulls each ECG embedding towards the embedding of its own paired report and away from everybody else's. The reports are clinician-overread — the physician-validated interpretation after review and editing of the automated analysis — rather than the machine-generated or structured label strings used by earlier signal-text work, which the authors identify as the study's central novelty.
Other catalog entries pair ECG waveforms with report text, including D-BETA and MELP. What separates ECG-CLIP is its evaluation design: pretraining happens once on Scripps Health data, and every headline result is measured on MIMIC-IV, a different health system's archive the model never saw.
The ECG encoder is a 12-layer transformer with eight attention heads and 512-dimensional embeddings, adapted to one-dimensional signals by tokenising each 10-second, 500 Hz recording into 80 non-overlapping 500-sample windows across the eight independent leads. A six-layer decoder handles reconstruction, and the text tower is BioClinicalBERT. Pretraining used the Scripps Health GE MUSE archive: 1,704,047 anonymised 12-lead ECGs from 542,288 patients recorded between January 2008 and January 2019, of which 1,679,884 from 538,028 patients survived filtering. Downstream, ECG-CLIP reached an AUC of 0.963 for acute myocardial infarction detection, 0.852 for cardiac amyloidosis, 0.861 for hypertrophic cardiomyopathy, and 0.867 for predicting atrial fibrillation from a normal sinus rhythm tracing — ahead of a ResNet baseline, a conventional-feature linear model and OpenAI CLIP on all four. Against ECG-pretrained comparators including ECGFounder, the gap concentrates in low-label regimes: at ten positive labels, 0.910 versus 0.884 for infarction detection and 0.777 versus 0.765 for atrial fibrillation prediction. Fine-tuning attaches a linear head to the final representation and updates the whole encoder with layer-wise learning-rate decay.
The practical target is any cardiac question for which a site has plenty of ECGs but few confirmed cases — rare cardiomyopathies such as amyloidosis, under 0.4% prevalent in the training set, are the clearest example. Because the pretrained encoder carries physiological structure learned without labels, a hospital can stand up a screening model from a few dozen adjudicated tracings rather than commissioning a labelling campaign. The single-lead results extend the same reasoning to patch monitors and consumer wearables. SHAP saliency maps localise the model's attributions to ST segments in the leads a cardiologist would inspect, which matters for anyone who has to justify a prediction clinically.
ECG-CLIP arrives into a crowded field of ECG foundation models and argues a specific point within it: the text a clinician already writes is a richer pretraining target than the structured labels the field has mostly used, and the payoff shows up as label efficiency rather than a higher ceiling on abundant data. On full datasets most comparators converge, and one of them edges ECG-CLIP on atrial fibrillation prediction; the separation sits in the low-label regime. Everything reported is retrospective, and the authors state that prospective clinical trials will be required to establish applicability in real-world settings. Training and fine-tuning code is public under a non-commercial licence; the pretrained weights and the Scripps corpus are not released, so reproducing the checkpoint requires an equivalent report-paired ECG archive of one's own.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.