An ECG 'foundation model' matched rivals using a fraction of the labelled data
Trained on 1.7 million ECGs paired with clinicians' report text, ECG-CLIP reached the same accuracy as the best comparator with about 90% less training data — a bid at the field's labelling bottleneck.
| Group | Value (value) |
|---|---|
| Acute MI detection — ECG-CLIP | 0.91 (0.9 to 0.92) |
| Acute MI detection — next best | 0.88 (0.88 to 0.89) |
| Cardiac amyloidosis — ECG-CLIP | 0.79 (0.77 to 0.81) |
| Cardiac amyloidosis — next best | 0.78 (0.75 to 0.8) |
| AF prediction — ECG-CLIP | 0.78 (0.77 to 0.78) |
| AF prediction — next best | 0.77 (0.76 to 0.77) |
Machine learning has been reading electrocardiograms (ECGs) for years, sometimes spotting things clinicians cannot — a 2019 model famously flagged patients prone to atrial fibrillation from an ECG taken while their rhythm looked normal [s2]. But most such tools are trained one task at a time on large sets of hand-labelled examples, which is slow and expensive to assemble. A study in The Lancet Digital Health, published in the September issue, describes a "foundation model" approach meant to loosen that bottleneck [s1].
The idea
Rather than learning a single labelled task, a foundation model is pretrained on a large, loosely structured dataset to build general-purpose representations that can then be adapted to many downstream tasks with relatively little extra data. The authors argue that traditional ECG machine learning is limited by an over-reliance on task-specific labels, and that this caps how much of the ECG's diagnostic and predictive potential can be reached [s1].
Their model, ECG-CLIP, was pretrained on more than 1.7 million ECGs and paired clinician-overread report text from 542,288 patients, collected at Scripps Health between 15 January 2008 and 15 January 2019 [s1]. The pretraining had two stages: reconstructing masked portions of the ECG, and contrastive learning that ties each ECG to its accompanying text description [s1]. It was then tested on MIMIC-IV, an independent dataset of more than 800,000 ECGs [s1].
What it was tested on
The evaluation spanned three kinds of task: detecting existing cardiovascular disease (acute myocardial infarction, cardiac amyloidosis and hypertrophic cardiomyopathy), predicting future disease (atrial fibrillation from a normal sinus-rhythm ECG), and predicting adverse outcomes (30-day emergency-department and post-surgical mortality, and three-year onset of chronic kidney disease and type 2 diabetes) [s1]. Performance was measured as area under the receiver operating characteristic curve (AUC) and compared against supervised baselines, a non-ECG foundation model, and three ECG-signal-only foundation models [s1].
The headline: doing more with less labelled data
ECG-CLIP outperformed the supervised and non-ECG baselines across the detection and prediction tasks, and — the finding the authors emphasise — did so with far less labelled data [s1]. It reached the same AUC on the cardiovascular-disease detection tasks as the best comparator model while using, on average, 90.8% less training data [s1].
Its edge was clearest when labelled examples were scarce. With only ten positive labels available, ECG-CLIP significantly outperformed the runner-up model on acute myocardial infarction detection (AUC 0.910, 95% CI 0.903–0.916, versus 0.884, 0.876–0.892), cardiac amyloidosis detection (0.790 versus 0.777), hypertrophic cardiomyopathy detection (0.772 versus 0.754) and atrial fibrillation prediction (0.777 versus 0.765) [s1]. It also detected anterior and inferior myocardial infarction using single ECG leads, outperforming supervised baselines even on leads usually considered less informative [s1].
Why label efficiency is the point
The practical significance is not that ECG-CLIP is dramatically more accurate — the AUC gaps over the next-best model are modest, and the confidence intervals overlap in places. It is that it reaches good performance with very few labelled cases. Rare conditions such as cardiac amyloidosis have, by definition, few examples to train on, and the ability to learn from ten positive labels rather than thousands is what could make AI feasible for diseases where labelled data barely exists — and for hospitals that lack the resources to build large annotated datasets.
The caveats
This is a development-and-validation study, not a clinical trial. It was trained on data from one US health system and tested on one external dataset; performance can degrade on ECGs from different machines, populations and care settings, and external validation on two sources is not the same as deployment across many. The outcomes are model metrics — AUCs on retrospective data — not evidence that using the tool changes what happens to patients. And a model that predicts three-year diabetes or kidney-disease onset from an ECG raises its own questions about what a clinician is meant to do with such a prediction.
What the study establishes is a capability: a single pretrained ECG model that transfers across many diagnostic and predictive tasks and, crucially, needs far less labelled data to get there. Whether that capability improves care is the next, unanswered question.
Sources
- Development and external validation of a contrastive learning foundation model for ECG interpretation — The Lancet Digital Health, 2026-09-01
- An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm — The Lancet, 2019
Sources
- Development and external validation of a contrastive learning foundation model for ECG interpretation — The Lancet Digital Health , September 1, 2026
- An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm — The Lancet , August 1, 2019
More on
An AI ECG model found the one in ten older patients for whom AF screening paid off
A secondary analysis of the VITAL-AF trial reports a screening benefit only in the top risk decile — with a confidence interval whose lower bound sits at 0.01.
A worrying ECG pattern was rare in screened high-schoolers — but not meaningless
Across 16,893 high school athletes screened over 15 years, 0.36% had T-wave inversions. Four were later found to have hypertrophic cardiomyopathy, a leading cause of sudden cardiac death in the young.
Ablation for atrial fibrillation beat a sham procedure by 2.6 points, which is nothing
The first double-blind sham-controlled trial of pulmonary vein isolation found quality of life improved in both arms. Most of the benefit patients feel appears not to come from the ablation.
A trial finally tested blood thinners in the atrial fibrillation grey zone
Guidelines have hedged on anticoagulation for people with one stroke risk factor. SINGLE-AF randomised 1,803 patients in South Korea, and the events it counted were very few in both arms.