According to research published in Lancet Digital Health on September 1, 2026, scientists at Scripps Research have developed a new artificial intelligence tool called ECG-CLIP to detect and predict heart diseases using less hand-labeled data than traditional models. Built using over 1.7 million electrocardiograms collected from more than 540,000 people and paired with clinicians’ notes, the model matches the performance of existing algorithms while using about 91% less training data.
How ECG-CLIP Outperforms Traditional Medical AI Models
Clinicians routinely use a 12-lead electrocardiogram (ECG) to record the heart’s electrical activity using chest and limb electrodes. While AI tools regularly assist with diagnoses, current systems rely heavily on large amounts of hand-labeled training data specifying the presence or absence of individual diseases. This reliance limits their adaptability across diverse clinical workflows, according to Scripps Research.
To overcome this bottleneck, the Scripps team designed ECG-CLIP as a foundation model capable of learning from diverse datasets before performing specific tasks. When tested on detecting acute myocardial infarction, cardiac amyloidosis, and hypertrophic cardiomyopathy across a new dataset of over 800,000 ECGs, ECG-CLIP consistently outperformed standard deep learning and linear baseline models using an area under the curve (AUC) evaluation method.
“Our new algorithm only needs to see on the order of a dozen confirmed ECGs of a specific disease to detect that disease in the future,” says Giorgio Quer, senior author and assistant professor of digital medicine at Scripps Research, as reported by Mirage News. Quer adds that the system learns similarly to a human clinician by first understanding general physiology.
Tackling Rare Diseases and Limited Resource Settings
The model’s efficiency proves especially valuable in environments with sparse data or limited hardware. When evaluated against three other ECG-trained foundation models lacking clinician notes, ECG-CLIP excelled in settings containing as few as 10 positive examples of a given condition. These performance gaps narrowed as labeled examples increased, but the advantage remains vital for rare diseases.
“This may be particularly useful in cases like rare diseases, where there are only a dozen or so positive examples of well-labeled ECGs that can be used for training the model,” Quer notes. Furthermore, tests demonstrated that ECG-CLIP performs well using single-lead ECG data for acute myocardial infarction detection, offering a practical advantage for resource-limited facilities.
Did you know? ECG-CLIP was trained on over 1.7 million ECGs paired directly with physician notes, allowing the algorithm to absorb complex clinical context rather than relying solely on raw electrical signals.
Predicting Atrial Fibrillation and Adverse Health Outcomes
Beyond immediate disease detection, researchers tested the algorithm’s predictive capabilities. Members of Quer’s lab evaluated the models on predicting future atrial fibrillation—an irregular heart rhythm—using 12-lead ECGs that displayed entirely normal heart rhythms. ECG-CLIP outperformed all competing models in this task.
The tool also demonstrated superior performance in forecasting broader adverse health outcomes. According to the study data, ECG-CLIP best predicted survival likelihood within 30 days of an emergency department visit or surgery, alongside the future development of chronic kidney disease and type II diabetes within a three-year window.
A persistent challenge for artificial intelligence in medicine is the “black box” problem, where the reasoning behind specific algorithmic outputs remains unclear. To address this, the Scripps team utilized saliency maps to visualize decision-making processes.
These maps highlight the exact regions within an ECG signal that contribute most heavily to the model’s predictions. By providing a clear view of the underlying signal features, the visualization technique aims to build trust with practicing cardiologists and facilitate eventual clinical deployment.
Next Steps for Real-World Validation
Future development will focus on expanding the variety of data available to the model to enhance performance in fast-paced environments like emergency departments. Researchers also plan to evaluate compatibility with alternative recording systems, including wearable devices designed for continuous remote monitoring.
Despite the promising results, broader adoption requires careful verification. “While these findings demonstrate substantial potential, rigorous validation in prospective clinical trials will be required to establish ECG-CLIP’s applicability in real-world clinical settings,” says co-first author Michael Ko, a graduate student and research assistant at Scripps Research.
This project builds on previous innovations from Quer’s lab, including a 2024 algorithm that diagnoses heart attacks using only three leads and a 2023 algorithm capable of identifying atrial fibrillation risks from two-week wearable patches.
Frequently Asked Questions
What is ECG-CLIP?
ECG-CLIP is an artificial intelligence foundation model developed by Scripps Research that pairs 1.7 million electrocardiograms with clinician notes to detect and predict heart diseases using minimal training data.
How much training data does ECG-CLIP require?
According to Scripps Research, the algorithm can detect specific diseases after seeing only about a dozen confirmed, positive ECG examples, using roughly 91% less hand-labeled data than traditional models.
Can ECG-CLIP work with single-lead ECG devices?
Yes. Testing showed the model performs well using single-lead ECG data for acute myocardial infarction detection, making it useful for resource-limited settings.
What are saliency maps in this context?
Saliency maps are visual overlays that highlight specific regions of an ECG signal contributing to the AI model’s predictions, helping clinicians understand and trust the output.
Keep reading