LA JOLLA, CA—Clinicians often use a medical test called a 12-lead electrocardiogram (ECG) to diagnose heart problems, which uses electrodes placed on the chest and limbs to record the heart's electrical activity. Artificial intelligence (AI) tools are commonly used to assist with diagnoses based on ECGs. But current tools typically require large amounts of training data, hand-labeled with the presence or absence of specific diseases, which can make the tools not very adaptable to other clinical tasks.
Now, scientists at Scripps Research have built a new AI model to improve detection and prediction of various heart diseases, which may be particularly useful in settings with limited labeled data or fewer available ECG leads. Published in Lancet Digital Health on September 1, 2026, the new tool, called ECG-CLIP, was trained using over 1.7 million ECGs collected from more than 540,000 people and paired with clinicians' notes, embedding the tool with knowledge that may make it more adaptable for different disease detection and prediction tasks in real-world clinical environments.
"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 senior author Giorgio Quer , an assistant professor of digital medicine at Scripps Research. "This is similar to how a clinician would learn: not from a million examples, but from understanding the general physiology behind an ECG first and then seeing a few specific cases."
ECG-CLIP is a "foundation model," which is a type of AI model that learns from diverse datasets, after which it can be used to perform many different tasks. Once ECG-CLIP was fully trained, Quer's team set out to test how it performed three clinical tasks compared with existing models. These included two supervised baseline models (a standard deep learning model and a linear model), a general foundation model (not trained on ECGs), and three ECG-trained foundation models.
The first task the team tested was detecting heart diseases. They selected three types to look for: acute myocardial infarction, cardiac amyloidosis and hypertrophic cardiomyopathy. The researchers then gave each model a new dataset containing over 800,000 ECGs and asked which showed signs of the three diseases. Using an "area under the curve (AUC)" method—measuring how well a model distinguishes between ECGs from people with and without a particular condition—Quer's team found that ECG-CLIP consistently performed better than the standard models at detecting all three diseases.
The team also found that across the three disease-detection tasks, ECG-CLIP matched the performance of the next-best model trained on the full dataset while using about 91% less hand-labeled training data on average.
Compared with the three ECG foundation models—which were trained on ECGs, but not clinician notes—ECG-CLIP also performed better in settings where the amount of labeled training data was limited (as few as 10 positive examples of a given disease). The differences in performance between ECG-CLIP and the other models generally disappeared as the number of labeled examples increased.
"We found that ECG-CLIP was better at detecting and predicting cardiovascular diseases, particularly in cases where there was much less data," says Quer. "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."
ECG-CLIP also performed well using single-lead ECG data in tests of acute myocardial infarction detection, which may make it helpful in settings where resources are limited.
The second task tested was predicting heart disease. Here, members of Quer's lab focused on the models' ability to predict atrial fibrillation—a type of irregular heart rhythm. They found that ECG-CLIP outperformed all other models at predicting future atrial fibrillation from 12-lead ECGs displaying normal heart rhythms.
The final task was predicting adverse health outcomes. ECG-CLIP demonstrated the best performance of the tested models at predicting the likelihood of survival in 30 days following an emergency department visit or surgery, and the likelihood of chronic disease development—specifically chronic kidney disease and type II diabetes—within three years.
A common problem with AI tools is that it can be unclear what specific features of the data the model uses to make decisions. To help visualize this, the researchers generated saliency maps. These maps highlight regions within the signal that are contributing most to the model's predictions, which gives clinicians a clearer view into ECG regions that the model relied on. Quer's team implemented saliency maps to help make ECG-CLIP's output more interpretable, which can help build trust with clinicians and increase the likelihood of deployment in the clinic.
Going forward, Quer's team aims to expand the type of data available to the model to improve its performance in specific settings, like the emergency department. The team also hopes to evaluate its compatibility with different types of ECG recording systems, like wearable devices—where integration could eventually allow continuous, remote heart disease monitoring.
"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 the study's co-first author Michael Ko, a graduate student and research assistant at Scripps Research.
The work adds to an expanding collection of AI algorithms developed by Quer's lab to screen for and diagnose heart conditions. This includes an algorithm developed in 2024 to diagnose heart attacks and rhythm disturbances just as accurately when only three out of 12 ECG leads are available; and another algorithm from 2023 that can identify patients most at risk for atrial fibrillation from an ECG patch worn for two weeks.
"For me personally, it's exciting that we can develop algorithms with the potential to really help clinicians and cardiologists," says Quer.
In addition to Quer and Ko, authors of the study, " Development and external validation of a contrastive learning foundation model for ECG-based prediction of cardiovascular diseases and outcomes ," include Matteo Gadaleta, Eric Topol and Evan Muse of Scripps Research.
This work was supported by funding from the VoLo Foundation, and the National Center for Advancing Translational Sciences (grant UM1TR004407).
