Researchers at Scripps Research have developed a new foundational AI model called ECG-CLIP, which improves the detection and prediction of various heart diseases while using about 91 percent less hand-labeled training data. Published in Lancet Digital Health on September 1, 2026, the algorithm requires only about a dozen confirmed examples to detect specific cardiovascular conditions.
How ECG-CLIP Works With Minimal Labeled Data
Clinicians routinely rely on a 12-lead electrocardiogram, or ECG, to record the heart’s electrical activity using chest and limb electrodes. While artificial intelligence tools frequently assist with reading these tests, traditional models demand vast amounts of hand-labeled training data specifying the presence or absence of individual diseases. To overcome this limitation, researchers at Scripps Research built ECG-CLIP by training a model on more than 1.7 million ECGs collected from over 540,000 individuals and paired with clinicians’ notes.
Senior author Giorgio Quer, an assistant professor of digital medicine at Scripps Research, explained the model’s design approach by noting that 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.
Quer added that the process mirrors human clinical learning not from a million examples, but from understanding the general physiology behind an ECG first and then seeing a few specific cases.
“We found that ECG-CLIP was better at detecting and predicting cardiovascular diseases, particularly in cases where there was much less data, 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.”
Giorgio Quer, senior author, assistant professor of digital medicine, Scripps Research
Detecting and Predicting Heart Conditions Across Multiple Clinical Tasks
Once fully trained, the research team tested ECG-CLIP across three distinct clinical categories against supervised baseline models, a general foundation model, and three alternative ECG-trained foundation models. In the first task, the models evaluated a new dataset of over 800,000 ECGs for acute myocardial infarction, cardiac amyloidosis, and hypertrophic cardiomyopathy. Using an area under the curve (AUC)
method to measure how effectively each model distinguished patients with and without a condition, the team found that ECG-CLIP consistently outperformed standard models while utilizing roughly 91 percent less hand-labeled training data on average.
Beyond immediate detection, the algorithm demonstrated predictive capabilities. When focused on 12-lead ECGs displaying normal heart rhythms, the model outperformed all competing systems at predicting future atrial fibrillation. Furthermore, it excelled at forecasting adverse health outcomes, including 30-day survival likelihood following an emergency department visit or surgery, alongside predicting the development of chronic kidney disease and type II diabetes within three years.
Clinical Interpretability and Emerging AI Screening Tools
To build trust among medical professionals and ensure transparency in hospital environments, the Scripps team incorporated saliency maps into the algorithm. These visual overlays highlight specific regions within the electrical signal that heavily influence the model’s predictions, providing clinicians with a clearer view of the underlying data features.
Concurrently, advances in rapid AI-driven cardiac screening were presented to thousands of delegates at the European Society of Cardiology annual congress in Munich. Reporting on a trial involving 67,000 patients in the United States, researchers detailed a separate AI tool capable of delivering an ECG read-out in less than two seconds. Dr. Sonya Babu-Narayan, clinical director of the British Heart Foundation, noted that technology like the AI ECG in this research, which has the potential to identify high-risk patients early, will not detect everyone with a heart condition. But it could be a solution to help fast-track the patients who are most likely to have a heart abnormality.
That trial demonstrated that the tool could identify up to 81 percent of individuals with heart failure and up to 90 percent of those with heart valve disease. While the technology cannot independently provide a definitive diagnosis, experts emphasize its utility in flagging high-risk individuals. Prof. Fu Siong Ng, a professor of cardiology at Imperial College London, pointed out that patients often wait several months for a scheduled heart ultrasound scan, stating that our technology could identify patients most at risk of heart failure and heart valve disease, so they could be prioritised for scans faster and more urgently.
По теме
