News
ECG-CLIP AI Model Detects Heart Disease From a Dozen ECG Examples
Scientists at Scripps Research have developed an AI model that identifies heart disease from as few as a dozen confirmed ECG recordings, using on average 91 percent less labeled data than existing systems.
A team at the Scripps Research Institute in California has published an AI model that learns to identify heart disease from as few as a dozen confirmed electrocardiogram recordings of a given condition. It's a radical departure from the previous approach, in which diagnostic AI models needed thousands of labeled examples to reach useful accuracy.
How the model works
ECG-CLIP is a so-called foundation model based on contrastive learning, similar to the technique used in models that combine image and text. During pretraining, the system paired ECG recordings with corresponding clinical notes written by doctors, learning general relationships between the heart's electrical signal and a patient's health status before it even encountered specific disease categories.
That meant that when researchers wanted to teach the model to recognize a specific, rarer disease, they didn't need to supply thousands of examples. Showing the system a dozen or so confirmed cases was enough for the model to generalize that knowledge to new, previously unseen ECG recordings.
Our algorithm needs only about a dozen confirmed ECGs of a disease to detect it - Giorgio Quer, Scripps Research
It's similar to how a clinician learns, not from a million examples but from an understanding of general physiology - Giorgio Quer, senior author of the study
What it can detect
The researchers tested ECG-CLIP on three types of tasks. The first was detecting diseases already present in a patient, including myocardial infarction, cardiac amyloidosis and hypertrophic cardiomyopathy. In each of these tasks, the model, measured by AUC, achieved results better than or comparable to systems trained on far larger amounts of labeled data.
The second area was predicting future events based on a normal, seemingly healthy ECG, including the risk of atrial fibrillation, survival over the following 30 days, and the development of type 2 diabetes and chronic kidney disease over a three-year horizon. The third test was the model's performance when analyzing data from a single ECG lead instead of the standard twelve, which matters for wearable devices and facilities with limited access to equipment.
Why less data is a breakthrough
The biggest barrier to building AI systems for medical diagnostics isn't a lack of computing power, it's a lack of large, precisely labeled datasets. Labeling ECG recordings requires the time of qualified cardiologists, and for rare diseases, gathering thousands of confirmed cases is often practically impossible.
Cutting the need for training data by 91 percent means similar models could be built for rare diseases that until now lacked datasets large enough to train classic neural networks. That opens the door to faster deployment of diagnostic tools in smaller hospitals and research centers that don't have databases of millions of patients.
What's next
Giorgio Quer's team has announced prospective clinical trials to verify ECG-CLIP's effectiveness under real-world medical practice conditions rather than just on historical data. The researchers also plan to test the model's compatibility with wearable devices such as smartwatches and portable ECG monitors, for continuous, remote patient monitoring.
For Polish medical facilities that have spent months testing various AI tools supporting cardiology, a model like this could mean cheaper deployments. Lower demand for training data reduces the cost of building local diagnostic systems tailored to a given population, without needing to acquire millions of ECG recordings.
