Scripps Research Develops ECG-CLIP AI to Detect Heart Disease With Minimal Data
Researchers at Scripps Research have developed ECG-CLIP, an AI foundation model that identifies heart diseases using significantly less labeled training data than previous tools. According to a study published in Lancet Digital Health on September 1, 2026, the model can detect specific cardiovascular conditions after seeing as few as a dozen confirmed examples, mimicking the way clinicians learn from general physiology rather than massive datasets.
How ECG-CLIP Reduces Reliance on Labeled Data
Traditional AI tools for electrocardiogram (ECG) analysis require vast amounts of “hand-labeled” data—records where clinicians have manually marked the presence or absence of a disease. This requirement often makes AI tools rigid and difficult to adapt to new clinical tasks. ECG-CLIP functions as a “foundation model,” meaning it learns from diverse datasets before being applied to specific tasks.
The model was trained using more than 1.7 million ECGs from over 540,000 people. Unlike previous models, ECG-CLIP paired these ECGs with clinicians’ notes. Senior author Giorgio Quer, an assistant professor of digital medicine at Scripps Research, stated that this approach allows the algorithm to detect a disease in the future after seeing only about a dozen confirmed cases.
Performance in Disease Detection and Prediction
The Scripps Research team tested ECG-CLIP against standard deep learning models, linear models, and other ECG-trained foundation models across three primary clinical tasks:
- Disease Detection: The team tested the model’s ability to identify acute myocardial infarction, cardiac amyloidosis, and hypertrophic cardiomyopathy. Using the “area under the curve” (AUC) method to measure accuracy, ECG-CLIP consistently outperformed standard models.
- Data Efficiency: Across the three detection tasks, ECG-CLIP matched the performance of the next-best model trained on a full dataset while using approximately 91% less hand-labeled training data on average, according to the researchers.
- Future Prediction: The model outperformed all other tested models in predicting future atrial fibrillation—a type of irregular heart rhythm—from 12-lead ECGs that appeared normal.
Clinical Applications for Rare Diseases and Limited Resources
The ability to function with minimal data provides a specific advantage for diagnosing rare cardiovascular diseases. Because few labeled examples exist for rare conditions, standard AI often fails; however, ECG-CLIP’s architecture allows it to maintain accuracy even with only 10 positive examples of a given disease, according to the study.
The researchers also found that ECG-CLIP performed well using single-lead ECG data when detecting acute myocardial infarction. This suggests the tool could be effective in resource-limited settings where a full 12-lead ECG machine is unavailable.
Comparison of AI Training Requirements
| Model Type | Training Data Requirement | Primary Learning Source |
|---|---|---|
| Standard Deep Learning | High (Thousands of labeled examples) | Hand-labeled ECGs |
| ECG-CLIP | Low (As few as 12 labeled examples) | ECGs paired with clinician notes |