2 September 2026 (Navroze Bureau) : Scientists at Scripps Research have developed a new artificial intelligence model called ECG-CLIP that can detect and predict several heart conditions while requiring significantly less manually labelled data than conventional AI systems.
The model was trained on more than 1.7 million electrocardiograms (ECGs) from over 540,000 people, with the ECGs paired with clinicians’ notes. Researchers say this multimodal training helps the model adapt to new cardiovascular tasks, particularly when only a small number of labelled examples are available.
AI Model Targets Multiple Heart Conditions
Researchers tested ECG-CLIP on three conditions: acute myocardial infarction, cardiac amyloidosis and hypertrophic cardiomyopathy.
Using a dataset containing more than 800,000 ECGs, the model consistently outperformed standard deep-learning and linear models in detecting all three diseases.
About 91% Less Labelled Data Needed
One of the most notable findings was the amount of labelled data required.
Across the three disease-detection tasks, ECG-CLIP achieved performance comparable to the next-best model trained using the full dataset while using approximately 91% less hand-labelled training data on average.
The advantage was particularly pronounced when only a handful of confirmed cases were available. Researchers reported that ECG-CLIP could perform effectively with as few as 10 positive examples for a disease.
Potential Advantage for Rare Diseases
The researchers say the technology could be especially useful for rare heart conditions, where collecting thousands of accurately labelled ECGs can be difficult.
Instead of requiring a separate large labelled dataset for every disease, a foundation model can first learn broader relationships between ECG patterns and clinical information and then adapt to specific tasks using relatively few examples.
Clinicians’ Notes Helped the Model
ECG-CLIP differs from some existing ECG foundation models because it was trained using both ECG signals and clinicians’ notes.
The model uses contrastive multimodal learning to connect electrical patterns in ECG recordings with clinical descriptions, allowing it to develop a broader representation that can be applied to different cardiovascular tasks.
Can Work With a Single ECG Lead
The model also showed promising performance when using single-lead ECG data to detect acute myocardial infarction.
This could eventually be useful in settings where conventional 12-lead ECG equipment is unavailable or difficult to deploy, although additional validation would be needed before clinical use.
ECG-CLIP Can Predict Future Heart Disease
The researchers did not limit their testing to detecting existing disease.
They also evaluated whether ECG-CLIP could predict future atrial fibrillation, an irregular heart rhythm, in people whose 12-lead ECGs initially showed normal rhythms.
The model outperformed the other systems tested for this prediction task.
AI Also Predicted Other Health Outcomes
ECG-CLIP performed strongly in predicting several adverse health outcomes.
Researchers found it performed best among the tested models at predicting 30-day survival following an emergency-department visit or surgery. It also showed the strongest performance in predicting the development of chronic kidney disease and type 2 diabetes within three years.
Researchers Added Saliency Maps
A major challenge with medical AI is understanding why a model reaches a particular conclusion.
To improve interpretability, the researchers generated saliency maps that highlight parts of an ECG signal that contributed most to the model’s prediction. Such visualisation could help clinicians examine what the AI is focusing on rather than treating its output as a black box.
More Testing Needed Before Clinical Use
Despite the promising results, ECG-CLIP is not yet a replacement for doctors or established diagnostic procedures.
The researchers emphasised that prospective clinical trials and further validation are needed to determine how well the model performs in real-world healthcare settings.
Could Expand Access to Cardiac Screening
The findings point to a potential future in which ECG-based AI tools can be adapted to different diseases without requiring enormous disease-specific labelled datasets.
That could be particularly important in healthcare systems where specialist cardiologists, high-quality labelled data or advanced ECG equipment are limited.
The Scripps team also plans to test the model across different ECG recording systems, including wearable devices, which could eventually support remote and continuous monitoring of cardiovascular health.
A New Direction for Medical AI
ECG-CLIP represents a shift from narrowly trained diagnostic algorithms towards broader foundation models capable of handling multiple clinical tasks.
Its ability to work with much less labelled data could make development of AI tools for rare diseases and under-resourced clinical environments more practical, provided future studies confirm its accuracy and safety.