Electrocardiograms, or ECGs, have long been an important tool for examining the electrical activity of the heart.

Modern artificial intelligence is now opening another possibility: using the information contained in ECG recordings to identify existing cardiovascular conditions, estimate future disease risk, and recognize broader health patterns.

Researchers at Scripps Research have developed an artificial intelligence system called ECG-CLIP that is designed to work effectively even when relatively few examples have been manually labeled for a particular medical condition.

A Foundation Model for ECG Data

Many medical AI systems are developed for a particular diagnostic purpose. Training such systems generally requires large collections of ECG recordings that specialists have labeled according to whether a specific condition is present. ECG-CLIP uses a different strategy. It is a foundation model, meaning that its initial training is intended to establish a broad understanding that can later be adapted to multiple applications.

The researchers trained the model with more than 1.7 million ECG recordings from 542,288 people. Each recording was paired with clinical information written by medical professionals. Combining electrical signals with written interpretations gave the system additional context during its training.

The model was developed using two stages of pretraining. One stage taught it to reconstruct portions of ECG signals, while another connected ECG information with corresponding clinical text. This approach was designed to create representations that could be transferred to different cardiovascular tasks.

Testing Detection of Cardiovascular Conditions

Researchers evaluated ECG-CLIP using an independent dataset containing more than 800,000 ECG recordings. The first group of experiments examined whether the model could distinguish recordings associated with acute myocardial infarction, cardiac amyloidosis and hypertrophic cardiomyopathy. Across these detection tasks, ECG-CLIP performed better than the standard comparison systems included in the study.

One of the most significant findings involved the amount of labeled information required. ECG-CLIP reached performance comparable to the strongest comparison model trained with the complete labeled dataset while using approximately 90.8% less training data on average. The advantage was particularly apparent when only a small number of positive examples were available. With just 10 positive examples for certain conditions, ECG-CLIP maintained stronger performance than competing approaches.

Expert Insight

Giorgio Quer, senior author of the study and an assistant professor of digital medicine at Scripps Research, explained the potential advantage of the approach: “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.”

Beyond Detecting Existing Disease

The research also examined whether ECG-CLIP could identify information associated with future cardiovascular conditions. One experiment focused on atrial fibrillation. The model analyzed 12-lead ECG recordings showing normal heart rhythms and attempted to identify individuals who would later develop atrial fibrillation.

ECG-CLIP produced better results than the other models evaluated for this prediction task. This finding suggests that an apparently normal ECG may contain patterns that become useful when analyzed with advanced computational methods. The researchers also examined broader health outcomes. The model was tested for its ability to estimate 30-day survival following an emergency department visit or surgery, as well as the likelihood of developing chronic kidney disease or type 2 diabetes over a three-year period.

Performance With Fewer ECG Leads

Another notable aspect of the research was the model's performance with single-lead ECG information. The system performed well when researchers used individual leads to detect acute myocardial infarction. This finding could be important for situations where a complete 12-lead ECG is not available. Portable and wearable devices increasingly use simpler ECG configurations than conventional clinical equipment.

Making AI Predictions More Understandable

An important challenge with artificial intelligence in medicine is determining why a model produces a particular prediction. The researchers addressed this issue by creating saliency maps. These visualizations indicate which portions of an ECG signal contributed most strongly to a prediction.

Greater interpretability could help clinicians examine whether an AI system is focusing on medically meaningful portions of an ECG rather than relying on irrelevant characteristics. It could also make future clinical validation easier by providing researchers with a way to investigate model behavior.

The research found that ECG-CLIP could achieve strong results with substantially less labeled training data than conventional approaches. Its performance with limited examples and single-lead recordings could be particularly valuable for applications where comprehensive labeled datasets are difficult to obtain.