ECG foundation model predicts heart disease with less data

Researchers at Scripps Research have developed an AI foundation model that can detect and predict several cardiovascular conditions using electrocardiogram (ECG) data. The model, called ECG-CLIP, requires substantially fewer labelled examples than conventional AI approaches and can also work with single-lead ECG data. This could make the technology particularly relevant for rare diseases and settings where clinical data or diagnostic resources are limited.

The research used more than 1.7 million ECGs from over 540,000 people. Unlike many existing models, ECG-CLIP was trained by pairing ECG recordings with clinicians’ notes. This allows the model to learn broader relationships between electrical heart activity and clinical information rather than relying solely on large datasets manually labelled for individual diseases.

Learning from ECGs and clinical notes

ECG-CLIP is designed as a foundation model, meaning that knowledge acquired during initial training can subsequently be applied to different clinical tasks. According to senior author Giorgio Quer, assistant professor of digital medicine at Scripps Research, the model can learn to identify a specific disease from roughly a dozen confirmed ECG examples.

Researchers compared ECG-CLIP with conventional supervised AI models, a general foundation model and three foundation models specifically trained on ECG data. For disease detection, they tested the models on a separate dataset containing more than 800,000 ECGs. The task was to identify acute myocardial infarction, cardiac amyloidosis and hypertrophic cardiomyopathy.

ECG-CLIP consistently outperformed the conventional models across all three conditions. On average, it matched the performance of the next-best model trained on the complete dataset while requiring approximately 91 percent less manually labelled training data.

The advantage was particularly evident when only a small number of positive examples were available. With as few as ten labelled cases of a disease, ECG-CLIP performed better than the other ECG foundation models. As more labelled examples became available, however, the performance differences generally disappeared.

Single-lead ECG-data

The researchers also examined whether ECG-CLIP could extract useful information from more limited ECG recordings. When detecting acute myocardial infarction, the model performed well using single-lead ECG data. This could potentially broaden its applicability in environments where full 12-lead ECG systems are unavailable.

Beyond diagnosis, ECG-CLIP was tested for its ability to forecast future disease. The researchers examined whether the model could predict atrial fibrillation from 12-lead ECGs that showed a normal heart rhythm at the time of recording. ECG-CLIP outperformed all other models included in this comparison.

A third set of tests focused on broader health outcomes. The model achieved the strongest performance among the tested systems in predicting 30-day survival following an emergency department visit or surgery. It also performed best in estimating the likelihood of developing chronic kidney disease or type 2 diabetes within three years.

To provide greater insight into how the AI reaches its conclusions, the researchers incorporated saliency maps. These visualisations highlight the regions of an ECG signal that contribute most strongly to a prediction. Such information could help clinicians understand which parts of the recording influenced the model’s assessment.

Clinical validation

Despite the results, ECG-CLIP is not yet ready for routine clinical use. The researchers stress that prospective clinical trials will be necessary to determine whether its performance translates into real-world patient care. The team plans to expand the range of data used by the model, including information relevant to specific environments such as emergency departments. Researchers also want to investigate whether ECG-CLIP can work with different ECG recording technologies, including wearable devices. In the longer term, such integration could potentially support continuous remote cardiovascular monitoring.

The researchers see particular potential in situations where high-quality labelled datasets are scarce. For rare cardiovascular diseases, where only a limited number of confirmed ECGs may be available for AI training, a model capable of learning from a small number of examples could make advanced ECG analysis more broadly applicable.

AI-ECG

Last year, an AI-model, developed at Mayo Clinic, enabled routine electrocardiograms (ECGs) to be used to detect advanced chronic liver disease at an early stage. By analysing subtle changes in cardiac electrical signals, the system identified approximately twice as many cases compared to standard clinical practice, often in patients without symptoms.

The approach leverages the physiological link between liver dysfunction and cardiovascular changes, which are difficult for clinicians to detect but recognizable through AI. Trained on data from over 11,000 patients, the model was validated using imaging and blood tests. A clinical trial involving 248 clinicians showed the technology can be integrated into routine care, enabling earlier intervention and potentially improving patient outcomes.