ECG-based AI algorithm could improve detection and prediction of heart diseases

Clinicians frequently employ a medical examination known as a 12-lead electrocardiogram (ECG) to diagnose heart conditions. This test involves placing electrodes on the chest and limbs to record the heart's electrical activity. AI tools are often employed to aid diagnoses based on ECG readings. However, these tools typically necessitate substantial quantities of labeled training data, which can hinder their adaptability to other clinical applications.

Researchers at Scripps Research have now developed an innovative AI model named ECG-CLIP, which aims to enhance the detection and prediction of diverse heart diseases. This model, described in Lancet Digital Health, was trained using over 1.7 million ECGs from more than 540,000 individuals, supplemented with clinicians' notes, thereby imbuing the tool with knowledge that could potentially make it more versatile for various disease detection and prediction tasks in real-world clinical environments.

Senior author Giorgio Quer, an assistant professor of digital medicine at Scripps Research, explains that ECG-CLIP only requires observing around a dozen confirmed ECGs of a specific disease to detect that disease in the future, akin to how a clinician would learn. This characteristic makes ECG-CLIP particularly advantageous for situations with limited data, such as rare diseases.

Additionally, ECG-CLIP outperformed existing models in detecting three distinct heart diseases—acute myocardial infarction, cardiac amyloidosis, and hypertrophic cardiomyopathy—in a new dataset containing more than 800,000 ECGs. Moreover, it matched the performance of the top-performing model trained on the entire dataset while utilizing approximately 91% less labeled training data on average.

ECG-CLIP demonstrated superior performance in limited data scenarios as well. In tests where only 10 positive examples of a given disease existed, the model's performance surpassed that of three ECG foundation models, which were trained solely on ECG data without considering clinicians' notes. In terms of predicting heart diseases, ECG-CLIP excelled in forecasting atrial fibrillation from normal 12-lead ECGs and outperformed all other models in predicting the likelihood of 30-day survival following emergency department visits or surgery, as well as the likelihood of chronic disease development within three years, including chronic kidney disease and type 2 diabetes.

To enhance the interpretability of the model's predictions, the researchers generated saliency maps, highlighting the specific regions of the ECG signal most influential in the model's decision-making process. This feature can help build trust with clinicians and facilitate the integration of ECG-CLIP into clinical practice.