ECG-CLIP AI Detects and Predicts Heart Disease With Few Labeled Examples

Researchers at Scripps Research have developed ECG-CLIP, a novel artificial intelligence foundation model designed to enhance the detection and prediction of cardiovascular diseases using electrocardiogram data. Published in Lancet Digital Health, the algorithm was trained on a dataset comprising 1.7 million ECGs paired with clinician notes from over 540,000 individuals. By integrating textual medical records with physiological signals, the model learns underlying cardiac principles rather than relying exclusively on vast quantities of hand-labeled examples. According to senior author Giorgio Quer, the system can identify specific cardiac conditions after exposure to merely a dozen confirmed cases, mirroring clinical learning patterns and addressing a major bottleneck in healthcare AI where labeled data is scarce. In comparative testing against standard deep learning architectures, linear models, and existing ECG-trained foundation models, ECG-CLIP demonstrated superior performance across three key clinical tasks: disease detection, future condition forecasting, and adverse outcome prediction. When screening for acute myocardial infarction, cardiac amyloidosis, and hypertrophic cardiomyopathy across a dataset of 800,000 ECGs, the model consistently outperformed baselines while utilizing approximately 91 percent less annotated training data. Its efficiency is particularly advantageous for identifying rare conditions, where it maintained high accuracy even with fewer than ten positive examples per disease. The algorithm also proved effective when processing single-lead ECG inputs, enhancing its utility in resource-constrained environments. Beyond diagnostic capabilities, ECG-CLIP excelled at predictive analytics. It accurately forecasted future atrial fibrillation episodes from normal sinus rhythm ECGs and outperformed competing systems in estimating 30-day survival rates following emergency department visits or surgical procedures. Additionally, the model successfully predicted the three-year onset of chronic kidney disease and type 2 diabetes. To address the transparency gap common in machine learning applications, the research team implemented saliency mapping, a visualization technique that highlights the specific signal segments driving each prediction, thereby fostering clinical trust and facilitating potential integration into electronic health record systems. While the algorithm represents a significant advancement in computational cardiology, researchers emphasize that prospective clinical trials remain essential to validate its real-world applicability and compatibility with diverse recording hardware, including wearable monitoring devices. This development follows a series of diagnostic tools from the same laboratory, reinforcing a growing portfolio of AI solutions aimed at augmenting cardiologist workflows and expanding accessible cardiac care.