Scripps Research scientists have built an artificial-intelligence model that connects electrocardiogram signals with the language clinicians used to interpret them, then adapts that shared representation to new prediction tasks. In retrospective testing, ECG-CLIP often needed far fewer hand-labeled examples than comparison models, the La Jolla institute reported Monday.
An electrocardiogram, or ECG, records the heart's electrical activity through electrodes placed on the body. A standard clinical test uses 12 leads, each providing a different view. Machine-learning systems can classify those signals, but conventional tools typically require large sets labeled for each particular diagnosis or outcome. Rare conditions and smaller health systems may not have enough confirmed examples to build such a set.
ECG-CLIP is a foundation model, meaning it first learns a general representation that can be adapted to several downstream tasks. The team trained it on more than 1.7 million ECGs from 542,288 patients in the Scripps Health GE MUSE system, paired with clinician-overread reports. The retrospective records were collected in clinics from January 2008 through January 2019, according to the study abstract.
Researchers then evaluated disease detection, future-risk prediction and adverse-outcome prediction. For detection, they used a separate dataset of more than 800,000 ECGs and tested acute myocardial infarction, cardiac amyloidosis and hypertrophic cardiomyopathy. Performance was measured with area under the receiver operating characteristic curve, which summarizes how well a model separates cases from non-cases across decision thresholds.
Across those three detection tasks, Scripps said ECG-CLIP matched the next-best model trained on the full data while using about 91% less labeled training data on average. Its advantage over other ECG foundation models was clearest when only about 10 positive examples were available. As the number of labeled examples grew, the performance differences generally disappeared.
Other tests examined whether a normal-rhythm 12-lead ECG could help predict future atrial fibrillation and whether the model could estimate 30-day survival after an emergency-department visit or surgery. The team also evaluated three-year development of chronic kidney disease and type 2 diabetes. ECG-CLIP performed best among the compared models on those reported tasks and also worked with single-lead data in a heart-attack detection test.
Those findings do not mean the system is ready to diagnose a patient. The work was retrospective, using existing records rather than decisions made in real time. Co-first author Michael Ko said prospective clinical trials are required to establish whether the model applies reliably in practice. Performance can also change across hospitals, ECG machines, patient populations and workflows that differ from the data used for development and validation.
The researchers added saliency maps, visual displays that identify regions of an ECG signal contributing strongly to a prediction. Such maps can help a clinician inspect what the model attended to, but they do not by themselves prove the model used a medically valid causal feature. Interpretability tools are an aid to review, not a replacement for external validation or clinical judgment.
The paper, led by Ko and Matteo Gadaleta with senior author Giorgio Quer, appeared Sept. 1 in The Lancet Digital Health. Funding came from the VoLo Foundation and the National Center for Advancing Translational Sciences. The immediate result is a promising efficiency finding under controlled retrospective tests; the next question is whether that advantage survives prospective use without introducing new errors or inequities.