AI-enabled ECG: from a screening alert to clinical confirmation

Mayo Clinic Minute: How AI can signal future heart issues

Mayo Clinic · 2019-10-03 · Source

This short English-language Mayo Clinic report from 2019 explains research combining a digital stethoscope ECG with AI. It is a historical example, not coverage of the 2025 clearance for the 12-lead device or the 2026 validation study discussed here.

Research: what changed in the clinical workflow?

Mayo Clinic’s account of the 2021 EAGLE randomized study, involving 22,641 adults, reports more diagnoses of low ejection fraction with AI-enabled ECG screening. The institution describes approximately five additional diagnoses per 1,000 people screened compared with usual care. A positive alert prompted consideration of echocardiographic assessment. This finding does not establish a definitive diagnosis by AI alone or demonstrate reduced mortality.

External validation: high AUROC does not make every alert correct

A four-site retrospective evaluation published in 2026 analyzed 13,960 records. AUROC was 0.92, while positive predictive value was 30.5%; low ejection fraction prevalence was 7.9%. Discrimination and the probability that a positive result identifies the condition are therefore different measures. These figures describe this study population. Anumana funded the study, and employment, consulting and licensing relationships were disclosed. Validation on external data does not mean financial independence.

Regulatory scope: a specific product and intended use

FDA document K250652, dated July 28, 2025, describes Anumana’s 12-lead ECG-AI algorithm as an aid to early detection of left ventricular ejection fraction at or below 40% in adults at risk for heart failure. It is not a standalone diagnostic or patient-monitoring tool; paced-rhythm ECGs are excluded. A positive result suggests further clinical evaluation. This US clearance is not blanket authorization for AI, authorization in other countries or approval of a RAI product.

Clinical use: do not treat different evidence as one result

These sources address three different questions: a workflow’s contribution to identifying diagnoses, model performance in a particular dataset and the boundaries of a product’s cleared use. Compressing them into one “accuracy” percentage would be misleading. Responsible AI assessment should keep research findings, product documentation and patient-specific clinical decisions distinct. This distinction is our editorial interpretation; the article does not provide individual diagnosis or treatment advice.

Editorial proposal: measure what follows the alert

When evaluating an implementation, we recommend making the model version, accepted ECG formats, intended population and review pathway visible. Track unnecessary investigations, records that cannot be evaluated and cases lost to follow-up alongside confirmed alerts. Responsibility for human review should be explicit. These are editorial suggestions for responsible implementation, not measured performance results or clinical-effectiveness claims for RAI products.