Heart tracing model improves predictions with limited data
With 100 training examples, a model trained mostly on children’s records scored 0.80 versus below 0.65 for competitors at distinguishing weak heart pumping.
Limit: Evaluation compared predictions against heart imaging; whether model-guided care improved health was not measured
Edited by Boris Djordjevic · AgingResearch.News
Researchers trained a computer model on 782,138 electrocardiograms—recordings of the heart’s electrical activity—from 255,271 patients with a median age of 10.9 years. The computational study used mostly pediatric records from Boston Children’s Hospital. Testing used separate patients whose heart scans provided comparison measurements.
With 100 training examples, the model scored 0.80 versus below 0.65 for competing models at detecting a left ventricle—the heart’s main pumping chamber—that expelled 40% or less of its blood per beat. These scores measured how well models distinguished people with and without this reduced pumping function. External validation used 82,543 paired heart tracings and ultrasound scans from adults with a median age of 64.0 years. There, the model scored 0.83.
Why it matters
Heart screening across the lifespan raises the question of whether models trained mostly in children can remain useful in older adults. The authors suggest their approach could support screening when labeled training data or imaging access are limited.
Caveats
This computational evaluation tested predictions against heart imaging, not whether model-guided care improved health. The work is a preprint and has not been peer-reviewed.
The paper
An ECG foundation model for generalizable cardiac function prediction across the lifespan
Boston Children's Hospital
medRxiv, May 27, 2026, Open access, Preprint, not peer-reviewed
Comments
Be the first to comment.