AHA Scientific Sessions 2026 AbstractNov 9, 2026, 01:19 GMT

Heart-trace models trained apart did slightly better on new data

Average absolute age-estimation error was 7.74 years with federated training versus 8.04 years with pooled training in external testing.

Paper by Shaaf Ahmad, Sriparna Mukherjee, Katherine Conners, +5Faisal Syed · Univ of North Carolina

Heart cross-section with branching electrical pathways beside small upright figures and a smaller gray heart with thinner walls and a wider cavity.

Our read

Biomarkers

Studied:: Computational study of 1,202,162 ECGs, against pooled training
What changed:: Average absolute age-estimation error was lower with federated training: 7.74 years versus 8.04 years with pooled training
The catch:: The federated setup was simulated; operation across institutions was not measured

Editor of record: Boris Djordjevic · AgingResearch.News

Paper

An independent cohort of 1,202,162 electrocardiograms—recordings of the heart’s electrical activity—provided the external test in a computational comparison of age-estimation models. One model used simulated federated learning: four separate data sources contributed to training without pooling recordings. The comparison model trained on pooled recordings using the same model design and preprocessing.

Although pooled training performed slightly better internally, federated training estimated age more accurately in external testing. For both models, older predicted age relative to actual age was associated with higher odds of heart failure and other diseases, after accounting for age, sex and race. The models’ ability to distinguish mortality outcomes was similar.

Why it matters

Using heart recordings to study age-related health depends on models that work beyond their training data. The authors supported federated modeling as an option when institutions cannot pool recordings.

Caveats

The federated setup was simulated, rather than tested in operation across institutions. The conference abstract did not give the external cohort’s ages or health profile, limiting assessment of applicability to particular patients.