This work investigates whether an adult electrocardiogram (ECG) foundation model can be transferred to pediatric ECGs. Using 339,563 12‑lead ECGs from 238,189 children, we compared three strategies under an identical architecture: direct transfer, fine‑tuning on pediatric data, and training from scratch. Direct transfer yielded a mean AUC of 0.8256, indicating poor performance. Fine‑tuning raised the AUC to 0.9695, surpassing de‑novo training (0.9661), and achieved 0.9079 with only 1% of the training data. On an external set of 10,999 ECGs, fine‑tuned model reached a mean AUC of 0.9385 across 30 labels. The model also generalized to age regression (MAE 1.12 years) and congenital heart disease classification (AUC 0.9562). In a reader study with 17 clinicians evaluating 100 cases, diagnostic accuracy improved from 56.8% to 65.4% (p < 0.05).
Review: The study demonstrates that adult ECG foundation models can be efficiently adapted to pediatric settings with minimal data, offering a practical pathway for pediatric ECG AI deployment.