AI Improves Pediatric ECG Diagnosis with Adult Data Transfer

Interpreting electrocardiograms (ECGs) in both adults and children requires age-specific criteria, and AI models primarily trained on adult ECGs often perform poorly when applied to pediatric populations, particularly when pediatric data is scarce. Existing multimodal methods that combine ECG waveforms and text typically align them at a global level, which can dilute specific diagnostic evidence and hinder effective transfer. To address this, a new framework called Pediatric-Adult ECG Alignment via Cross-modal Enhancement (PEACE) has been introduced. PEACE is pretrained on the extensive adult MIMIC-IV ECG corpus and uses a knowledge-guided approach. It describes each diagnosis along rhythm, morphology, and ST-T axes, composing only positive-label descriptors into axis tokens and a fused embedding for each recording. A label query network (LQN) then uses diagnostic labels to cross-attend over ECG and axis tokens. Furthermore, PEACE employs label set aware bidirectional contrastive learning (LSBC) to align pooled ECG features with the fused embedding when recordings share diagnoses. Curriculum adaptive fusion (CAF) dynamically adjusts alignment strength based on classification loss and training progress, preventing disruption during early optimization. The knowledge branch is used only during training, with inference relying solely on ECG signals. On the ZZU-pECG dataset, PEACE achieved substantial macro average AUC improvements, especially under zero-shot and 50-shot fine-tuning, demonstrating clear gains over baseline models and comparable performance to domain adaptation methods with limited supervision.