Label-Conditioned Cross-Modal Fusion for Adult-to-Pediatric ECG Transfer via Curriculum-Gated Contrastive Alignment

Automated pediatric electrocardiogram (ECG) interpretation remains challenging because developmental differences in heart rate, intervals, and waveforms limit the transferability of models trained mainly on adult data, while expert-labeled pediatric ECG cohorts are scarce. We propose PEACE (Pediatric-Adult ECG Alignment via Cross-modal Enhancement), an adult-to-pediatric ECG transfer framework pretrained on MIMIC-IV ECGs and adapted to pediatric targets. PEACE integrates label-specific bidirectional contrastive learning (LSBC) to align ECG representations with diagnostic semantics and curriculum adaptive fusion (CAF) to stabilize optimization under limited pediatric supervision. Label-conditioned short text descriptors provide auxiliary semantic supervision during training, whereas inference requires ECG signals only. On ZZU-pECG, PEACE achieves macro-average AUCs of 59.39%, 81.74%, and 91.56% under zero-shot, 50-shot, and full fine-tuning settings, respectively, outperforming ECG-only, multimodal, and generic domain adaptation baselines including DANN and MMD. On PTB-XL, it reaches 96.90% macro-average AUC after full fine-tuning over nine harmonized labels with nonzero mapped incidence. Gradient-based attention maps show increased saliency around QRS voltage and morphology regions for chamber-related RVH and around QRS-to-T/repolarization intervals for LQTS, broadly consistent with ECG regions commonly inspected during routine interpretation. These results suggest that adult-scale ECG pretraining coupled with rhythm, morphology, and ST-T repolarization semantic descriptors improves transferable pediatric diagnosis under label scarcity while preserving clinically interpretable waveform focus.