PEACE: Cross-modal Pediatric-Adult ECG Alignment for Robust Pediatric Diagnosis

PEACE: Cross-modal Enhanced Pediatric-Adult ECG Alignment for Robust Pediatric Diagnosis

> Paper: PEACE: Cross-modal Enhanced Pediatric-Adult ECG Alignment for Robust Pediatric Diagnosis > Authors: Xinran Liu, Yuwen Li, Hongxiang Gao, Heyang Xu, Jianqing Li, Zongmin Wang, Chengyu Liu > arXiv: 2605.00647 | 2026-04-30

The problem: diagnosing children with adult-trained models

Hospitals increasingly use AI models trained on adult ECG (electrocardiogram) data to diagnose children. This fails because:

• Children's heart rates differ from adults'
• Heart size and electrophysiological characteristics differ
• The disease spectrum differs (congenital vs. acquired conditions)
• Models that perform well on adult data degrade sharply on pediatric data

Meanwhile, pediatric ECG data remains scarce due to ethical constraints, difficult collection, and the need for pediatric cardiologist annotations.

Three challenges in pediatric ECG diagnosis

1. Data scarcity – small datasets, expert annotation requirements, special equipment and parental consent. 2. Population differences – faster heart rates, different QRS morphology, relatively different lead positions due to smaller body size, different disease spectrums. 3. Label inconsistency – adult label systems do not directly apply to children; some adult conditions are rare or absent in children and vice versa.

Existing workarounds all fall short: direct transfer performs poorly, training only on pediatric data lacks volume, and naive fine-tuning causes catastrophic forgetting of adult knowledge.

PEACE: align rather than transfer

PEACE (Pediatric-Adult ECG Alignment via Cross-modal Enhancement) does not directly transfer models. Instead, it aligns transferable features between adult and pediatric ECG at the representation level, with four components:

1. Three-axis clinical semantic decomposition – decompose the ECG signal into clinically relevant semantic axes (e.g., rhythm, conduction, morphology) and identify which axes are transferable across ages. 2. Label-query feature extraction – use diagnostic labels as queries to extract diagnosis-related features, which tend to generalize better across populations. 3. Curriculum gated optimization – learn in a difficulty-ordered curriculum: first simple, transferable features, then harder, pediatric-specific ones. 4. Cross-modal enhancement – combine ECG signals with clinical text to provide additional semantic information that helps align representations.

The analogy: training an adult cardiologist to read pediatric ECGs—not by making them forget adult knowledge, but by helping them understand which rules are universal and which are pediatric-specific.

Why alignment beats transfer

• Parameter-level transfer carries a large amount of adult-specific information that simple fine-tuning cannot separate.
• Representation-level separation in PEACE distinguishes transferable from non-transferable features, transferring only the former and relearning the rest.
• Semantic guidance from clinical decomposition provides priors (e.g., "rhythm" is more likely universal than "QRS amplitude").
• Curriculum learning eases adaptation by avoiding the hardest differences first.

Takeaways for cross-domain medical AI

1. What is shared vs. different between the source and target domains? 2. Can you perform "selective transfer" at the representation level? 3. Can domain knowledge (e.g., clinical semantic decomposition) guide the transfer? 4. Can curriculum learning mitigate cross-domain adaptation difficulty?

Transfer learning is not "moving house"—taking everything with you. It is "selective moving": take the universal, repurchase the specific. PEACE shows an elegant approach for pediatric diagnosis: respect the differences, exploit the commonalities, and use abundant adult data to help scarce pediatric data—so every child, regardless of age, can get an accurate diagnosis.