PEACE: Cross-Modal Alignment to Transfer Adult ECG Diagnostic Models to Pediatrics

Forum topic · 小凯 · 2026-05-04

Summary

PEACE (Pediatric-Adult ECG Alignment via Cross-modal Enhancement) is a research framework addressing the scarcity of pediatric ECG data by leveraging abundant adult ECG data for robust pediatric diagnosis. Because children differ from adults in heart rate, cardiac size, electrophysiology, lead placement, and disease spectrum, directly transferring adult-trained models performs poorly, while simple fine-tuning risks catastrophic forgetting. PEACE instead aligns adult and pediatric ECG representations rather than transferring model parameters wholesale. Its technical approach combines: (1) three-axis clinical semantic decomposition of ECG signals (e.g., rhythm, conduction, morphology) to identify transferable components; (2) label-guided query feature extraction that uses diagnostic labels to pull diagnosis-relevant, cross-population features; (3) curriculum gating optimization that learns easier transferable features before harder pediatric-specific ones; and (4) cross-modal enhancement that incorporates clinical text alongside ECG signals to supply additional semantics for alignment. The key insight is 'selective transfer': identify which ECG characteristics are shared across ages, align those representations, and relearn pediatric-specific patterns. PEACE exemplifies how domain knowledge and curriculum learning can guide transfer learning in sensitive medical AI applications, potentially enabling accurate pediatric ECG diagnosis despite limited pediatric training data.

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 Models

Imagine a hospital using an AI model trained on adult ECG (electrocardiogram) data to diagnose children. The model performs well on adults but poorly on pediatric data, because:

Children's heart rates differ from adults'
Cardiac size and electrophysiological characteristics differ
The disease spectrum differs (congenital vs. acquired conditions)

Yet pediatric ECG data is scarce due to ethical constraints, difficult data collection, specialist annotation requirements, and parental consent barriers.

The Triple Dilemma of Pediatric ECG Diagnosis

1. Data scarcity: small datasets, expert annotation needs, special equipment 2. Population differences: faster heart rates, different QRS morphology, different relative lead placement, different disease spectra 3. Label inconsistency: adult label systems don't directly apply to children—some adult conditions are rare or absent in children, and vice versa

Existing approaches fall short: direct transfer performs poorly; training only on pediatric data leaves insufficient data; naive fine-tuning causes catastrophic forgetting of adult knowledge.

PEACE: Alignment Instead of Transfer

PEACE (Pediatric-Adult ECG Alignment via Cross-modal Enhancement) proposes aligning transferable features between adult and pediatric ECG at the representation level, rather than transferring model parameters directly.

Its technical components:

1. Three-axis clinical semantic decomposition — decompose ECG signals into clinically meaningful semantic axes (e.g., rhythm, conduction, morphology) and identify which axes transfer across age groups 2. Label-guided query feature extraction — use diagnostic labels as "queries" to extract diagnosis-specific features that generalize across populations 3. Curriculum gating optimization — learn easier, transferable features first, then harder, pediatric-specific ones 4. Cross-modal enhancement — combine ECG signals with clinical text to provide extra semantic information for alignment

Why Alignment Beats Transfer

Model-level transfer carries adult-specific information inside its parameters that simple fine-tuning cannot separate
Representation-level separation distinguishes transferable from non-transferable features, transferring only the former and relearning the latter
Semantic guidance: clinical decomposition provides priors (e.g., "rhythm" is more likely universal than "QRS amplitude")
Curriculum learning eases cross-domain adaptation by progressing from simple to complex

Key Takeaway

Transfer learning is not "moving house"—relocating everything. It is selective moving: take what is universal, reacquire what is specific. PEACE shows an elegant path for pediatric diagnosis: respect the differences, but exploit the commonalities—using abundant adult data to help scarce pediatric data reach accurate diagnosis for children.