acceptodds
Withdrawn from ICLR 2027

PedECG: Adult-to-Pediatric ECG Transfer Through Representation Reuse, Target Reinterpretation, and Evidence Validation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pediatric ECG interpretation is inherently development-dependent: the diagnostic significance of heart rate, conduction intervals, voltage, axis, and repolarization changes throughout childhood. Although large adult ECG datasets provide rich cardiac knowledge, directly transferring adult models to children can also transfer diagnostic rules calibrated to adult physiology. To address this challenge, we propose PedECG, a decomposable framework for adult-to-pediatric ECG transfer that separates representation reuse, target-side reinterpretation, and record-specific evidence validation. PedECG preserves adult pretrained waveform representations with a frozen encoder, reinterprets them using quantitative ECG measurements and age-indexed pediatric reference context, and verifies candidate diagnoses against evidence from the current ECG record. Experiments show that adult cardiac representations remain highly reusable in the pediatric domain, while structured target-side reinterpretation and record-specific verification further improve how this transferred knowledge is converted into pediatric predictions; the learned interpretation also adapts to changes in the pediatric physiological reference frame. By separating reusable physiological knowledge from population-specific diagnostic interpretation, PedECG offers a general strategy for extending medical AI from data-rich source populations to clinically distinct target populations, helping make large-scale learned medical knowledge usable beyond the cohorts on which it was originally developed.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Pediatric ECG, Transfer Learning, Representation Learning, Domain Adaptation, Multimodal Learning, Developmental Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_azqCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the transfer of knowledge from adult ECG models to pediatric ECG interpretation, with a particular focus on which knowledge learned from adult ECGs can be retained and which diagnostic rules need to be adjusted according to developmental stage. The authors propose PedECG, which decomposes the transfer process into three stages: Preserve–Reinterpret–Validate. The framework preserves representations learned from adult ECGs, reinterprets adult-derived evidence using age and pediatric physiological reference ranges, and then validates candidate diagnoses against the current ECG record.

The main contributions can be summarized as follows:

  1. The paper decomposes adult-to-pediatric ECG transfer into three components: representation reuse, target-domain reinterpretation, and record-level validation.
  2. PedECG freezes an adult ECG encoder to preserve source-domain knowledge, introduces age, quantitative ECG measurements, and PEDNORM to construct a pediatric-specific readout, and uses a multimodal verifier to reassess candidate diagnoses.
  3. The paper provides stage-wise experiments to analyze the role of different components in the proposed framework.

The main empirical findings are as follows. A linear probe trained on top of the frozen adult encoder performs comparably to a ResNet trained from scratch on pediatric data, while a randomly initialized frozen encoder performs substantially worse. The record-specific verifier also provides a noticeable performance improvement. However, the paper currently provides limited direct evidence for its central claim that developmental reinterpretation is a necessary component of adult-to-pediatric transfer.

Critical strengths

The paper identifies an important issue in cross-population transfer: source-domain representations and target-domain decision rules may exhibit different transfer properties. Waveform structures learned from adult ECGs may remain useful in pediatric populations, while the diagnostic interpretation of the same evidence can vary across developmental stages. This question has relevance beyond pediatric ECG classification and may be of broader interest to representation transfer and medical domain adaptation. The stage-wise formulation also enables a more informative analysis than standard end-to-end fine-tuning.

The linear-probe result on the frozen adult encoder is particularly informative. Its performance is close to that of a pediatric model trained from scratch and clearly better than that obtained from randomly initialized frozen features. This supports an interesting conclusion: the adult-to-pediatric population shift may not require relearning the underlying ECG representation from scratch, and a substantial part of the domain difference may instead arise at the interpretation and decision stages.

Critical weaknesses

The main concern is that the necessity of developmental reinterpretation has not yet been sufficiently established. The central claim of the paper is that adult representations can be preserved, while their interpretation should be adjusted according to pediatric developmental stage. The current experiments mainly show that the model responds to age and PEDNORM, but they do not yet demonstrate that this mechanism is necessary for the final predictive performance. Several structured-readout variants reported in the appendix achieve very similar results, and the full model does not clearly outperform several simpler alternatives. The authors also correctly note that the age-substitution experiment only demonstrates developmental responsiveness. It does not establish that the induced changes are physiologically correct, nor does it quantify the independent contribution of developmental conditioning to the final performance. Therefore, the motivation that pediatric physiological reference ranges vary with age is well supported, but the evidence is still insufficient to establish that the proposed developmental reinterpretation mechanism is the key component required to address this issue.

A second concern is that the necessity of the full three-stage framework remains open to alternative explanations. The verifier provides a clear performance gain, but the current results primarily show that a pediatric-adapted VLM with access to ECG images and quantitative measurements is useful. The blank-image and mismatched-image experiments demonstrate that the verifier uses information from the current ECG record, but they do not establish that a candidate-generation-then-verification design is preferable to directly training a multimodal pediatric predictor. Similarly, the adult-representation linear-probe experiment shows that adult features are useful, but it does not directly demonstrate that the final improvement of PedECG depends on preserving the adult representation. A simpler interpretation of the current results remains plausible: the performance gain may primarily come from strong adult ECG features, pediatric structured information, and a 4B VLM verifier, while the full Preserve–Reinterpret–Validate decomposition may not itself be necessary.

Questions for authors
  1. Under strictly matched architecture and model capacity, what is the performance of the full system when removing age, PEDNORM, and all developmental conditioning, respectively? If the performance changes only marginally, this would directly weaken the paper's central reinterpretation claim.
  2. Using the same Qwen3-VL-4B, the same pediatric supervision, and the same input modalities, how does a directly trained one-stage pediatric predictor perform? This comparison is important for determining whether the Validate stage provides a methodological advantage beyond the capacity of the multimodal model itself.
  3. Within the full PedECG pipeline, what happens when the frozen adult representation is replaced with a pediatric representation trained from scratch or another capacity-matched representation? This experiment would help determine whether the final system genuinely benefits from the Preserve stage.
Reviewer_38AcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose PedECG for adult-to-pediatric ECG transfer. The framework uses frozen adult ECG representations, incorporates age-dependent physiological attributes to generate pediatric diagnosis candidates, and validates them using a VLM. Experiments on a pediatric ECG dataset show improvements over the evaluated baselines.

Critical strengths
  • The paper addresses a meaningful transfer learning problem where adult ECG knowledge must be adapted to age-dependent pediatric diagnostic criteria.
  • The experiments provide useful analyses of representation reuse, structured readout, and evidence verification, with several ablations supporting the effectiveness of the overall framework.
Critical weaknesses
  • Although the structured readout improves over concatenation, simpler variants achieve slightly better performance (App. B.1). This raises questions about the necessity of the proposed more complex components.

  • The paper compares frozen adult representations with pediatric models trained from scratch, but does not evaluate fine-tuning of the adult-pretrained encoder, a conventional transfer learning strategy.

  • The empirical evidence supporting its broader significance remains limited. The evaluation uses a single pediatric dataset with relatively small performance gains, and the qualitative examples lack systematic expert validation. It remains unclear whether the proposed method improves the correctness of pediatric ECG interpretation beyond aggregated diagnostic metrics. Also, the developmental context substitution analysis (Sec. 4.4) demonstrates changes in predictions without establishing whether these changes are physiologically appropriate or improve diagnostic accuracy.

Questions for authors
  • Could the authors further clarify the necessity of the individual structured readout components, given the comparable performance of simpler variants?
  • How does PedECG compare with conventional fine-tuning of the adult-pretrained encoder on pediatric data?
  • Could the authors provide more qualitative analysis to demonstrate how the proposed framework improves pediatric ECG interpretation?
Reviewer_PNQWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PedECG, a framework for transferring adult ECG models to pediatric diagnosis. PedECG consists of three main stages: separates representation reuse, target-side reinterpretation, and record-specific evidence validation. To address age-related domain shifts between adult and pediatric ECG data, PedECG incorporates gating mechanisms, dynamic mapping, age-dependent decision thresholds, and Top-K selection for candidate generation. The generated candidates are subsequently verified, and their scores are fused to produce final diagnoses based on predefined decision thresholds. Experiments show the performance of PedECG.

Critical strengths

1.Clinically meaningful motivation. 2.Introduce PedNorm to normalize different measurements.

Critical weaknesses

The core concerns are based on the assumption that adult models can provide enough information. Considering the gap between adults and children, I don’t think adult models can capture sufficient information for children's ECG. Then, the following questions are raised naturally:

  1. What are the domain shifts between adult and pediatric ECG data, and can an adult-pretrained encoder capture sufficient information for pediatric ECG diagnosis? How can pediatric-specific information be learned?
  2. Confusion about gate. Can such a simple gating mechanism truly achieve effective domain adaptation? The gating mechanism may fail to recover information that has already been lost by the adult-pretrained encoder. Beyond those questions, some questions need further considered.
  3. Age already influences the final disease predictions through three different mechanisms. Is it necessary to introduce so many separate modules for age-dependent adaptation?
  4. The paper introduces too many modules and hyperparameters, making it difficult to identify which components actually contribute to the performance improvements.
  5. H1 in experiments:  The results in Table 2 only demonstrate that adult pretraining provides some benefits. However, the adult-pretrained model achieves performance comparable to that of the pediatric model trained from scratch. Why not simply train the model from scratch using pediatric data?
  6. H2 in experiments: The marginal performance improvement reported in Table 2 is insufficient to convincingly support the authors' conclusion.
Questions for authors

Please refer primarily to Questions 1 and 2 in Critical Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.