acceptodds
Withdrawn from ICLR 2027

ECG2Act: From Abnormality Recognition to Clinical Action Concern Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Electrocardiogram (ECG) analysis informs not only the diagnostic evaluation of cardiac abnormalities but also care planning, which often must proceed before a definitive diagnosis is established. Yet existing ECG deep learning methods mainly focus on abnormality recognition, without explicitly identifying the diagnostic and care-process concerns that guide subsequent clinical assessment and management. We therefore introduce ***clinical action concern prediction***, which jointly identifies problem-oriented diagnostic concerns and care-process concerns from a single ECG to inform subsequent clinical assessment and management. This task presents **two challenges**: the two target families share the same ECG input yet may rely on *different waveform features*, while the concern-level supervision indicates *which concerns are present* without specifying *which ECG characteristics support them*. To address these challenges, we develop **ECG2Act**, which learns a shared ECG representation and uses concern-specific action queries to retrieve target-relevant evidence, while utilizing measurement regression and evidence classification supervision to anchor this representation to interpretable ECG characteristics. On patient-disjoint evaluation of **179,100 ECGs from 82,486 subjects**, ECG2Act achieves diagnostic and care-process AUROCs of **0.938 and 0.713**, respectively. Crucially, care-process concerns remain distinguishable among ECGs sharing the same diagnostic profile, demonstrating that ECG-based care assessment captures information not represented by diagnostic recognition alone. These results support extending current ECG analysis beyond abnormality recognition to jointly assess diagnostic and care-process action concerns.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
ECG, Screening, Clinical action

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_xxPaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In ECG2Act, the authors propose a benchmark based on MIMIC-IV-ECG to assess the value of ECG data to inform clinical care decisions therefore extending beyond usual diagnostic prediction settings. The authors use an auxiliary ECG feature loss and report and slight improvement in predictive accuracy through it.

Critical strengths
  • The idea of extending clinical analyses beyond diagnostic/abnormality recognition to care processes is meaningful (but care processes will also depend on clinical judgement, local guidelines etc).
Critical weaknesses

Main

  • Main concern I: The study represents a purely synthetic use-case; care decisions won't be made based on the ECG alone but will incorporate clinical context.
  • Main concern II: "Together, these results support extending ECG analysis beyond abnormality recognition to jointly assess both diagnostic and care-process concerns." The study raises a question and could develop into an interesting direction (if extended by clinical context), but does not seem aligned with the topics of ICLR, i.e., provides little insights that extend beyond the use-case itself.
  • training encoders from scratch instead of using pretrained encoders severely limits the applicability of the study in the age of FMs.
  • "The diagnostic–care gap persists across concerns and ages." The gap between cardiac diagnostic conditions and care decisions is a result of the setup of the study: The ECG is a very informative modality for cardiac diagnostics, the restriction to ECG only leaves out relevant clinical information that would have been informative for care decisions.
  • The authors claim that the auxiliary losses lead to physiological grounding. There is no direct way for jointly predicted to influence the main prediction task (apart from a reorganization of the joint feature space). There is no evidence for the added value of auxiliary tasks beyond macro AUROC performance increases.
  • The discussion section is minimal and does not discuss a single limitation of the work.
  • No source code is provided, which severly limits the reproducibility of benchmarking studies like this.

Minor

  • Introduction: "Rates, intervals, and recognizable waveform patterns inform ECG interpretation (Kligfield et al., 2007; Mason et al., 2007), yet concern labels alone do not explicitly supervise these properties." but we do not want to connect them explicitly, rather learn it (and also features beyond these)
  • Presentation of ECG FMs is very selective and does not align with the state of the field. Consider rather citing reviews or recent benchmarking studies instead of selecting specific models.
  • Action query is very close to attention probes in V-JEPA (https://arxiv.org/abs/2404.08471) and should also be referenced as such.
  • There is no evidence for the added value of auxiliary tasks beyond macro AUROC performance increases, which is not

Summary

  • The study is a potentially interesting direction for a biomedical journal (when extending towards the inclusion of clinical context).
  • Basing the benchmark on ECG only is from my perspective results in an ill-defined prediction task.
  • The study provides little insights that extend beyond the presented use-case and seems poorly aligned with ICLR.
Questions for authors
  • Why should the ECG be relevant as sole observable to inform care process decisions?
  • While this is a potentially study for a biomedical journal, what is its insight for ICLR/machine learning community?
  • What is the added value of MEETI instead of MIMIC-IV-ECG itself if only samples are retained?
Reviewer_AjMAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ECG2Act, a framework for jointly predicting diagnostic and care-process concerns from a single ECG. It combines concern-specific action queries with physiological auxiliary supervision. Experiments demonstrate its predictive performance and examine care-related information beyond diagnostic labels.

Critical strengths

S1. The paper extends ECG analysis beyond abnormality recognition to jointly predict diagnostic and care-process concerns, addressing a clinically relevant problem.

S2. The paper shows that ECGs with the same diagnostic labels can still be used to distinguish different care-process outcomes, providing useful evidence for the proposed task.

Critical weaknesses

W1. The proposed framework mainly combines existing techniques, including label-specific attention and physiological auxiliary supervision. More importantly, the experiments lack comparisons with existing ECG models or standard multi-label classifiers, making it difficult to assess the advantages of ECG2Act.

W2. The authors explicitly state that the evaluation set was used to compare and select model configurations. This compromises the independence of the reported evaluation results and raises concerns about potential selection bias.

Questions for authors

Q1. Could the authors provide comparisons with existing ECG classification and clinical outcome prediction methods under the same experimental setting? Such comparisons are important for evaluating the advantages of ECG2Act over existing approaches.

Q2. Section 4.1 states that the evaluation set was used for model selection. Could the authors clarify how it was used and whether the reported results can be validated on an independent test set?

Q3. Could the authors provide an ablation study comparing the proposed action queries with standard classification heads using a shared ECG representation? This would help establish whether the action-query mechanism provides additional benefits.

Q4. The evidence classification objective uses the average representation of all eight action queries rather than supervising each query individually. Could the authors explain the rationale for this design and clarify how it supports concern-specific physiological grounding?

Q5. How were the auxiliary loss weights of 0.20 and 0.05 selected? Have the authors evaluated the sensitivity of model performance to these weights?

Reviewer_E7PCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper defines a task called clinical action concern prediction: from a single 12-lead ECG, predict four diagnostic labels and four care-process labels. The proposed ECG2Act uses a conv-Transformer encoder with one learnable query per label, plus auxiliary losses on machine measurements and report findings. It reports macro-AUROCs of 0.938 (diagnostic) and 0.713 (care) on MIMIC-IV-ECG.

Critical strengths
  1. Linking ECG analysis to downstream care is a clinically relevant direction, and the study uses a large cohort with patient-disjoint splits.
  2. The within-profile analysis is a reasonable way to check whether the care predictions go beyond the diagnostic labels.
Critical weaknesses
  1. The diagnostic "action concerns" are not really about action. They are coarse groupings of machine-report abnormalities by ECG phenotype, and each group mixes benign and urgent findings: acute ST elevation with old Q waves, sinus tachycardia with AF/SVT, first-degree AV block with high-grade AV block. The distinctions that would actually change management are lost, and findings such as QT prolongation are left out entirely.

  2. The care labels record what happened in the hospital, not what the ECG should flag. A repeat ECG within 24 h or an admission is largely decided by protocol, troponin and vital signs. Three of the four labels are essentially nested (ICU ⊂\subset monitoring, and both imply admission). I also do not see a realistic setting in which disposition would be predicted from one ECG alone.

  3. There is no comparison with any existing ECG model (e.g., Ribeiro et al., CLOCS, ECG-FM) or with simple variants of the architecture, so the value of the method cannot be judged. The model is a standard query decoder with auxiliary losses. ECG-based prediction of admission and deterioration has been studied before, so the gap claimed in the introduction is overstated.

  4. The test partition was used for model selection (Sec. 4.1). The claim that care predictions carry information beyond the diagnostic labels is tested only against four coarse binary labels. Age, sex, triage acuity, and the machine measurements and findings that are already available should be controlled for.

  5. Some smaller issues. Per-lead z-scoring removes absolute voltage, which matters for LVH, low voltage and ST elevation. Phase scrambling and time reversal are too coarse to show what the model relies on. In the attribution figures, every care concern lists the same findings, and the red/yellow severity is never defined. Label names change across figures (Serial/Repeat ECG, High-acuity care/review), and the learning rate, batch size and number of epochs are not reported.

Questions for authors

Mostly as in Weaknesses. In particular:

  1. How do standard ECG baselines, and a simple model using age, sex and triage acuity, perform on the same split?
  2. What are the results when model selection uses only the validation set?
  3. Does ECG2Act still add care information after controlling for machine measurements, report findings and demographics?

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

Related papers

Discussion (0)

Sign in to comment.