acceptodds
Under review as a conference paper at ICLR 2027

PhysioTrace: Tracing Zero-Shot ECG Decisions through Clinical Observations to Waveform Evidence

Abstract

ECG-language models enable flexible zero-shot diagnosis, yet their decisions remain difficult to interpret. Existing ECG explanations often focus on signal-level attribution that identifies *where* evidence lies in the waveform, while emerging ECG-language methods begin to expose *what* clinical observations support a diagnosis. However, these two views remain disconnected. We introduce **PhysioTrace**, a hierarchical explanation framework that makes ECG predictions traceable from diagnosis to clinical observation to waveform evidence. Its key idea is to use task-agnostic clinical observations as a shared intermediate representation: each diagnosis is decomposed into observation-level contributions, and each observation is grounded in the multilead ECG through observation-conditioned counterfactual attribution. Experiments across multiple ECG datasets and backbones show that PhysioTrace largely preserves zero-shot performance while providing faithful observation-level explanations and observation-specific waveform grounding. PhysioTrace therefore turns disconnected semantic and signal-level explanations into a unified and auditable reasoning structure. Code is available at https://anonymous.4open.science/r/PhysioTrace/.

Primary area:
learning on time series and dynamical systems
Keywords:
Physiological Time Series, ECG-Language Models, Zero-Shot Diagnosis, Explanation, Waveform Grounding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XHbgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PhysioTrace is a hierarchical explanation framework that makes zero-shot ECG predictions traceable from diagnosis to clinical observation to waveform evidence. Its central idea is to use task-agnostic clinical observations as a shared intermediate representation: a learnable projection maps both ECGs and diagnostic prompts into a common observation coordinate system, trained with matching and reconstruction losses, so that a diagnosis decomposes into observation-level contributions. Each observation is then grounded in the multilead signal via observation-conditioned counterfactual attribution, which ranks lead-time waveform regions by their evidence for that observation. Because the observation vocabulary is derived from real ECG reports rather than diagnosis labels, explanations are intended to transfer across diagnoses and datasets. The method operates on frozen ECG-language backbones and is evaluated on zero-shot classification retention, intervention experiments, and grounding.

Critical strengths
  • Using a task-agnostic observation space as the intermediate layer is a well-chosen design. It makes explanations reusable across diagnoses and datasets rather than bespoke per label, which is what allows a single projection to serve multiple backbones and tasks. This is the element that gives the framework scope beyond a single-model interpretability trick, and it is the main reason I view the paper as above clear rejection.
  • Joining semantic and signal-level explanation into one chain addresses a real gap. Attribution-only methods localize evidence on the waveform without saying what clinical observation it supports, while concept-based methods identify supporting observations without patient-specific localization. Connecting the two into an auditable diagnosis→observation→waveform path is a genuine contribution if the links hold.
Critical weaknesses
  • The report-derived observation space may be systematically incomplete, and the risk is neither discussed nor mitigated. The vocabulary is parsed from real ECG reports, so it can only encode what reports happen to state. It has been reported that a substantial share of ECG reports omit detailed waveform descriptions (e.g., FG-CLEP, Li et al., 2025), and the natural remedy — supplying LLM-generated, retrieval-grounded cardiological text as additional supervision (e.g., ESI, Yu et al., TMLR 2024) — is not discussed. Since the paper itself notes that evidence outside the vocabulary may remain unrepresented, completeness directly bounds explanation fidelity: an observation absent from the vocabulary cannot be recovered by any downstream attribution. MERL, D-BETA, MELP, SGERA, and ZETA are all cited, which makes the omission of this particular strand conspicuous. This is the weakness I would most want addressed, as it bears on soundness rather than positioning.
  • The evaluation does not establish that the explanation links are faithful. Counterfactual perturbation is applied to a frozen cross-modal encoder, and perturbed signals can readily fall outside the distribution that encoder was trained on, in which case attribution maps reflect out-of-distribution artifacts rather than true evidence; no diagnostic for this is reported. Compounding it, the central cost of the method — the change in zero-shot performance relative to the frozen backbone — is hedged as "largely preserves" rather than quantified per dataset and backbone, and the paper concedes that clinical validity of the observation space requires independent evaluation that is not provided. A framework whose contribution is explanation quality needs the price and the fidelity both stated numerically.
Questions for authors
  1. How do you ensure the observation vocabulary is complete, given that many ECG reports lack explicit waveform descriptions? Was any LLM-based or retrieval-augmented text enhancement applied to supplement the report corpus?
  2. How does your observation↔waveform grounding differ from prior fine-grained alignment between localized waveform segments and individual clinical tags? What does the counterfactual stage add beyond what such alignment already provides?
  3. Can you show that counterfactually perturbed ECGs remain in-distribution for the frozen encoder — e.g., via embedding-space distance, reconstruction diagnostics, or a comparison against in-distribution controls?
  4. Please quantify the zero-shot performance change versus the frozen backbone, per dataset and per backbone, rather than as an aggregate hedge.
  5. Was the observation vocabulary validated by clinicians, or checked against a curated cardiology ontology?
Reviewer_HFM2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PhysioTrace, which aims to make zero-shot ECG-language model predictions more interpretable by tracing a diagnosis through clinical observations to waveform regions. The method builds a task-agnostic vocabulary of 351 report-derived clinical observations and reuses it across different diagnostic prompts. It decomposes the learned projected diagnostic score into observation contributions and uses waveform perturbations to identify the lead-time regions associated with each observation. The paper evaluates the framework across six frozen ECG-language backbones and multiple datasets. The experiments examine zero-shot performance preservation, clinical observation recognition, observation-level faithfulness, and waveform localization.

Critical strengths
  • The first strength is that the paper aims to provide an interpretable chain from diagnosis to clinical observations and then to waveform regions.

  • Another strength is the ability to scale by reusing clinical observations, which the paper describes as task-agnostic. The paper basically builds one vocabulary of 351 report-derived observations and reuses it for different diagnostic prompts. This reusability can help the framework scale and hopefully work well on downstream tasks. However, it is worth noting that this does not imply unlimited generalization to unseen clinical settings.

  • Another interesting aspect is that the method largely preserves zero-shot performance, and this is well supported by Table 3. For example, the six-task mean macro-AUROC results are very clear.

  • The report-processing pipeline is described well enough to understand its overall design.

  • Another good thing about this method is that the observation phrases are assessed by clinical experts.

  • The experiments are quite good in that they use six frozen ECG-language backbones. This makes the results more convincing, as it shows that the observation-based approach is not entirely unique to a single architecture.

  • It is a good thing that the authors evaluate clinical observation recognition using expert annotations collected without access to model predictions or diagnostic labels. This makes the clinical observation recognition test more convincing and supports the validity of the reported results.

  • The paper provides readable observation contributions and corresponding lead-time maps. For example, these could be useful for ECG-based hypothesis generation and model debugging, without making strong claims about clinical deployment. I think this is a good contribution to the community. However, it is worth noting that making predictions easier to inspect is not itself a measure of improvement in clinical workflows.

Critical weaknesses

Weaknesses

  • The first major weakness is that the paper claims to provide faithful observation-level explanations. However, this is not entirely established because the explanation decomposes the learned projected score, while the frozen ECG-language model uses its own original score. The two scores are encouraged to agree, but they are not necessarily identical. Although the authors acknowledge this approximation in Sections 3.3 and 7.7, its implications for individual-patient explanations are not entirely clear. For example, Appendix Table 7 already reports original-versus-projected top-1 agreement across six tasks for MERL ResNet18, but the authors could extend this analysis to the other five backbones. They could also show explanation fidelity conditional on agreement, which would help readers understand the divergence in patient-level explanations.

  • Around line 153, it generally makes sense to use a frozen text encoder to obtain the semantic direction for each observation. However, although Table 2 validates recognition for seven observations and Figure 4(c) examines cross-observation responses, these experiments do not establish that all 351 coordinates retain their intended semantics. I suppose the authors could expand the existing concept-level recognition tests or cross-activation analysis to more observations to show whether the learned coordinates really retain their intended semantics.

  • Around lines 211–219, the authors divide a 10-second, 12-lead ECG into 40 windows per lead, creating 480 regions. Estimating the observation perturbation maps requires rerunning the frozen ECG encoder on the modified input for each region. Although the resulting representation can be reused for many observations, which is a good property, this can still be computationally expensive. Because the paper describes a practical explanation framework, it would be good if the authors reported wall-clock runtime and scalability so that readers could estimate the usability of this framework in practice.

  • Around line 242, the authors choose the largest positive values when visualizing waveform evidence. However, the actual diagnostic contribution can also be negative. A region whose removal reduces the observation activation can therefore increase the diagnostic contribution when the text coordinate is negative. Although Equation (12) preserves the sign of the text coordinate, visualizing only large positive observation responses does not provide a complete view of the diagnosis-level contributions. I think the authors should discuss this more clearly.

  • Table 2's patient-level evaluation is not really complete, although Table 1 assesses the clinical meaning of all 351 observation terms. It reports only seven observations from a vocabulary of 351, and the expert-label evaluation uses only 100 ECG samples. The problem is that a model can perform well on these representative observations but still fail on rare terms. The reported mean also covers only these seven observations. It would be better if the authors provided a more holistic evaluation of the 351 observation terms.

  • Table 3 could be improved if the authors bolded the best values. Without highlighting the best values, it is harder for readers to quickly identify the strongest results.

  • Around line 90, the authors claim that the method makes predictions systematically traceable. However, this is not entirely accurate because the explanations are tied to the projected prediction, not necessarily the original frozen model's prediction.

  • This is a minor point, but around line 336, could the authors explain why the average performance of the proposed method is actually worse than D-BETA + ZETA?

  • Around line 375, Figure 4(c) visually suggests that target-observation effects exceed off-target effects. Appendix Section 7.4 clarifies that this matrix tests observation specificity rather than anatomical localization, while Table 6 separately reports expert-verified localization. Without numerical diagonal-versus-off-diagonal effect sizes, uncertainty estimates, and sample coverage, it is harder to judge the strength of the observation-specificity claim. A useful addition would be to report the aggregate diagonal and off-diagonal contrast, confidence intervals, and the number of tested observation pairs.

  • Table 4 is a little strange because the reconstruction-only objective achieves almost perfect prediction fidelity, close to 1.0. Its classification performance is also generally higher than that of the full objective, but the full objective produces a higher Top-5 flip rate. Section 5.6 mainly describes these observed results without explaining why they happen. I wish the authors would explain more about the motivation or intuition behind combining the two objectives. For example, matching-only performs much worse than the original baseline. However, adding the matching objective to reconstruction improves the Top-5 intervention results while slightly reducing predictive performance. It is difficult for me to determine whether the improved Top-5 flip rate reflects better clinical interpretability or simply greater sensitivity to deleting large contributions. Is there some kind of synergy or trade-off between the two objectives? I think this interaction needs to be studied or explained more clearly.

Questions for authors

The following questions refer to the weaknesses discussed above. Clarifying these points will affect my assessment of the paper.

  • Could the authors clarify how faithful the explanations remain when the projected model disagrees with the original model? Can the results in Appendix Table 7 be extended across all six backbones?

  • How representative are the seven evaluated observations in Table 2 of the full 351-observation vocabulary? Is there additional evidence that the remaining observations retain their intended clinical semantics?

  • Given that removing the largest positive contributions directly reduces the projected score, how does the observation-deletion experiment establish clinical faithfulness beyond score sensitivity?

  • In Table 4, why does the full objective improve Top-5 deletion sensitivity while slightly reducing predictive performance compared with reconstruction-only? Does this improvement reflect better clinical interpretability?

  • Could the authors clarify the strength of observation specificity in Figure 4(c), particularly the quantitative difference between diagonal and off-diagonal effects and its uncertainty?

Reviewer_Ly1YOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper introduces physiotrace, a hierarchical explanation framework that makes ECG prediction traceable from diagnosis to clinical observation to waveform evidence. Compared to previous methods such as ZETA, this method introduces Observation-Conditioned Counterfactual Grounding to ground them to patient-specific waveform evidence. It improves the interpretability of ECG diagnosis model.

Critical strengths
  1. This paper builds on ZETA to improve the explainability and transparency of ECG diagnosis. It evaluates the proposed method across a broad range of models and datasets and includes expert evaluation of explanation quality.
Critical weaknesses
  1. The method used to ground patient-specific waveform evidence relies on well-established masking-based sensitivity analysis, as demonstrated by Zeiler and Fergus (2014). More broadly, feature attribution has been extensively studied through frameworks such as SHAP, although SHAP computes contributions differently from single-region occlusion. I encourage the authors to clarify how their approach addresses the specific characteristics of ECG signals, such as temporal structure and cross-lead dependencies, and what methodological advances it offers beyond adapting existing attribution techniques.
  2. The presentation is also difficult to follow in places. For example, the “Evidence Decomposer” appears to be a central component in Figure 2, but this term is not explicitly defined in the main text, making it difficult to connect the diagram to the method description. Figure 2 also depicts two text encoders. While the first is clearly part of the pretrained ECG–language model, the role of the second—and whether it shares weights with the first—is unclear from the figure. Consistent terminology and explicit descriptions of these components would improve readability.
Questions for authors
  1. What is the evidence decomposer?
  2. In Figure 2, there are 2 frozen text encoders. What is the difference between them?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.