acceptodds
Under review as a conference paper at ICLR 2027

ECG-EGR: Event-Grounded Relational Reasoning for Clinical ECG Interpretation

Abstract

Electrocardiography (ECG) is fundamental to cardiovascular diagnosis, and deep learning has substantially advanced automated ECG classification and report generation. However, most existing approaches encode ECGs into latent representations optimized for downstream prediction, leaving the correspondence between model representations, localized waveform evidence, and clinical findings largely implicit. We introduce ECG-EGR, an Event-Grounded Relational framework that factorizes ECG interpretation through a sparse, variable-cardinality set of waveform-supported Events. Each Event carries explicit lead-time support, an immutable semantic address, and a mutable reasoning state, separating evidence localization from contextual interpretation. ECG-EGR first discovers grounded Events and aligns their semantics with weak clinical supervision through unbalanced optimal transport, then relationally composes Event states to capture temporal and cross-lead dependencies while preserving their grounded identity. A language model accesses ECG information only through the composed Event representations, while a finding-conditioned retriever links generated findings to immutable Event semantics, enabling structured provenance from clinical findings to source Events and waveform regions. Extensive experiments demonstrate strong report generation performance, including a BLEU-1 score of 0.876, while achieving the strongest cross-task representation transfer among the compared methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
ECG MLLM, Report Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8FY9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ECG-EGR reformulates ECG interpretation as reasoning over a sparse, variable-cardinality set of waveform-supported Events. Each Event carries explicit lead-time support, an immutable semantic address, and a mutable reasoning state, separating evidence localization from contextual interpretation. Events are discovered competitively and aligned to weak clinical supervision via unbalanced optimal transport, then relationally composed to capture temporal and cross-lead dependencies. A language model accesses the ECG only through composed Event representations, while a finding-conditioned retriever links generated findings back to Event semantics, giving structured provenance to source waveform regions. Validation covers report generation and cross-task representation transfer.

Critical strengths
  • The separation of an immutable semantic address from a mutable reasoning state is a genuinely clean architectural idea. It lets evidence localization stay fixed while contextual interpretation varies, which is a principled way to make grounding and reasoning coexist rather than compete. This design is what makes the provenance claim structurally possible at all, and it is the main reason I consider the paper above clear rejection.
  • Tracing a generated finding back to a specific lead-time waveform region targets a real clinical need. Most ECG report generators produce text without evidence, so a mechanism that links findings to source waveform regions is consequential if it works, and the reported cross-task transfer suggests the representation is not overfit to a single benchmark.
Critical weaknesses
  • Provenance faithfulness — the paper's central claim — is asserted but not validated. The headline contribution is that findings can be traced to source Events and waveform regions, yet no clinician-annotated evidence-location ground truth appears to be used. Without it, there is no way to tell whether an "immutable semantic address" denotes a diagnostically meaningful region or merely a stable learned cluster. Since the entire value proposition rests on traceability being genuine rather than nominal, this materially limits what the experiments can support.
  • The relationship to fine-grained waveform-level ECG-text alignment is not clarified. The paper grounds semantics at sub-recording granularity, which sits close to a line of work on grounded and fine-grained ECG-text alignment. GEM and D-BETA are discussed, but the strand aligning localized waveform segments with individual clinical tags (e.g., FG-CLEP, Li et al., 2025) is not covered. As the Event-level grounding operates on the same waveform-to-finding correspondence, it is hard to judge what the relational composition contributes beyond what tag-level alignment already achieves. This is the point I would most want addressed, as it bears directly on the novelty assessment.
Questions for authors
  1. Is any clinician-validated evidence-location ground truth used to verify provenance? If not, could you report agreement between retrieved waveform regions and expert annotations on even a small subset?
  2. Do discovered Events correspond to interpretable morphological categories, or are they learned clusters? Is there a quantitative measure of Event semantic consistency across recordings?
  3. How does Event-level grounding differ from prior tag-level fine-grained waveform alignment — are there findings the relational composition produces that alignment-based methods cannot?
  4. BLEU-1 of 0.876 is the headline metric, but it is largely unigram overlap. Could you report clinical-entity-level metrics, e.g., finding-level F1? A convincing response to (1) and (3) would meaningfully improve my assessment, since both concern whether the core claim is substantiated rather than merely demonstrated.
6 comments
AuthorsOpenReview ↗

Author Response to Reviewer 8FY9 (1/4)

Dear Reviewer 8FY9:

We sincerely thank the reviewer for raising this important concern. We fully agree that clinician-annotated, abnormality-specific evidence locations would provide the most direct ground truth for validating finding-level provenance. However, such annotations are currently unavailable in the ECG–report datasets used in our study. We are actively seeking collaborations with clinical institutions to establish a dedicated dataset with clinician-validated, diagnosis-specific waveform evidence annotations, which we believe would be a valuable resource for the community.

In the meantime, to provide independent quantitative evidence, we evaluate ECG-EGR on LUDB v1.0.1, a dataset comprising 200 twelve-lead ECG recordings with expert-delineated P-wave, QRS-complex, and T-wave boundaries. By comparing the learned Event supports against these expert annotations, we assess whether our representations capture clinically meaningful waveform structures without relying on Event-level supervision. The results have been added to Appendix C.3.

(1) Agreement with Expert-Annotated Waveform Regions (Appendix C.3.1).

We extract Event supports using the frozen Stage-I ECG encoder and CED; expert annotations are never used for training, parameter selection, or inference. For each recording, we aggregate the supports of all active Events into a lead-time localization map and compare it against expert annotations. AP and AUROC are computed on the 40-ms encoder grid (a bin is positive when at least 50% overlaps an annotated region), while soft Dice and soft IoU are computed at the original 500-Hz resolution. Mean instance coverage measures the aggregate Event support within each annotated waveform interval. All metrics are macro-averaged across recordings.

Table R1. Agreement between ECG-EGR Event supports and expert annotations on LUDB.

WaveAP ↑AUROC ↑Soft Dice ↑Soft IoU ↑Mean Inst. Cov. ↑
P0.380.780.320.1910.43
QRS0.610.840.470.2960.61
T0.530.800.430.2770.59

The strongest agreement is observed for QRS complexes (AP 0.61, AUROC 0.84, soft Dice 0.47), followed by T and P waves. These results provide independent expert-annotation-based evidence that the learned Event supports capture physiologically meaningful ECG waveform structures and exhibit consistent localization correspondence across different physiological components.

(2) Temporal Null-Control Analysis (Appendix C.3.2).

To assess whether the observed agreement depends on the original temporal alignment, we compare the Event supports against two within-record controls, each repeated 32 times. Circular Shift shifts all Event supports together in time, preserving support mass, lead composition, and temporal continuity. Time Permutation applies the same random temporal permutation to all Events and leads, disrupting temporal order while preserving their joint structure.

Table R2. Comparison of original ECG-EGR Event supports with temporal null controls on LUDB.

WaveMethodAP ↑Soft Dice ↑
PECG-EGR0.380.32
Circular Shift0.090.04
Time Permutation0.080.03
QRSECG-EGR0.610.47
Circular Shift0.100.06
Time Permutation0.090.05
TECG-EGR0.530.43
Circular Shift0.180.08
Time Permutation0.170.07

ECG-EGR consistently outperforms both temporal null controls across all three waveform components, demonstrating substantially stronger agreement with expert-annotated ECG regions. For QRS complexes, the original Event supports achieve 6.1–6.8 times the AP (0.61 vs. 0.10/0.09) and 7.8–9.4 times the soft Dice (0.47 vs. 0.06/0.05) of the controls. These substantial improvements provide strong evidence that ECG-EGR learns temporally grounded Event supports aligned with physiologically meaningful waveform structures, rather than relying solely on incidental overlap or the overall support distribution.

We also recognize the importance of rigorous clinical validation of finding-level provenance. The LUDB evaluation compares aggregate supports of all active Events with expert-annotated waveform regions, rather than finding-specific retrieved evidence with clinician-annotated diagnostic locations. Nevertheless, these results provide valuable independent evidence that our learned Event supports are meaningfully aligned with physiological waveform structures. We agree that diagnosis-specific expert annotations would enable a more comprehensive assessment of finding-level provenance, and regard such clinician-guided validation as an important next step in our efforts to strengthen the clinical reliability and interpretability of ECG-EGR.

We have added these analyses to Appendix C.3.1–C.3.2 (Tables 12–13, lines 1393–1457).

AuthorsOpenReview ↗

Author Response to Reviewer 8FY9 (2/4)

Dear Reviewer 8FY9,

Thank you for this important question. Discovered Events are learned, waveform-grounded evidence units, not predefined morphological classes or hard clusters with globally fixed identities. Their semantic Keys remain immutable during relational composition, but neither the Keys nor Event indices are universal labels shared across recordings. We therefore assess whether similar Event Keys exhibit consistent clinical associations across patients, rather than assume a one-to-one mapping to expert-defined morphologies.

Response to Q2: Do Events correspond to interpretable morphological categories, or learned clusters? Is their cross-recording semantic consistency measurable?

(1) Text-free Cross-recording Event Consistency (Appendix C.7.1).

Using the frozen Stage-I model, we construct a patient-disjoint PTB-XL retrieval benchmark (Gallery: folds 1–8; Query: fold 10; fold 9 excluded; one ECG per patient). Although Stage I was trained with report supervision, no clinical text is used at retrieval time: each Query Event ranks Gallery ECGs by its maximum cosine similarity to their active Events. We measure top-5 agreement using record-level SCP-label Jaccard similarity, positive-label recall, and prevalence-adjusted label lift. Results are averaged over Events within each Query ECG.

Table R3. Text-free cross-recording retrieval (PTB-XL).

RepresentationJaccard@5 ↑Recall@5 ↑Lift@5 ↑
Random0.34810.45661.006
Geometry-only0.44820.56261.577
Initial Value (V0V^0)0.48930.60612.003
Pooled-Key0.50010.61952.132
Semantic Key (KK)0.55090.66742.326

Semantic Keys outperform all controls. Relative to Pooled-Key, they improve Jaccard@5 by 0.0508 and Recall@5 by 0.0479; their Lift@5 of 2.326 indicates substantially greater clinical-label agreement than expected from Gallery-label prevalence. This suggests that individual Event Keys retain cross-recording clinical information beyond support geometry or record-level pooling.

(2) Finding-conditioned Cross-recording Event Retrieval (Appendix C.7.1).

We additionally test concepts from the PTB-XL diagnostic, form, and rhythm categories (up to six per category; ≥80 positive Gallery ECGs). For up to 60 positive Query patients per concept, we select the Event Key most similar to the concept description encoded by frozen BioLORD-2023; the selected Event then ranks Gallery ECGs through Event-Key similarity. Gallery SCP labels, rather than Event-level annotations, define the retrieval targets. Results are macro-averaged over concepts.

Table R4. Category-wise finding-conditioned retrieval (PTB-XL).

CategoryRepresentationAUROC ↑AUPRC ↑
DiagnosticInitial Value (V0V^0)0.7520.370
Semantic Key (KK)0.7810.379
FormInitial Value (V0V^0)0.6590.205
Semantic Key (KK)0.6860.213
RhythmInitial Value (V0V^0)0.7780.458
Semantic Key (KK)0.8060.499

Semantic Keys improve both AUROC and AUPRC across all three clinical categories. Overall, they achieve 0.758 AUROC / 0.364 AUPRC, versus 0.730 / 0.344 for Initial Values; hard-negative AUROC also improves (0.703 vs. 0.676). Direct Text Event-Max performs better (0.808 / 0.432 overall); unlike Event-Key matching, it directly ranks Gallery Events using the target concept embedding.

These results show that ECG-derived Event Keys support quantifiable cross-recording clinical-label consistency without predefined Event categories or fixed slot identities. We have added the analyses to Appendix C.7.1 (Tables 17–19, lines 1809–1889).

AuthorsOpenReview ↗

Author Response to Reviewer 8FY9 (3/4)

Dear Reviewer 8FY9,

Thank you for highlighting the distinction between Event-level reasoning and fine-grained ECG–text alignment.

Response to Q3: How does Event-level grounding differ from prior tag-level fine-grained waveform alignment — are there findings the relational composition produces that alignment-based methods cannot?

(1) Distinction from FOCAL and direct comparison.

FOCAL (Li et al., 2026; formerly FG-CLEP) uses Semi-UOT to align clinical tags with localized ECG patches for contrastive representation learning. ECG-EGR instead discovers a variable-cardinality set of lead–time-supported Events (CED), aligns findings with multiple Events via UOT (CET), and explicitly composes temporal, cross-lead, synchrony, and overlap relationships through relation-aware message passing (REC). During composition, supports and semantic Keys remain fixed, while a lineage matrix approximates source-Event information routing for finding-to-waveform provenance.

To compare downstream representations rather than decoder capacity, we adapt FOCAL to ECG report generation and use the same Qwen3-8B decoder for both methods. FOCAL was not originally designed for this task; the comparison therefore evaluates task-adapted representations, not FOCAL's original classification setting.

Table R5. Same-backbone report generation (↑ higher is better; ↓ lower is better).

Dataset / ModelB-1↑B-F1↑F-Pre↑F-Rec↑F-F1↑Hall.↓Omit.↓
MIMIC-IV-ECG (In-domain)
FOCAL (adapted)0.4410.9350.6770.6420.6590.1290.358
ECG-EGR0.8760.9620.8490.8290.8390.0450.171
PTB-XL (Out-of-domain)
FOCAL (adapted)0.0560.8460.6820.5330.5980.0580.467
ECG-EGR0.4120.8610.7610.6660.7100.0460.334

ECG-EGR improves Finding F1 by 0.180 on MIMIC-IV-ECG and 0.112 on PTB-XL, with better Finding Precision/Recall and lower error rates on both. We do not claim that tag-alignment models cannot predict particular findings; rather, ECG-EGR explicitly represents interactions among grounded evidence units. This comparison supports the full representation pipeline; the ablation below isolates REC's contribution. Full results are in revised Table 6 (lines 498-511).

(2) REC ablation.

We isolate REC by comparing the full model with w/o REC while keeping the remaining framework unchanged. Removing REC degrades report quality and clinical precision, increases hallucinations, and reduces both evidence-faithfulness metrics, supporting the contribution of relational composition beyond Event discovery and semantic alignment.

Table R6. Effect of relational composition.

SettingB-1↑B-F1↑F-Pre↑Hall.↓EG-Nec↑EG-Suf↑
Full0.8760.9620.8490.0450.0940.131
w/o REC0.5860.9240.5010.1370.0610.083

(3) Intervention validation (Appendix C.8).

We test whether lineage-retrieved evidence affects prediction, using teacher-forced NLL on the fixed target finding rather than relying on attribution visualization alone.

In C.8.1, we remove the highest-attribution pre-REC source Event and rerun REC and the decoder. Controls are matched on support mass, per-lead distribution, and temporal centroid. Target removal increases NLL by 0.2586, versus 0.0115 for controls. Across source-Event removals, mean per-finding Spearman correlation between attribution and NLL impact is 0.8282.

In C.8.2, we select positions covering 50% of attribution mass (capped at 15% of all positions), dilate by 5 samples, feather over 15, and attenuate to 10% amplitude. Three circularly shifted control masks preserve lead selection, coverage, and temporal structure. We rerun the ECG encoder, REC, and decoder. Target attenuation increases NLL by 0.0949, versus 0.0278 for controls.

Table R7. Intervention results (patient-level inference).

InterventionTarget ΔNLLControl ΔNLLPaired difference (95% CI)Positive patientsAdj. p
Source Event removal0.25860.01150.2471 [0.2294, 0.2669]94.5%0.0002
Raw waveform attenuation0.09490.02780.0671 [0.0611, 0.0732]87.1%0.0002

Confidence intervals and positive-patient rates are patient-level; p-values are Holm-adjusted across both endpoints. Together, the task-matched comparison, REC ablation, and controlled interventions show that relational composition improves finding-level prediction and that retrieved source Events and waveform regions are prediction-relevant, without equating attribution with clinician-verified localization.

These analyses appear in revised Table 6 and Appendix C.8 (Table 20, lines 1935–2030).

Reviewer_k7uiOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes representing ECG signals using a sparse set of Events instead of relying on global embeddings. Each Event is represented by waveform support, a semantic key, and a mutable reasoning state. The authors introduce several modules to support this representation, including Competitive Event Discovery (CED) and Compositional Evidence Transport (CET). The framework follows a two-stage training process. After Stage I is trained and frozen, the Relational Event Composer (REC) updates Event states through temporal, cross-lead, synchrony, and support relationships. During inference, the model first generates a clinical finding, then retrieves relevant Events and projects their weights back to waveform regions. The evaluation covers report generation, linear probing, zero-shot clinical concept transfer, Event stability, and the dependence of generated findings on retrieved Events.

Critical strengths
  • S1. Well-motivated and coherent network design. I think the method follows a good principle for designing the network. The authors start by representing each Event as an entity with its own identity, including waveform support, a semantic key, and a mutable state. This design is relatively easy to follow and provides a structured way to organize ECG information. As a result, using Events in this way also improves the inspectability of the AI model. I think Competitive Event Discovery shares some principles with previous approaches, particularly slot-based methods. However, the combination of waveform-constrained evidence reading, semantic alignment, and relational composition is quite novel in my opinion.

  • S2. Effective use of weak supervision. Another strength is the use of weak supervision, which is a good choice because Event-level annotations are not readily available for ECG data. Since the proposed Events are learned representations, they may not directly correspond to individual clinical findings. Therefore, learning their semantic alignment through weak supervision is a reasonable approach. I also appreciate that the semantic keys are constructed from ECG-derived information, while report embeddings provide supervision rather than being directly inserted into the key construction. This makes the design more convincing.

  • S3. Very strong report-generation performance. The evaluation results are impressive and quite convincing. The performance gains are very large. For example, ECG-EGR achieves a BLEU-1 score of 0.876 and a BLEU-4 score of 0.734, which are extremely strong results. The Finding F1 score of 0.839 is also impressive. These improvements are not limited to linguistic metrics but extend to clinical-content metrics as well. I hope the authors will release their code and checkpoints because these results could be very useful to the research community.

  • S4. Broad experimental evaluation across multiple tasks. The experimental evaluation is comprehensive, which is another major strength. Rather than evaluating the framework only on report generation, the authors also evaluate it across multiple tasks, including representation transfer, zero-shot concept recognition, and evidence-related analysis. I think obtaining such large improvements across different tasks provides stronger support for the proposed method than improvements on a single benchmark. The component ablation experiments are also valuable because they demonstrate that CED, CET, and REC contribute meaningfully to the overall performance.

  • S5. Strong representation transfer across tasks and supervision levels. Another particularly interesting result is the representation transfer performance. ECG-EGR achieves the best performance in 11 out of 12 PTB-XL settings, including experiments using only 1%, 10%, and 100% of the labeled data. The results under limited supervision are especially impressive. This suggests that the learned Event representations contain useful information beyond the report-generation task and can transfer effectively to different ECG classification tasks.

  • S6. More explainable and structured evidence representation. By design, the proposed network provides a structured mechanism for associating individual clinical findings with waveform-supported Events. Compared with conventional global embeddings, this representation makes it easier to inspect which Events and waveform regions are associated with a generated finding. Technically, the Event structure also allows us to intervene on particular Events and study how they affect the model's predictions. I think this is a valuable direction toward more explainable ECG interpretation. However, the connection between the final waveform attribution and the actual contribution of source Events is not fully established by the current experiments, as discussed in my weaknesses below.

Critical weaknesses
  • W1. The source-level intervention does not fully support the finding-to-waveform tracing claim.

    One of the main claims of the paper is that ECG-EGR can connect a generated clinical finding to the original ECG regions supporting it. The authors achieve this by tracing information from composed Events back to the original source Events and their waveform regions.

    To evaluate the relevance of retrieved evidence, the authors introduce EG-Nec and EG-Suf, which test whether retrieved Events matter to the generated findings. However, there is a subtle issue with where these interventions are performed. The experiments remove the retrieved composed Events, or retain only those Events, and compare the resulting prediction loss against matched control subsets. This demonstrates that the retrieved Events matter to the model's predictions.

    The problem is that these interventions are performed after information has already been combined across Events, rather than at the original source-Event or waveform level. Therefore, although the results demonstrate that the composed Events are important, they do not directly establish whether the final attribution to individual source Events or waveform regions is faithful.

    For example, a composed Event associated with region A may contain information received from region B. Removing that composed Event can show that its combined information contributes to the prediction, but it does not establish how much of the contribution should be attributed to region A or region B. This allocation depends on the REC lineage matrix, whose computation is still described at a relatively high level.

    I suggest that the authors clarify how this matrix is constructed and introduce an additional intervention on source Events before REC, or directly on the traced waveform regions, followed by recomputation of the downstream predictions. Specifically, can the authors evaluate the final source-Event lineage by intervening on high-weight source Events before REC, or on their waveform supports, and comparing the downstream effects against matched controls? This would more directly test the additional source-attribution step claimed by the method.

  • W2. Minor numerical inconsistencies in the Event-stability results.

    In the appendix, the numbers reported in Figure 5 do not fully match the corresponding descriptions on page 18, particularly for the different-ECG control condition. I consider this a minor issue because the discrepancies do not appear to change the main qualitative conclusion regarding Event stability. Nevertheless, the authors should check the reported values and make the figure and text consistent.

  • W3. The large report-generation gains are encouraging, but the contribution of the backbone is not fully isolated.

    I find the large improvements in report generation very encouraging. However, I would like to see a more controlled baseline comparison. The paper reports substantial improvements over existing methods, but the proposed framework uses Qwen3-8B as its language-model backbone. The component ablations in Table 4 are very good and demonstrate that each proposed module contributes meaningfully within ECG-EGR.

    However, even the model without REC already achieves a BLEU-1 score of 0.586, which exceeds the best external baseline result of 0.534 in Table 1. This raises the question of how much of the overall improvement comes from the proposed Event representation versus the language-model backbone and training configuration. Since Qwen3-8B is already a strong language model, it is possible that part of the performance improvement comes from the decoder rather than the proposed modules.

    Although the component ablations are useful, I would like to see a comparison in which a conventional ECG representation and the proposed Event-based representation are connected to the same Qwen3-8B decoder under comparable training conditions. This would help establish how much of the improvement is specifically attributable to the proposed representation.

    I would also appreciate more detailed ablations for report generation, potentially in the appendix. The reported BLEU-1 of 0.876 and BLEU-4 of 0.734 are exceptionally high. BLEU-4 is generally a more demanding metric than BLEU-1 because it requires longer n-gram matches, so achieving such a high BLEU-4 score is particularly impressive. Together with the strong clinical-content metrics, these results are almost too good to be true at first glance. I am not suggesting that the results are incorrect, but their magnitude makes a more controlled comparison especially valuable. Additional experiments could help clarify why the proposed framework achieves such large improvements.

  • W4. The data preparation and evaluation protocols need more detail for reproducibility.

    Assuming the code is not released at the time of review, there are several aspects of the paper that I would like the authors to clarify. First, even if all methods are trained on the same data, the description of how the training data is prepared is not sufficiently clear.

    For example, how are clinical reports segmented into individual findings? How are the findings normalized, and how are negations and uncertain findings handled? What preprocessing steps are used to remove irrelevant or non-informative text? These details are particularly important because the proposed framework relies on report-derived supervision for Event semantic alignment.

    Another concern is the dataset split. The authors mention an 80/10/10 subject-level split, but the exact patient IDs assigned to each partition are not provided. Even with the same splitting ratio, different implementations may produce different training and test sets. The paper specifies a random seed, but it is still unclear how to reproduce the exact subject allocation. Given the large performance gains and the lack of extensive multi-seed evaluations, I think providing the exact split files or splitting procedure would be helpful for reproducibility.

    The paper also briefly mentions report normalization for evaluation, but the exact preprocessing steps are not sufficiently documented. These choices can affect the reported BLEU, Finding F1, and other clinical-content metrics. Therefore, I would encourage the authors to provide more details about the data preparation and evaluation pipelines.

  • W5. Several important implementation details are missing from the model description.

    Moving from the data to the model architecture, I also find that several components are not described in enough detail for independent reproduction.

    For the ECG encoder, the authors mention local and contextual feature extraction but do not clearly specify the architecture, feature dimensions, or receptive fields. For CED, I could not find the exact number of Event queries, query-refinement iterations, or pruning thresholds.

    For CET, the authors describe the use of BioLORD and unbalanced optimal transport, but important optimal-transport hyperparameters and solver settings are not clearly provided. For REC, the exact graph construction procedure and update equations are also not sufficiently detailed.

    Finally, for the Qwen3 interface, the authors mention LoRA training but do not specify important configuration details such as the LoRA rank and target modules. Similarly, the architecture of the Set Memory Adapter is not described clearly enough.

    I understand that the paper has limited space and that not every implementation detail needs to appear in the main text. However, these components are central to the proposed method, and their exact configurations can significantly affect reproduction. I would encourage the authors to provide pseudocode, configuration tables, or implementation details in the appendix, especially if the code will only be released after acceptance.

Questions for authors

Most of my questions are in the weaknesses already. But here are some important ones:

  1. Regarding W3 (Backbone comparison): Could the authors clarify whether any baseline uses the same Qwen3-8B decoder and comparable training settings? A same-backbone comparison would help isolate the contribution of the proposed Event representation.

  2. Regarding W4 (Data preparation): Which MIMIC-IV-ECG version and report fields were used? Were all baselines trained and evaluated using identical ECG–report pairs, preprocessing, and splits?

  3. Regarding W1 (Source-level attribution): Could the authors clarify how the REC lineage matrix is constructed? Can source-level interventions verify whether the final waveform attribution faithfully reflects the information used by the model?

  4. Regarding W5 (Reproducibility): Could the authors provide the missing architecture details, key hyperparameters, and exact data-splitting procedure, or release configurations sufficient to reproduce the main results?

6 comments
AuthorsOpenReview ↗

Author Response to Reviewer k7uiW (1/5)

Dear Reviewer k7uiW:

We thank the reviewer for highlighting the distinction between composed-Event relevance and source-level attribution faithfulness. We clarify REC lineage construction and add source-Event and raw-waveform interventions.

Response to W1

"Could the authors clarify how the REC lineage matrix is constructed? Can source-level interventions verify whether the final waveform attribution faithfully reflects the information used by the model?"

(1) REC Lineage Matrix Construction (Appendix A).

The REC lineage matrix tracks how original Stage-I source-Event information propagates into composed Events through residual connections and attention. It is initialized as an identity matrix over active source Events, AREC(0)=IA_{\mathrm{REC}}^{(0)}=I. At layer rr, Aij(r)A_{ij}^{(r)} denotes the routing mass from source Event jj to composed Event ii. The relative magnitude of the gated residual update is:

ηi(r)=RMS⁡(ΔVi(r))RMS⁡(Vi(r))+RMS⁡(ΔVi(r))+ϵ\eta_i^{(r)} = \frac{\operatorname{RMS}(\Delta V_i^{(r)})}{\operatorname{RMS}(V_i^{(r)}) + \operatorname{RMS}(\Delta V_i^{(r)}) + \epsilon}

Let Dη(r)D_\eta^{(r)} be the diagonal matrix of residual-update coefficients and αˉ(r)\bar{\alpha}^{(r)} the head-averaged, neighbor-normalized attention matrix. The lineage update is:

AREC(r+1)=RowNorm⁡[(I−Dη(r)+Dη(r)αˉ(r))AREC(r)]A_{\mathrm{REC}}^{(r+1)} = \operatorname{RowNorm}\left[\left(I-D_\eta^{(r)}+D_\eta^{(r)}\bar{\alpha}^{(r)}\right)A_{\mathrm{REC}}^{(r)}\right]

RowNorm normalizes each row to sum to one. The two terms preserve residual lineage and propagate it through attention, respectively. For finding ff, source attribution is Psrc=PcompARECP_{\mathrm{src}}=P_{\mathrm{comp}}A_{\mathrm{REC}}; source weights are then aggregated over immutable waveform supports. Importantly, ARECA_{\mathrm{REC}} represents detached Event-value routing provenance, not an exact causal decomposition: pairwise relation embeddings affect message construction but are not separately decomposed. The matrix is used only for provenance analysis, not as language-model input.

(2) Pre-REC Source-Level Intervention (Appendix C.8.1).

For each finding, we remove the highest-attribution source Event before REC by zeroing its active indicator, Event Key, and initial Event Value, then rerun REC and the decoder with other Stage-I outputs fixed. A control Event is matched by support mass, per-lead support distribution, and temporal centroid. We measure teacher-forced mean NLL on the fixed target finding's lexical tokens, retaining preceding findings. Removal of the attributed Event increases NLL by 0.2586, versus 0.0115 for the control, yielding a paired difference of 0.2471 (95% CI: [0.2294, 0.2669]; 94.5% positive patients; adjusted p=0.0002p=0.0002). Source-sweep analysis further gives a mean Spearman correlation of 0.8282 between attribution weights and leave-one-Event-out NLL increases. These results show that lineage weights identify source Events with greater actual predictive contributions.

(3) Raw-Waveform Intervention (Appendix C.8.2).

For each finding, we attenuate the smallest set of attributed lead-time positions covering 50% of attribution mass, capped at 15% of all positions. The regions are dilated by five samples, feathered over 15 samples, and attenuated to 10% amplitude. Three circularly shifted masks preserving perturbation geometry serve as controls. We recompute Stage-I, REC, and the decoder with the target finding and preceding context fixed. Attributed-region attenuation increases NLL by 0.0949, versus 0.0278 for shifted controls, yielding a paired difference of 0.0671 (95% CI: [0.0611, 0.0732]; 87.1% positive patients; adjusted p=0.0002p=0.0002). This demonstrates that the attributed waveform regions contribute more to finding prediction than matched perturbations at other temporal locations.

Table R1. Source-level and raw-waveform intervention results on MIMIC-IV-ECG.

InterventionTarget ΔNLL ↑Control ΔNLLPaired Difference (95% CI)Positive Patients (%)Adjusted p
Pre-REC Source Removal0.25860.01150.2471 [0.2294, 0.2669]94.50.0002
Raw-Waveform Attenuation0.09490.02780.0671 [0.0611, 0.0732]87.10.0002

Target and control ΔNLL denote NLL increases relative to the unmodified input. Confidence intervals and positive-patient percentages use patient-level bootstrap; adjusted pp-values use Holm correction across the two primary endpoints.

We have added the lineage initialization and update equations to Appendix A (REC Source Lineage and Waveform Provenance, lines 875-952) and the intervention protocols and results to Appendix C.8.1–C.8.2 (Table 18, lines 1938-2033).

AuthorsOpenReview ↗

Author Response to Reviewer k7uiW (3/5)

Dear Reviewer k7ui:

We thank the reviewer for raising the important concern about the contribution of the language model backbone. To better isolate the contribution of ECG representations, we have added a same-backbone comparison using a shared Qwen3-8B decoder, including Dense-Context, text-only Event-REC, and FOCAL-based representations.

Response to W3

"Could the authors clarify whether any baseline uses the same Qwen3-8B decoder and comparable training settings? A same-backbone comparison would help isolate the contribution of the proposed Event representation."

(1) Same-Backbone Comparison of ECG Representations (Table 6).

All four configurations use the same pretrained Qwen3-8B decoder backbone, allowing us to compare different ECG representations while controlling for backbone differences. Specifically, we compare waveform-grounded Event representations (ECG-EGR), a text-only Event-REC variant trained without citation supervision, Dense contextual representations without explicit Event discovery, and FOCAL-based representations adapted for report generation.

To ensure a consistent comparison, all configurations use the same underlying ECG–report pairs and patient-level data partitions, together with identical reference-report preparation and evaluation normalization. Model-specific input transformations are retained as required by each representation architecture.

Table R2. Report generation performance of different ECG representations using a shared Qwen3-8B decoder.

DatasetRepresentationB-1 ↑B-4 ↑M ↑R-L ↑B-F1 ↑F-Pre ↑F-Rec ↑F-F1 ↑Hall. ↓Omit. ↓
MIMICECG-EGR (Full)0.8760.7340.8710.8430.9620.8490.8290.8390.0450.171
MIMICEvent-REC (Text-only)0.6880.5360.7440.7880.9590.8370.8120.8240.0400.188
MIMICDense-Context (Text-only)0.6770.5280.7370.7860.9510.8350.8070.8210.0500.193
MIMICFOCAL0.4410.3110.5610.5830.9350.6770.6420.6590.1290.358
PTB-XLECG-EGR (Full)0.4120.2390.4680.4710.8610.7610.6660.7100.0460.334
PTB-XLEvent-REC (Text-only)0.3040.2010.3670.4190.8540.6410.6450.6430.0490.355
PTB-XLDense-Context (Text-only)0.1040.0160.2020.3110.8630.5970.6450.6200.0390.355
PTB-XLFOCAL0.0560.0170.2030.3030.8460.6820.5330.5980.0580.467

On MIMIC-IV-ECG, ECG-EGR achieves a Finding F1 of 0.839, outperforming Event-REC (0.824), Dense-Context (0.821), and FOCAL (0.659). It also achieves the lowest Omission Rate of 0.171, compared with 0.188, 0.193, and 0.358, respectively.

On the out-of-domain PTB-XL dataset, ECG-EGR further achieves the highest Finding F1 of 0.710, compared with 0.643, 0.620, and 0.598, respectively, while maintaining the lowest Omission Rate of 0.334, compared with 0.355, 0.355, and 0.467.

Notably, the Finding F1 improvement over Dense-Context increases from 0.018 on MIMIC-IV-ECG to 0.090 on PTB-XL, indicating a particularly strong advantage under out-of-domain evaluation. We also note that Dense-Context achieves a slightly higher BERTScore-F1 (0.863 vs. 0.861) and a lower Hallucination Rate (0.039 vs. 0.046) on PTB-XL, indicating that the advantage of ECG-EGR is particularly evident in clinical finding generation and omission reduction rather than uniformly across all metrics.

These results demonstrate that the improvements are not solely attributable to the Qwen3-8B backbone. Under the same decoder backbone and matched data and evaluation protocols, waveform-grounded Event representations provide additional benefits for clinical finding generation, with particularly pronounced gains under out-of-domain evaluation.

We have added this same-backbone representation comparison and its analysis to Section 5.4, Table 6 (Same-Backbone Comparison of ECG Representations, lines 481–519).

AuthorsOpenReview ↗

Author Response to Reviewer k7uiW (4/5)

Dear Reviewer k7uiW:

We thank the reviewer for raising this important question regarding dataset specification and experimental consistency. These details were provided in Appendix B.1 (Datasets and Data Preparation) in the original manuscript. We have further clarified and expanded this section to make the data sources, report construction, and baseline evaluation protocol more explicit.

Response to Q1/W4

"Which MIMIC-IV-ECG version and report fields were used? Were all baselines trained and evaluated using identical ECG–report pairs, preprocessing, and splits?"

(1) MIMIC-IV-ECG Version and Report Preparation (Appendix B.1).

We use MIMIC-IV-ECG v1.0 as our training dataset, containing approximately 800,000 diagnostic 12-lead ECG recordings from nearly 160,000 patients. Each recording is sampled at 500 Hz for 10 seconds. We extract machine-generated interpretation text from the report 0–report 17 fields in machine_measurements.csv, matching reports to ECG recordings using subject_id and study_id. Nonempty report entries are retained in their original order. We remove predefined data-quality warnings, acquisition errors, lead-reversal indicators, pacemaker-related statements, formatting artifacts, boilerplate, and exact duplicate findings, while preserving clinical wording, negations, and uncertainty expressions. The retained findings are combined into a unified natural-language report for generation.

(2) Consistency Across Baselines.

All report-generation baselines are trained and evaluated using the same ECG–report pairs and patient-level partitions. We use a deterministic 80/10/10 subject-level training, validation, and test split (random seed 42), ensuring that recordings from the same patient never cross partitions. The same reference-report preparation and evaluation normalization, including text formatting and clinical finding extraction, are applied across methods. Although model-specific ECG input transformations may differ according to architectural requirements, the underlying ECG recordings, reference reports, and evaluation protocol remain unchanged.

We have further clarified these dataset specifications, report preprocessing procedures, split protocols, and baseline data-consistency measures in Appendix B.1 (Datasets and Data Preparation, lines 956-1017).

Reviewer_A17QOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes ECG-EGR, a framework that represents ECG signals as sparse, waveform-grounded Events. It introduces Competitive Event Discovery (CED) for adaptive Event identification, Compositional Evidence Transport (CET) for semantic alignment, and a Relational Event Composer (REC) for modeling inter-Event relationships. Experiments demonstrate improvements in clinical report generation and cross-task representation transfer.

Critical strengths

S1. The paper introduces an adaptive Event representation grounded in local ECG waveforms. This approach seems promising for capturing diverse ECG abnormalities and helping trace clinical findings back to the original signal.

Critical weaknesses

W1. It remains unclear whether the learned Events actually capture meaningful ECG abnormalities. Although the paper provides visualizations and Event-level intervention results, it lacks direct validation of whether the selected waveform regions match clinical evidence. Also, since the model generates findings before retrieving supporting Events, it is uncertain whether the retrieved regions actually influenced the model's decisions. I would like to see comparisons with expert-annotated ECG regions and controlled experiments testing how predictions change when the retrieved waveform regions are removed.

W2. I noticed several numerical inconsistencies in the paper. For example, Figure 5 reports a Support Soft-IoU of 0.891 under 30 dB Gaussian noise, while Appendix B.3 claims it remains above 0.91. Under 20 dB noise, the figure shows a matched V0V^0 cosine of 0.990, but the text reports 0.757. In Table 1, ECG-Chat has a Finding Precision of 0.698 and a Hallucination Rate of 0.394. According to Appendix A.3, contradicted findings should be a subset of false positives. If both metrics use the same prediction set and micro-aggregation protocol, the Hallucination Rate should therefore not exceed (1-0.698=0.302). I hope the authors can clarify these discrepancies and verify the reported results.

Questions for authors

Q1. How do the authors verify that the learned Events correspond to clinically meaningful ECG abnormalities? Could they provide a quantitative comparison with expert-annotated waveform regions?

Q2. Since supporting Events are retrieved after the findings are generated, how can we know whether the model actually relies on these regions? Could the authors test how predictions change when the corresponding ECG waveform regions are removed?

Q3. Could the authors clarify the numerical inconsistencies between Figure 5 and Appendix B.3, as well as the apparent inconsistency between Finding Precision and Hallucination Rate for ECG-Chat in Table 1, as discussed in W2?

Q4. What value of Kmax⁡K_{\max} is used in the experiments, and how many Events are typically discovered per ECG? Does allowing the number of Events to vary actually improve performance compared with using a fixed number?

Q5. Given the large improvements in BLEU scores, could the authors clarify the format of the generated reports and provide examples comparing ECG-EGR, baseline outputs, and reference reports? It would also be helpful to understand whether differences in report length or formatting affect these results.

Q6. The report generation experiments only use Qwen3-8B. Have the authors tested other language models or model sizes to see whether the improvements generalize across different backbones?

9 comments
AuthorsOpenReview ↗

Author Response to Reviewer A17Q (1/6)

Dear Reviewer A17Q:

We thank the reviewer for recognizing ECG-EGR's waveform-grounded Events. We address the concerns on clinical meaning and waveform localization through semantic, functional, and expert-localization evaluations.

Response to Q1

"How do the authors verify that the learned Events correspond to clinically meaningful ECG abnormalities? Could they provide a quantitative comparison with expert-annotated waveform regions?"

(1) Clinical semantic validation (Appendix C.5).

Learned Events encode discriminative information associated with specific ECG findings, particularly in their semantic Keys. We evaluate 17 ECG findings defined by PTB-XL SCP codes. For each ECG, we average active Event Keys (semantics), Values (waveform features), or Geometry (lead-time supports) and train class-balanced logistic-regression probes. We also match individual Keys to BioLORD-2023 embeddings of finding descriptions without training classifiers; Key-Max takes the highest Event–finding similarity per ECG.

Table R1. Clinical finding prediction and zero-shot Event-to-concept association on PTB-XL (macro-averaged over 17 findings).

EvaluationMethodAUROCAUPRCHard AUROCHard AUPRC
SupervisedGeometry0.82530.32780.77530.4518
Value0.89150.53900.85920.6345
Key0.92080.58960.89360.6878
Zero-shotKey-Max0.78770.42510.72700.5327
Key-Random0.57270.23430.54220.3508
Mismatched Text0.49610.12720.44770.2481

Event Keys outperform Values and Geometry (0.9208/0.5896 AUROC/AUPRC) and retain strong performance on hard negatives with related, non-target SCP diagnoses (0.8936/0.6878). Key-Max exceeds Key-Random by 0.2150 AUROC and 0.1908 AUPRC; mismatched descriptions yield 0.4961 AUROC.

(2) Functional validation (Appendix C.5).

Removing clinically selected Events more strongly affects their target findings than matched alternatives, with effects specific to corresponding findings. Using selectors and independent evaluators trained on disjoint PTB-XL folds, we measure target-finding logit decreases after removing top-ranked Events from held-out ECGs, against count-matched controls, including Events matched by lead-time support geometry.

Table R2. Significant comparisons in Event-removal and concept-specificity analyses.

ComparisonSignificant findings
Top-ranked vs. geometry-matched Events16/17
Top-ranked vs. random Events16/17
Top-ranked vs. low-importance Events17/17
Corresponding vs. other finding evaluators17/17

Only sinus tachycardia (STACH) is non-significant in the first two comparisons after Holm correction. Specificity analysis using all 17 evaluators shows greater effects on the corresponding finding in all 17 cases (95% CIs above zero; Holm-adjusted p = 0.0017), indicating finding-specific rather than generic effects.

(3) Expert-annotated waveform localization (Appendix C.3).

Event supports align with independently annotated ECG waveform structures, and this agreement diminishes when temporal alignment is disrupted. We lack clinician-annotated regions for specific diagnostic abnormalities; addressing this gap motivates our work. We therefore use LUDB v1.0.1 (200 twelve-lead ECGs) with expert-delineated P-wave, QRS, and T-wave intervals. The Stage-I encoder and CED are frozen; annotations are not used for training, selection, or inference. Active Event supports are aggregated into lead-time maps for comparison.

Table R3. Localization agreement with expert annotations and temporal controls on LUDB (macro-averaged across recordings).

WaveMethodAPSoft Dice
PECG-EGR0.380.32
Circular Shift0.090.04
Time Permutation0.080.03
QRSECG-EGR0.610.47
Circular Shift0.100.06
Time Permutation0.090.05
TECG-EGR0.530.43
Circular Shift0.180.08
Time Permutation0.170.07

Original supports outperform both controls for all waves. For QRS, AP/Soft Dice fall from 0.61/0.47 to 0.10/0.06 and 0.09/0.05 under Circular Shift and Time Permutation (32 repetitions each).

LUDB annotates physiological components, not diagnosis-specific abnormalities; thus, this does not establish precise localization of individual diagnoses. Together, these experiments support the clinical relevance and waveform grounding of learned Events.

We have added these analyses in Appendix C.5 (Tables 15–16 and Figures 10–11, lines 1575–1753) and Appendix C.3 (Tables 12–13, lines 1393–1442).

AuthorsOpenReview ↗

Author Response to Reviewer A17Q (2/6)

Dear Reviewer A17Q:

We thank the reviewer for raising the important question of whether retrieved waveform evidence genuinely contributes to finding generation. We address this concern through raw-waveform intervention experiments and complementary Event stability analyses.

Response to Q2

"Since supporting Events are retrieved after the findings are generated, how can we know whether the model actually relies on these regions? Could the authors test how predictions change when the corresponding ECG waveform regions are removed?"

(1) Finding-conditioned Waveform Intervention (Appendix C.4.1).

To directly examine the contribution of retrieved evidence, we intervene on the original ECG waveforms at the regions attributed to each generated finding. On the full held-out MIMIC-IV-ECG test set, we first generate findings and obtain their finding-specific waveform attribution maps. We then zero out highly attributed samples, targeting up to 50% of attribution mass, with maximum occupancies of 15% across all lead-time positions and 40% per lead. We compare against five random masks with matched per-lead sample counts and no more than 5% overlap with retrieved regions, and a disjoint low-attribution control.

We keep the original findings fixed and measure their teacher-forced NLL change on finding-text tokens. We evaluate two pathways: Full-Pipeline, which re-encodes the modified ECG and reruns CED, Event grounding, and REC; and Fixed-Support, which preserves the original Event supports and assignments while recomputing waveform-derived Event values and semantic Keys before REC. All model parameters remain frozen.

Table R4. Finding NLL increases after waveform intervention and paired differences against control regions (95% CIs).

InterventionRetrieved ΔNLLRandom ΔNLLRetrieved–RandomRetrieved–Low
Full-Pipeline0.078190.045530.03265 [0.02508, 0.04019]0.02561 [0.01857, 0.03303]
Fixed-Support0.022480.009880.01260 [0.00985, 0.01544]0.01316 [0.01034, 0.01588]

Removing retrieved regions consistently increases finding NLL more than removing either control region, with all four paired 95% CIs excluding zero. Notably, the effect persists under Fixed-Support intervention: even when Event locations remain unchanged, removing their underlying waveform content increases finding NLL. This directly supports the predictive relevance of the retrieved waveform evidence. The larger Full-Pipeline effect additionally reflects the impact of waveform perturbations on subsequent Event discovery and reasoning.

(2) Complementary Event Stability Analysis (Appendix C.4.2).

Learned Event representations also remain consistent under common ECG signal perturbations. We evaluate 5,000 held-out ECGs under Gaussian noise, baseline wander, and temporal shifts, matching Events using Hungarian assignment based exclusively on support Soft-IoU. Under 20 dB Gaussian noise, support Soft-IoU is 0.757, while matched initial Event value and Key similarities remain at 0.990 and 0.996, respectively. In contrast, the different-ECG control yields a support Soft-IoU of only 0.042. These results complement the intervention experiments by demonstrating the stability of learned Event representations under nuisance perturbations.

Together, the raw-waveform interventions and stability analyses provide evidence for the predictive relevance of retrieved waveform regions and the robustness of learned Events.

We have added these analyses in Appendix C.4.1 (Table 14, lines 1478–1511) and Appendix C.4.2 (Figure 9, lines 1512–1565).

AuthorsOpenReview ↗

Author Response to Reviewer A17Q (3/6)

Dear Reviewer A17Q:

We thank the reviewer for carefully identifying these numerical inconsistencies. We have thoroughly rechecked the original experimental results, numerical outputs, and metric calculations. We confirm that these discrepancies were typographical errors in the manuscript, rather than inconsistencies in the underlying experiments. We have corrected the corresponding descriptions and table entries as detailed below.

Response to Q3

"Could the authors clarify the numerical inconsistencies between Figure 5 and Appendix B.3, as well as the apparent inconsistency between Finding Precision and Hallucination Rate for ECG-Chat in Table 1?"

(1) Event stability results (Appendix C.4.2).

We rechecked the original Event-matching results under Gaussian noise and verified all three measurements: support Soft-IoU, matched initial Event value similarity (V0V^0), and matched Event Key similarity (KK). The correct values are:

ConditionSupport Soft-IoUMatched V0V^0 cosineMatched KK cosine
Gaussian noise (30 dB)0.8910.9980.999
Gaussian noise (20 dB)0.7570.9900.996

The statement that support Soft-IoU remains above 0.91 under 30 dB noise was incorrect; the correct value is 0.891. Under 20 dB noise, 0.757 corresponds to support Soft-IoU, not matched cosine similarity; the correct matched V0V^0 and KK similarities are 0.990 and 0.996, respectively. These corrections align the textual descriptions with Figure 5 and the original experimental outputs.

(2) Finding Precision and Hallucination Rate (Table 1).

We also rechecked ECG-Chat's original evaluation results and the definitions in Appendix A.3. The Hallucination Rate in Table 1 was mistakenly reported as 0.394; its correct value is 0.294. This resolves the apparent inconsistency:

1−Finding Precision=1−0.698=0.302\mathbf{1-\mathrm{Finding\ Precision}=1-0.698=0.302}

Since contradicted findings are a subset of false positives, the Hallucination Rate must not exceed the false-positive proportion. The corrected value satisfies this relationship:

Hallucination Rate=0.294≤0.302\mathbf{\mathrm{Hallucination\ Rate}=0.294\leq 0.302}

Both metrics use the same generated finding set and micro-aggregated evaluation protocol, consistent with the definitions in Appendix A.3.

We sincerely apologize for these reporting errors. All corrections have been verified against the original experimental results; no underlying experimental results or evaluation protocols have been changed. We have corrected the Event stability descriptions in Appendix C.4.2 (lines 1512-1565) and the ECG-Chat Hallucination Rate in Table 1 (lines 270-285).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

Dear Reviewers,

We sincerely thank all reviewers for their thoughtful reviews, constructive suggestions, and the time devoted to evaluating our work. We are particularly grateful for their recognition of the adaptive, waveform-grounded Event representation and its potential for clinical traceability (Reviewer A17Q), the principled separation of immutable semantic addresses from mutable reasoning states and the clinical significance of waveform-level evidence tracing (Reviewer 8FY9), and the coherent and novel architectural design, effective weak supervision, strong report-generation performance, and broad cross-task representation transfer (Reviewer k7ui). These valuable assessments and suggestions have guided us in substantially strengthening the paper.

During the rebuttal period, we conducted extensive additional experiments, analyses, and revisions to address the reviewers' concerns:

  • Clinical Meaning and Expert-Annotated Waveform Validation: We added clinical semantic and functional evaluations across 17 ECG findings, with Event Keys achieving AUROC = 0.921 and finding-specific intervention effects across all 17 findings. Independent evaluation against expert-delineated P/QRS/T regions on LUDB further confirms meaningful waveform localization (QRS AP = 0.61, versus 0.10/0.09 under temporal controls).

  • Cross-Recording Event Semantic Consistency: We introduced patient-disjoint PTB-XL retrieval experiments to assess whether learned Events capture clinically consistent information across recordings. Semantic Keys achieve Jaccard@5 = 0.551 and Lift@5 = 2.326, outperforming geometry-only, initial-value, pooled-Key, and random controls. Finding-conditioned retrieval further demonstrates consistency across diagnostic, form, and rhythm concepts.

  • Source-Level Provenance Faithfulness: We clarified the REC lineage construction and added controlled interventions on both pre-REC source Events and original ECG waveform regions. Removing attributed source Events produces a substantially greater finding-NLL increase than matched controls (paired ΔNLL = 0.2471, 95% CI [0.2294, 0.2669]). Directly attenuating attributed waveform regions yields a similarly significant effect (paired ΔNLL = 0.0671, 95% CI [0.0611, 0.0732]). These results provide direct functional evidence that lineage-attributed source Events and waveform regions contribute to finding predictions.

  • Novelty and Same-Backbone Representation Comparisons: We clarified the distinction from fine-grained ECG–text alignment methods, particularly FOCAL (formerly FG-CLEP), emphasizing relational Event composition and preserved source lineage. We further introduced controlled comparisons using the same Qwen3-8B decoder with Dense-Context, text-only Event-REC, and FOCAL representations. ECG-EGR achieves higher Finding F1 than adapted FOCAL on both MIMIC-IV-ECG (0.839 vs. 0.659) and out-of-domain PTB-XL (0.710 vs. 0.598). REC ablation additionally confirms the contribution of relational composition beyond Event discovery and semantic alignment.

  • Adaptive Event Selection and Report-Generation Robustness: We added controlled Adaptive-K versus Fixed-K comparisons, demonstrating improvements in Finding F1 (0.8392 vs. 0.8269) and BLEU-4 (0.7341 vs. 0.6896) while retaining fewer Events. We also provided qualitative report comparisons, report-length-stratified and format-normalized evaluations, and additional backbone experiments using Qwen3-4B and Llama 3.1-8B, supporting the effectiveness of ECG-EGR across different model sizes and families.

We are grateful for the reviewers’ feedback, which has helped us further strengthen the paper. We hope that the additional experiments, analyses, and clarifications address the remaining concerns, and we welcome any further feedback. If the reviewers find these responses satisfactory, we would sincerely appreciate their support.

Warm regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.