List of Papers Browse by Subject Areas Author List
Abstract
Multimodal large language models (MLLMs) have significantly advanced the automatic generation of electrocardiogram (ECG) reports. However, current approaches handle diagnosis as a simple signal-to-text mapping task, overlook clinical reasoning, and thus suffer from hallucination and poor interpretability.
We propose an evidence-mediated ECG report generation and evaluation (EmERGE) approach that reframes the task as structured causal reasoning. It incorporates six clinically validated ECG features, such as rhythms and rates, P waves, and P-R intervals. These features act as explicit mediators, enabling MLLMs to perform diagnostic conditional reasoning using interpretable electrophysiological evidence.
Experiments demonstrate that EmERGE substantially improves ECG report quality, outperforming state-of-the-art grounded report generation methods by 13\% in clinical efficacy and 9\% in interpretability.
To verify whether diagnostic conclusions rely on causal evidence, we developed a rule-guided counterfactual evaluation protocol. The protocol applies signal-level interventions to clinically relevant leads and waveform segments to validate the link between selective evidence and diagnostic outcomes. EmERGE establishes an evidence-driven, causality-focused approach, providing a viable solution for real-world clinical applications.
The source code and prompts are available at \url{https://github.com/hustselab511/EmERGE.git}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4833_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hustselab511/EmERGE.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YaoKey_EmERGE_MICCAI2026,
author = { Yao, Keyu AND Liu, Yuchen AND Wang, Xiaohan AND Shen, Gang},
title = { { EmERGE: A Multimodal LLM-Based Evidence-Mediated ECG Report Generation and Evaluation Approach } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The manuscript proposes EmERGE, an evidence-mediated ECG report generation framework that reformulates ECG interpretation as a two-step reasoning process rather than a direct signal-to-text mapping task. The central idea is to introduce structured electrophysiological evidence as an explicit intermediate representation between multimodal ECG input and the final report. Its main contributions are: 1) Evidence-mediated report generation: six clinically defined ECG evidence dimensions (e. g. , rhythm/rate, P wave, PR interval, QRS complex, ST segment, T wave/QT interval) are extracted first and then used to condition diagnostic report generation. 2) Clinically motivated causal formulation: the method explicitly models diagnosis as being mediated by interpretable ECG evidence, which improves transparency relative to prior end-to-end MLLM approaches. 3) Rule-guided counterfactual evaluation: the paper introduces signal-level perturbation tests designed to assess whether diagnostic outputs change selectively when evidence-relevant waveform regions are modified. 4) Evaluation with two strong backbones: the framework is instantiated with PULSE and GEM backbones (EmERGE(P), EmERGE(G)) and shows improved performance on both in-domain (MIMIC-IV) and out-of-domain (PTB-XL) settings.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) Clinically meaningful problem formulation: The paper addresses an important weakness of existing ECG report generation systems: they may produce fluent reports without demonstrating that the conclusions are grounded in electrophysiological evidence. Reframing the task around explicit evidence extraction is well motivated and clinically sensible. 2) Interpretable intermediate representation: The use of six structured ECG evidence dimensions is a strong design choice. It aligns with how cardiologists reason about ECGs and provides a more auditable interface than direct end-to-end text generation. 3) Interesting evaluation beyond surface-level text metrics: The rule-guided counterfactual protocol is a meaningful addition. In particular, evaluating Targeted Sensitivity, Non-target Invariance, and Diagnosis Response Consistency goes beyond standard report-generation metrics and attempts to probe whether the model’s diagnostic behavior is causally aligned with waveform evidence. 4) Consistent empirical gains across multiple protocols: The method improves over strong baselines not only on LLM-based clinical metrics (e. g. , AccD, Rel, Fid), but also on classification-oriented metrics (macro/micro F1) and counterfactual consistency measures. This makes the overall empirical case stronger than if only one metric family had improved. 5) Cross-backbone applicability within the tested setting: The framework is evaluated with two representative backbones, PULSE and GEM, and yields gains in both cases. This suggests that the approach is not narrowly tied to one specific model implementation. 6) Clinical relevance: Improving evidence grounding and interpretability is highly relevant for eventual real-world deployment of ECG reporting systems, where unsupported but plausible-sounding conclusions are particularly problematic.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) Methodological novelty is stronger in formulation than in implementation: While the use of structured ECG evidence as an intermediate representation is novel and meaningful in this application domain, the underlying machinery—multi-round prompting, staged generation, and conditioning on intermediate text—is still relatively close to existing multi-step LLM reasoning patterns. The manuscript would be stronger if it more clearly distinguished what is conceptually new from what is inherited from general LLM prompting practice. 2) Dependence on LLM-generated supervision: The structured evidence annotations are partially constructed using GPT-4o together with machine measurements. This is practical, but the paper does not sufficiently analyse how noisy or biased these pseudo-labels may be, nor how sensitive the final system is to errors in this supervision. 3) Counterfactual perturbation realism is not fully established: The counterfactual evaluation is one of the most interesting parts of the paper, but also one of the least fully validated. It is not entirely clear whether the proposed signal-level interventions always correspond to physiologically realistic ECG changes, and therefore how strongly one can interpret these results as evidence of true causal validity. 4) Evaluation emphasis remains skewed toward LLM-derived scores: Although the paper reports classification and counterfactual metrics in addition to LLM-based clinical evaluation, the headline improvements are primarily emphasized through LLM-derived scores. A more balanced presentation of the non-LLM evaluation results would strengthen the empirical claims. 5) Limited failure-mode analysis: The qualitative examples are helpful, but the manuscript does not provide a sufficiently systematic breakdown of where the method still fails. For example, it would be useful to disentangle errors caused by incorrect evidence extraction from those caused by evidence-to-diagnosis reasoning. 6) Reproducibility details are incomplete at the current stage: Although the authors state that code will be released after acceptance, the manuscript does not disclose important implementation details such as the exact prompt templates, evidence-extraction criteria, or thresholding decisions used in constructing the six-dimensional evidence space. These omissions make exact replication difficult. 7) Reference formatting is inconsistent and should be carefully audited: Several citations appear to correspond to recent preprints, but they are not consistently identified as such, and some entries lack clear bibliographic details such as venue information or arXiv identifiers. This does not affect the core technical contribution, but it reduces clarity and makes it harder to assess the status and maturity of the cited prior work. The authors should carefully revise the bibliography and clearly distinguish peer-reviewed publications from preprints.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The paper tackles an important and timely problem in medical report generation, and the evidence-mediated perspective is a promising direction. A few suggestions that could further strengthen the work: 1) Clarify more explicitly what part of the method is genuinely novel relative to existing multi-step LLM prompting frameworks, as opposed to being a careful domain adaptation of those ideas. 2) Provide a more detailed justification of the signal-level counterfactual intervention design, ideally including discussion of physiological plausibility and boundary cases where such perturbations may not correspond to realistic pathology changes. 3) Expand the analysis of failure modes, for example by separately reporting cases where the evidence extraction is correct but the final reasoning is wrong, versus the reverse. 4) Improve the bibliography by ensuring that all preprints are properly labeled and that all references contain complete and accurate publication information.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The manuscript presents a well-motivated and clinically relevant approach to ECG report generation. Its main strength lies in introducing explicit structured evidence as an intermediate reasoning layer, which improves interpretability and leads to consistent gains across multiple evaluation settings. The counterfactual evaluation framework is also a valuable contribution, as it moves beyond standard report-generation metrics and attempts to assess whether diagnostic behavior is selectively responsive to waveform evidence. At the same time, the work has several limitations. The implementation is somewhat incremental from a general LLM-reasoning perspective, the supervision pipeline depends partly on LLM-generated evidence annotations, and the realism of the counterfactual perturbations is not yet fully established. In addition, while the empirical results are strong, many of the most prominent gains are reported through LLM-derived clinical metrics, and the manuscript would benefit from broader objective or expert-based validation. Overall, I place the paper marginally above the acceptance threshold. The clinical motivation is strong, the empirical evidence is consistently positive, and the counterfactual analysis introduces an interesting discussion point for the community. However, the paper would benefit from clarification in rebuttal regarding reproducibility details, evaluation balance, and the validity of the proposed counterfactual design.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes an evidence-mediated ECG report generation framework that introduces a structured intermediate representation (six ECG phenotypes) between input signals and generated reports. The design aims to enable interpretable, evidence-driven report generation from ECG with reasoning based on ECG phenotypes.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper evaluates the model across multiple reasonable and clinically relevant scenarios and tasks, comparing it with other baselines. 2.If released, the GPT-4o-generated evidence/reasoning data could be useful for the community. 3.The design and implementation of the counterfactual evaluation are well-motivated and technically sound, apart from some details mentioned below.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.While there are many works aiming to generate ECG reports from signals, and this work proposes a structured reasoning approach, I do not fully understand the motivation. In clinical practice, ECG reports are typically generated by rule-based systems embedded in ECG machines. Since the model is trained on these reports as ground truth, it effectively learns to approximate them. What is the advantage of this approach compared to existing rule-based systems, which are simpler and do not hallucinate?
2.In Section 2.2, please define what h(⋅) represents. Is it an LLM?
3.It would be much clearer if the authors explicitly showed how the structured evidence E looks. Is each Ei a single label (Normal / Abnormal / Borderline), or a richer textual description? Also, how are these labels assigned to each phenotype? What cutoffs are used? Following this, why is GPT-4o used to construct E* if the representation is structured and could potentially be formed manually with regex?
4.How were the 12,000 MIMIC-IV and 6,000 PTB-XL samples selected (randomly?), and what is their class distribution? Could happen that most cases are normal. I guess using a smaller sample size was due to GPT-4o API limits and costs, but this can be avoided.
5.In Section 2.2, both Le and Lr appear to be cross-entropy losses, but different notations are used (CE vs. −logp). Why is this inconsistent? Also, the formula formatting is broken. More importantly, is Lr conditioned on predicted E or ground-truth E* during training?
6.For the counterfactual analysis, the perturbations are not clearly described. What exactly are “relevant” and “irrelevant” perturbations? How are they applied? For the DRC metric, what representation or encoder is used to compute delta(D) and delta(E)?
7.Why is a single projector used for both ECG signals and images? Since these are different modalities, a separate design would seem more natural.
8.The implementation details are insufficient. Which LLM is used? Which parts are trained or frozen? Is full fine-tuning performed? How is catastrophic forgetting addressed? Referring to another paper for core details makes reproduction difficult.
9.Evaluation details are unclear. For LLM-based evaluation: which LLM is used, and what are the exact inputs? Why do the reported metrics differ from those in the GEM paper?
10.For CE-based diagnostic accuracy: which labels are used, and how are they extracted from generated reports (e.g., rule-based or regex-based)?
11.In Figure 3, it would be helpful to show the actual ECG signal (e.g., Lead II) for better interpretation.
12.The core idea is similar to reasoning with LLMs. Why not incorporate reasoning directly into supervision and train the model in a single stage instead of a two-step pipeline?
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Thanks to the authors for the submitted work.
Overall, I find the paper’s motivation insufficiently justified. The practical value of learning to generate ECG reports is unclear, given that ECG machines already reliably produce them.
In addition, many implementation details are omitted and referred to another paper. Several design choices are not clearly motivated, for example, how the proposed evidence-mediated framework differs from standard reasoning approaches in LLMs, and why a two-step pipeline is preferred over directly incorporating reasoning into supervision.
This decision reflects my current understanding of the manuscript. I would be open to revising my assessment if the authors clarify the points raised and address the questions above.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes EmERGE, a framework for automated ECG report generation. The core idea is simple. Instead of letting the model jump straight from the ECG signal to a diagnosis, the authors add a middle step. The model first reads the signal, extracts clinical evidence, and then writes the diagnosis. So the flow goes from signal, to evidence, to diagnosis. To make this work, the authors use six ECG dimensions that cardiologists actually look at: rhythm and rate, P wave, PR interval, QRS complex, ST segment, and T wave with QT interval. The model must reason through these six dimensions, which makes its thought process easier to follow. The authors frame this whole pipeline as structured causal inference, where the evidence acts as a causal mediator between the signal and the diagnosis. They also propose a new evaluation protocol. It is a rule-guided counterfactual test that checks whether the model really depends on the intended evidence instead of taking shortcuts. It uses three metrics: targeted sensitivity, non-target invariance, and diagnosis response consistency. The results look solid. EmERGE improves LLM-based metrics on benchmark datasets, for example AccD 79.96 percent versus 74.23 percent for GEM. It is backbone-agnostic, working on top of both PULSE and GEM, generalizes well on PTB-XL, and is also evaluated by expert cardiologists.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Here are the major strengths I see in this paper. [S1] The six evidence dimensions are clinically well grounded. They follow the way cardiologists actually read an ECG, going through rhythm, P wave, PR interval, QRS, ST segment, and T wave with QT. Because of this, the intermediate representation is not an arbitrary engineering choice. It reflects real clinical practice, which makes the whole framework easier to trust from a clinician’s point of view. [S2] The causal framing is not just narrated, it is enforced by the design. The model has to produce evidence first in Step 1, and the diagnosis is only allowed in Step 2.The training loss also supervises the evidence before the report. This structurally blocks a common failure mode where the model decides the diagnosis first and then invents matching evidence. This is a small but important design decision. [S3] The counterfactual evaluation is the most original part of the paper. Most ECG report generation works only check text or label accuracy, so it is hard to tell if the model is really using the ECG evidence. Here the authors change specific parts of the signal and watch how the model reacts. They look at three things, targeted sensitivity, non-target invariance, and diagnosis response consistency. The LAFB experiment is especially convincing. When the key evidence is removed, EmERGE drops the diagnosis, while the baseline mostly does not.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
[W1] The automated numbers and the human scores do not match. The abstract claims big improvements, and Table 1 looks good on LLM-based metrics. But in Table 6, EmERGE scores 3.92 on Overall Satisfaction, while GPT-4o scores 4.38 and DeepSeek-R1 scores 4.44.The same pattern shows up in Analytical Relevance, Analytical Accuracy, and Reasoning Quality. So on every item rated by real cardiologists, EmERGE is below the two LLM baselines. The text only says cardiologists rated EmERGE positively, and never mentions this gap. The setup is also a bit confusing. Section 3.5 says GPT-4o and DeepSeek-R1 were used as ground truth, but the same two models still appear with their own scores in Table 6.Why did the authors not discuss this gap or clarify the evaluation setup? [W2] The six evidence labels are made by GPT-4o and featureDB, and the whole method is trained on these labels. But no cardiologist ever checks whether the labels are actually correct. There is no verified subset and no agreement score with human experts. GPT-4o is a general language model, not an ECG specialist, and featureDB gives machine-based features. If the labels are off in some systematic way, every claim about causal grounding is also off. The authors mention this in one line at the end, but the whole method rests on this supervision. Why did the authors not include a cardiologist-verified subset or report inter-rater agreement?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The counterfactual evaluation protocol is, to my knowledge, a genuinely new way to probe whether ECG report generators respond to the evidence they claim to use, and I think it is the paper’s most reusable contribution. The six-dimensional clinical evidence schema is sensibly aligned with AHA standards, and the backbone-agnostic and cross-domain results give reasonable confidence that the framework is not tightly specialized to one setup. What keeps this from a clear Accept is the gap between the automated-metric story and the only human evaluation in the paper. Table 6 places EmERGE below both GPT-4o and DeepSeek-R1 across every dimension, including Reasoning Quality and Overall Satisfaction, and the text does not acknowledge or explain this. The evidence supervision is also produced entirely by GPT-4o and featureDB without any cardiologist-verified subset or inter-rater agreement, which weakens the core claim that the model is grounded in clinical evidence rather than in another model’s outputs. These are not fatal issues, but they are central enough to the paper’s own thesis that I cannot place this above Weak Accept without stronger clarification or validation. Overall, the novel evaluation protocol and consistent methodological gains justify a recommendation above the rejection line, while the unresolved human-evaluation and supervision-integrity concerns keep the score on the weaker side of acceptance. The pull in the other direction is the expert-rating inversion in Table 6 — EmERGE rates lower than GPT-4o and DeepSeek-R1 on cardiologist scoring, which directly pushes against the “13% clinical efficacy” headline. Together with the lack of any validation of the GPT-4o + featureDB evidence supervision, this leaves the clinical-grounding claim less defended than the quantitative claims.
Overall I lean Weak Accept. The methodological contribution (counterfactual protocol) carries the recommendation; the clinical-claim calibration and the supervision-validation question are the main items I would want addressed in the final version.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank the Meta Reviewer and all reviewers for your professional comments. Below, we shall clarify your key concerns.
1.Research Motivation (R2.1) Traditional rule-based ECG methods are simple to implement and can generate plausible reports in routine cases. However, they rely heavily on predefined rules and are sensitive to signal quality. They fail in cases with severe waveform noise or complex comorbidities. To address these limitations, data-driven deep learning methods have been developed for ECG report generation (please see DOI: 10.1038/s41467-020-15432-4 for reference). They outperform conventional rule-based baselines consistently. EmERGE targets grounded reports by linking waveform patterns to diagnostic conclusions via interpretable evidence. Experimental results further verify that our evidence-driven approach surpasses end-to-end methods, achieving a 13% improvement in the CE metric.
2.Supervision and Expert Validation (R1.2, 1.4, 3.1, 3.2) We clarify potential misunderstandings in experimental settings. Our framework builds upon LLM-based generation and evaluation; thus, we conduct human expert assessment in Table 6.GPT-4o and DeepSeek-R1 are not used as comparative baselines. They serve only to validate the quality of silver labels adopted for EmERGE training. Strong performance of GPT-4o confirms that our supervisory signals are clinically reliable and free of systematic bias. Intermediate features extracted by FeatureDB are derived from validated rule-based beat-level algorithms in prior studies. The LLM acts merely as a reasoning engine. Nevertheless, Table 6 provides solid indirect evidence to support the reliability of our framework.
3.Counterfactual Intervention (R1.3, 2.6) Due to page limits, the initial submission lacks complete elaboration on rule-based interventions, which may cause ambiguity. We will supplement more details in the camera-ready version. Our rule-based interventions guarantee clinical plausibility. Rules are defined as structured operations on specific leads and segments toward predefined clinical goals. We define two intervention types: Targeted Intervention: expected to alter diagnostic conclusions; Non-Target Intervention: expected to preserve original diagnoses. The intervention replaces segmented waveforms with normal templates at the beat level, followed by smoothing and cropping. All rules are formulated based on official AHA and Philips guidelines. Interventions are deterministically executed via Neurokit2 at the signal level, effectively mitigating LLM hallucinations. This protocol examines whether the model truly relies on clinical evidence for diagnosis. We acknowledge that real-world distribution and clinical alignment require further validation, which we will in our future work.
4.Dataset Configuration (R2.4) For fair comparison with baseline methods, we adopt the same data split ratio as GEM. We select a smaller training set for few-shot learning and verify our structural evidence-driven approach. By explicitly grounding report generation on structured clinical evidence, our method outperforms GEM with far less training data (18k vs. 120k). This outcome demonstrates that our evidence-driven paradigm extracts denser and more effective supervision signals than vanilla end-to-end learning. We will continue to adopt larger datasets in future optimizations.
5.Definition of Six-Dimensional Features (R1.6, 2.3) Following AHA guidelines, we define six feature dimensions covering rhythm, P wave, PR interval, QRS complex, ST segment, and T wave & QT interval. Each dimension employs semi-structured natural language descriptions. Every entry contains a categorical label (Normal, Abnormal, or Borderline) and a concise textual characterization, without involving final diagnostic judgments.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This is an quite novel study to reformulate ECG interpretation as a two-step reasoning process to make diagnosis with explicit clinical evidence.