We appreciate your recognition of the innovation and contributions in our work! We are glad that you found the paper novel and the studied problem is important. We address your concerns point by point as follows.

[Weakness 1] Parameter-efficient fine-tuning may limit model adapting to new grounding task

Thank you for raising this important point. We agree that unfreezing the encoders could potentially improve the model’s ability to adapt feature extraction to the grounding task, particularly in a multimodal setting where fine-grained cross-modal alignment is essential.

However, fully training all encoders incurs substantial computational cost, as it requires training two additional Transformer encoders. For this reason, we designed a cross-modal alignment learning module and adopted a widely used parameter-efficient training strategy (e.g., as in LLaVA). This design choice significantly reduced training cost and memory usage while maintaining good performance.

We appreciate this suggestion and see end-to-end fine-tuning as a valuable next step. We plan to include this experiment in the revised version to better understand its impact on grounding performance.

[Weakness 2] Data quality evaluation with human experts

We conduct a quantitative human evaluation on 200 unique ECG cases, randomly sampled from ECG-Grounding generated by GPT-4o.

The evaluation is performed by eight board-certified cardiologists, each independently assessing a subset of cases using seven predefined criteria designed to measure both reliability and clinical usefulness.

Table 1 summarizes the results. The expert evaluation shows that GPT-4o consistently achieves high scores across both reliability and clinical usefulness. These findings indicate that, under our knowledge-guided instruction data generation, GPT-4o can produce ECG interpretations that are both clinically reliable and practically valuable.

Table 1: Human Evaluation on GPT-4o Generated Training Data

Reliability metrics

Usefulness metrics

Scoring criteria (Note: Due to space limitations, we only present scores 4–5 here. The full scoring criteria (including 1–3) will be provided in the updated version):

Analytical Relevance: Do the model’s analyses closely support the diagnosis, and is there corresponding ECG evidence?
5: Every analysis point is highly relevant to the diagnosis, with clear supporting evidence.
4: Most analyses are strongly relevant, with minor insufficiencies.

Analytical Accuracy: Are there any medical factual errors in the model's output?
5: Completely accurate
4: Mostly accurate

Analytical Completeness: Does the model comprehensively discuss key ECG components relevant to the diagnosis, including rhythm, intervals, and waveforms?
5: All relevant ECG features (rhythm, PR, QRS, ST, T waves, intervals, etc.) are accurately discussed.
4: Most key ECG features are covered, with minor omissions.

Reasoning Quality: Does the model provide a clear, evidence-based reasoning process similar to that of a clinician, logically deriving the diagnosis from ECG features?
5: Clear and coherent reasoning structure, explaining each step from ECG to diagnosis causally.
4: Overall reasonable reasoning, but some steps lack detail.

Findings Novelty: Does the model provide insights or findings not noticed by the clinician?
5: Important new diagnoses or findings.
4: Novel and somewhat insightful content.
3: Some new findings, but of limited value.

Clinical Value: Does the model output help in clinical decision-making?
5: Direct and significant support for clinical judgment; content is clear and reliable.
4: Most content is helpful and practically useful.

Overall Satisfaction: Subjective rating of the overall quality of this analysis.
5: Very satisfied.
4: Satisfied

[Weakness 3] IMG-Only baseline and performance gain from adding images

Thank you for the insightful suggestion. We did not train a separate IMG-only variant of GEM since PULSE can be seen as the IMG-only baseline, as it shares the same configuration used in our ablation study.

The only difference is that PULSE was trained for three epochs, whereas our ablation variants were trained for one epoch due to resource constraints. Despite this, the TS+IMG still outperforms PULSE in most tasks, suggesting that adding modalities improves both training efficiency and performance, and that the two modalities are complementary. We will clarify this point in the revised version.

In addition to performance, we believe that adding the image modality also improves usability, for example, by making it easier for users to interact with the system using a photo of their ECG.

[Weakness 4] Failure case analysis

Thank you for this insightful and valuable comment. We conduct a systematic analysis of failure cases, particularly informed by the expert feedback collected during the human evaluation. Based on these observations, we categorize the major failure types and analyze potential causes as follows:

• Incorrect diagnosis

  These errors may result from limitations in the representation stage. For example, certain morphological features such as subtle ST-segment changes or P-wave abnormalities were sometimes missed by the encoders.

  Potential causes include:

  • Representation deficiency: Some features were not captured due to encoder limitations.

  • Training data limitations: The current training set may not cover the full spectrum of ECG variations, especially when data distribution shifts.

• Overstating the severity of findings

  GEM occasionally exaggerates the severity of certain cardiac conditions, which some cardiologists noted could lead to unnecessary patient anxiety.

  Potential causes include:

  • Limited context on patient history: In real-world settings, physicians contextualize ECG findings using patient history, which GEM does not have access to, potentially leading to overdiagnosis in isolation.

In the revised version, we will include this failure case analysis to provide a more systematic view of where errors arise. These insights will not only contextualize current performance but also inform future directions for improving model safety, reliability, and alignment with clinical practice.

[Question 1] Robustness of Diagnosis Guider and how to handle rare or complex cardiac conditions

Thank you for this important question regarding the robustness of the Diagnosis Guider. We address your concerns as follows:

First, rule-based interpretation is the foundation of clinical ECG analysis in current medical practice. For example, the widely adopted AHA/ACCF/HRS guidlines [1] provide structured diagnostic criteria for arrhythmias, conduction delays, and ischemic changes. Commercial ECG systems such as GE MUSE [2] and Philips DXL [3] also rely on deterministic rule sets for generating diagnostic statements. Our Diagnosis Guider is developed with reference to standard clinical guidelines, aiming to analyze ECGs in a way aligned with cardiology practice.

Second, we acknowledge that rule-based systems, like clinical interpretation itself, can be fallible, especially in rare disease or complex conditions. To improve the overall robustness, our framework does not rely solely on the Diagnosis Guider. It contains learning-based components that can capture broader data-driven patterns. This hybrid design enhances the overall robustness compared to standalone rule-based interpretation. Furthermore, our human evaluation shows that cardiologists found Diagnosis Guider outputs clinically reliable and useful. Together, these results indicate that the system is capable of handling a wide range of ECG interpretation scenarios with performance approaching clinical applicability.

Third, failure modes may still arise, particularly for rare, under-represented, or novel cardiac conditions that fall outside both the current rule base and training data distribution. While we recognize that the system is not yet perfect, the overall performance, both in benchmark evaluations and in expert review, indicates that it has reached a level that helpful in many settings and can provide practical clinical value.

In the future, we plan to further strengthen the Diagnosis Guider by expanding the rule set with expert input, incorporating more diverse training data, and iteratively refining the framework to improve its coverage and adaptability over time.

[1] Rautaharju, Pentti M., Borys Surawicz, and Leonard S. Gettes. "AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram." Journal of the American College of Cardiology 53.11 (2009): 982-991.

[2] GE HealthCare. (2019). Implementing MUSE to Streamline ECG Workflow Case Study.

[3] Philips DXL ECG Algorithm Physician’s Guide.

[Question 2] Explanation for SFT LLaVA’s performance compared to SFT PULSE

This finding can be explained by the fact that PULSE was not pretrained on any high-granularity grounding data. As a result, PULSE lacks the necessary capabilities for grounded ECG interpretation. The knowledge embedded during its pretraining provides limited support for the evidence-based grounding task we propose.

Thank you very much for your thoughtful comments! We are glad you recognized the value of our method and appreciated the open-source contributions in our work. We respond to your comments in detail below.

[Weakness 1, Question 1] Generalization beyond ECG

We appreciate the reviewer’s suggestion to consider generalization beyond ECG. While our work is intentionally focused on ECG, a modality that is not only foundational in cardiovascular diagnostics but also highly accessible and widely adopted across care settings, we agree that exploring generalizability is valuable.

Following your suggestion, we conduct experiments extending our methods to respiratory spirogram data for pulmonary disease detection. The dataset was extracted from UK Biobank field 3066. As shown in the Table 1 below, our approach outperforms both deep learning SOTA and LLaMA3.1-8B across multiple metrics. These results suggest that our methods are not inherently limited to ECG and hold promise for broader clinical modalities.

While these results demonstrate the potential generalizability of our framework, this paper is intentionally scoped to ECG interpretation. All modeling components and instruction data were designed around the specific structure and diagnostic reasoning challenges associated with ECG data. We believe ECG offers a rich and clinically meaningful testbed for studying multimodal reasoning in depth.

In the revised version, we will add a discussion on how our framework could be adapted to other medical modalities. This primarily involves adjusting the knowledge-guided instruction data generation pipeline to account for modality-specific representations and clinical reasoning structures.

Table 1: Experiment results on respiratory spirogram data for pulmonary disease detection

[Weakness 2, Question 2] Dependence on GPT-4o generated training data

Thank you for raising this important point.

To fully address your concern, we conduct a structured human evaluation covering three sources of outputs: GPT-4o generated data, Deepseek-R1 generated data, and GEM generated interpretations.

In total, 400 ECG-Grounding data (200 from GPT-4o and 200 from Deepseek-R1) and 200 GEM's interpretation were independently reviewed by eight board-certified cardiologists, using seven predefined clinical criteria designed to assess both reliability and usefulness. This unified evaluation protocol allows us to (1) verify the quality of GPT-4o generated training data, (2) test the effectiveness of open-source substitutes, and (3) confirm the clinical utility of the GEM model.

We report the evaluation results in the Table 2 below. The expert evaluation shows that GPT-4o consistently achieves high scores across both reliability and clinical usefulness, with particularly strong performance in analytical completeness and reasoning quality. These results demonstrate that, when using our knowledge-guided instruction data generation, GPT-4o is capable of generating high-quality ECG interpretations that are both clinically reliable and practically valuable.

Deepseek-R1 is also capable of generating clinically acceptable, high-quality ECG interpretations with our methods. This demonstrates that our method is adaptable to alternative LLM backbones and remains applicable in settings without commercial API access. We will include these results in the revised version of the paper.

Due to time and resource constraints, we have not yet completed GEM training and evaluation on Deepseek-R1 generated data. We plan to include this experiment in the revised paper to quantify the performance changes.

Table 2: Human Evaluation on GPT-4o/Deepseek-R1 Generated Training Data and GEM's Interpretation

Reliability metrics

Usefulness metrics

Scoring criteria (Note: Due to space limitations, we only present scores 4–5 here. The full scoring criteria (including 1–3) will be provided in the updated version):

Analytical Relevance: Do the model’s analyses closely support the diagnosis, and is there corresponding ECG evidence?
5: Every analysis point is highly relevant to the diagnosis, with clear supporting evidence.
4: Most analyses are strongly relevant, with minor insufficiencies.

Analytical Accuracy: Are there any medical factual errors in the model's output?
5: Completely accurate
4: Mostly accurate

Analytical Completeness: Does the model comprehensively discuss key ECG components relevant to the diagnosis, including rhythm, intervals, and waveforms?
5: All relevant ECG features (rhythm, PR, QRS, ST, T waves, intervals, etc.) are accurately discussed.
4: Most key ECG features are covered, with minor omissions.

Reasoning Quality: Does the model provide a clear, evidence-based reasoning process similar to that of a clinician, logically deriving the diagnosis from ECG features?
5: Clear and coherent reasoning structure, explaining each step from ECG to diagnosis causally.
4: Overall reasonable reasoning, but some steps lack detail.

Findings Novelty: Does the model provide insights or findings not noticed by the clinician?
5: Important new diagnoses or findings.
4: Novel and somewhat insightful content.
3: Some new findings, but of limited value.

Clinical Value: Does the model output help in clinical decision-making?
5: Direct and significant support for clinical judgment; content is clear and reliable.
4: Most content is helpful and practically useful.

Overall Satisfaction: Subjective rating of the overall quality of this analysis.
5: Very satisfied.
4: Satisfied

[Question 3] Validation of GPT-4o's clinical reasoning depth

Directly using GPT-4o for board-style ECG interpretation tasks that require specialist-level reasoning typically yields suboptimal results. As shown in Table 3 of our paper, when prompted to generate free-text clinical reports on the PTB-XL dataset, GPT-4o achieves a report score of only 50.2, significantly lower than our method (67.1). This task requiring comprehensive interpretation and is close to to board-style questions. Moreover, GPT-4o also performs poorly on multiple-choice diagnostic tasks (as shown in main paper Table 2 ), highlighting its difficulty in making correct diagnosis even under simplified settings. This further underscores the challenge it faces in generating accurate and complete reasoning in free-text ECG interpretation.

This is the reason we designed the knowledge-guided instruction data generation, where the GPT-4o is used as a controlled generation tool to construct training data, guided by ground-truth diagnostic reports and heartbeat-level descriptions. With our method, GPT-4o is able to generate high-quality reasoning data. This is supported by human evaluation on GPT-4o outputs, where cardiologists judged the reasoning to be clinically acceptable in most cases.

[Limitation 1] Knowledge inheritance and error propagation risk

We appreciate the reviewer’s insightful comment regarding the risk of inherited errors from GPT-4o.

We agree that if GPT-4o’s factual understanding or clinical knowledge is flawed, these errors could propagate into the generated training data. To mitigate this risk, we designed our data generation pipeline to anchor GPT-4o’s outputs to ground-truth clinical labels. Specifically, GPT-4o is prompted to explain or elaborate on known diagnoses and lead-level features from expert-labeled datasets such as MIMIC-IV-ECG. This constrains the model to generate reasoning that reflects verified clinical diagnosis, rather than freely recalling potentially unreliable knowledge.

Our human evaluation on both GEM’s interpretation and GPT-4o generated training data verified that our data generation pipeline help substantially mitigate the risk of major factual inaccuracies. In future work, we plan to integrate clinical validation tools and expert feedback loops to further improve the reliability and safety of model-generated content.

We will explicitly include this discussion in the revised manuscript.

Dear Authors,

Thank you for the thorough and constructive rebuttal. I appreciate the additional experiments and clarifications. Below I summarize my current assessment and list a few points where further detail would strengthen the revision.

1. Generalizability beyond ECG

Commendation

Your new spirogram experiment is a valuable addition and demonstrates that the proposed architecture is not intrinsically tied to ECG signals.

Clarification requests

1. Training protocol – Was the same ECG‑trained model fine‑tuned on spirograms, or did you train a fresh model from scratch using the spirogram corpus?

2. Data scale and diversity – How many spirogram records and diagnostic labels were used, and how does their clinical and linguistic diversity compare with your ECG datasets (e.g., number of unique diagnoses, class balance, range of signal morphologies)?

3. Prompt / instruction adaptation – Did you modify the knowledge‑guided instruction templates for spirograms? If so, please provide two or three concrete examples.

4. Interpretation of results – Given that spirogram signals typically exhibit lower morphological variability than ECGs, could you elaborate on what aspects of the architecture or prompting strategy you believe enabled the observed performance gains?

Including these details (even in an appendix) will help readers reproduce and interpret the generalization experiment.

2. Dependence on GPT‑4o‑generated data

Your human evaluation comparing GPT‑4o and DeepSeek‑R1 outputs is welcome. To maximise reproducibility and future benchmarking:

• Would it be feasible to release the DeepSeek‑R1‑generated corpus and your cardiologist‑graded labels? Even a de‑identified subset would allow independent verification.

• For transparency, please report inter‑rater reliability statistics (e.g., Fleiss’ κ) for the eight cardiologists across the seven clinical criteria.

3. Validation of GPT‑4o’s clinical reasoning

Your strategy of anchoring GPT‑4o outputs to ground‑truth labels is sensible. Have you catalogued any systematic errors that still leak through? Even a short qualitative summary would inform practitioners about residual risks.

Overall, the rebuttal addresses the majority of concerns, and I am encouraged by the additional evidence supplied. Clarifying the points above will, in my view, make the final manuscript even more impactful and reproducible.

Interpretation of results

We believe that two factors collectively contribute to the observed performance gains:

(1) Reliable extraction of core morphological features:

In Table 1, our method employs the Deep Learning SOTA as the encoder to capture morphological features from the spirogram curves, and uses Llama 3.1-8B as the base LLM to process the complementary textual inputs.

The comparison between the Deep Learning SOTA and Llama3.1-8B highlights the importance of robust morphological feature extraction for accurate spirogram interpretation.

(2) Incorporation of domain-specific diagnostic logic aligned with clinical guidelines:

The comparison between the Deep Learning SOTA and the full model demonstrates the importance of incorporating domain-specific diagnostic logic aligned with established clinical guidelines (i.e., knowledge-guided instruction data generation).

These results further confirm the generalizability and robustness of our approach.

Dependence on GPT‑4o‑generated data

Would it be feasible to release the DeepSeek‑R1‑generated corpus and your cardiologist‑graded labels?

Yes, we will release these data after appropriate de-identification.

Inter‑rater reliability statistics.

We appreciate the suggestion to report inter-rater reliability to further imporve the transparency. While Fleiss’ κ is commonly used, it requires that (1) each item is rated by the same number of raters, and (2) the ratings are nominal.

In our human evaluation, each cardiologist assessed a different subset of the 200 cases from a given model (i.e., not all items were rated by the same raters). Additionally, the seven criteria were scored on ordinal 5-point scales rather than categorical labels. Therefore, Fleiss’ κ is not applicable in current setting.

Instead, to ensure transparency, we report in Table 3 the mean and standard deviation of scores across all cardiologists for each clinical criterion.

The mean reflects the overall evaluation of the model's outputs, and the standard deviation serves as a proxy for inter-rater variability (i.e., higher std indicates greater variability across raters). The standard deviation will be added in our revision.

We plan to include the inter-rater reliability statistics such as Krippendorff’s α in future large-scale evaluations where more overlapping annotations are available.

Table 3: Human Evaluation on Generated Data and GEM's Interpretation (Mean and STD)

Reliability metrics

Usefulness metrics

 Validation of GPT‑4o’s clinical reasoning

Have you catalogued any systematic errors that still leak through?

So far, we have not observed any systematic errors in GPT-4o outputs that lead to severe factual inaccuracies or propagate into GEM’s training data under our data generation pipeline.

As discussed in Section 4.4 of the main paper, the most notable deviation we have observed is occasional disagreement between GPT-4o explanations and cardiologist interpretations. These discrepancies may arise in finer-grained feature judgments, such as waveform morphology or interval thresholds.

Such disagreements are not unexpected, as they reflect the inter-observer variability commonly seen in real-world clinical interpretation, even among experienced cardiologists. Importantly, GPT-4o explanations are explicitly conditioned on ground-truth diagnostic labels derived from original clinical reports, which themselves represent expert assessments made at the time of care. As such, GPT-4o outputs reflect one valid expert viewpoint, while the reviewing cardiologist may provide another.

We consider this inter-observer variability inherent to clinical diagnostic tasks, especially in morphology-sensitive settings, rather than a failure of the generative process.

Thank you for your reply! Here are our clarifications on the points you raised.

Generalizability beyond ECG

Training protocol

We trained a fresh model (i.e., no prior domain-specific training) from scratch using the spirogram corpus to ensure a meaningful and modality-specific evaluation. Directly applying the ECG-trained model to spirograms can result in suboptimal and uninterpretable performance due to domain-specific feature dependencies and differences in signal morphology. Training from scratch ensures that the model learns spirogram-specific representations without being biased by ECG-specific priors. Moreover, in practical clinical applications, it is uncommon to deploy a single model to simultaneously interpret both ECGs and spirograms, as each modality requires distinct input structures, clinical features, and reasoning frameworks.

Data scale and diversity

We trained on a cohort of 18,416 patients selected from the UK Biobank (field 3066). The target task is the diagnosis of Chronic Obstructive Pulmonary Disease (COPD), formulated as a binary classification task. The dataset's diagnostic labels are categorized into two groups: confirmed COPD (8,245 patients) and non-COPD (10,171 non-COPD controls). In contrast, our ECG-based tasks are multi-label classification tasks covering a broader set of unique diagnosis.

Compared to ECGs, which consist of high-frequency, multi-phase waveforms with distinct temporal segments (e.g., P wave, QRS complex, T wave), spirogram signals typically exhibit lower-frequency, continuous curves such as flow-volume or volume-time curves. While spirograms appear morphologically smoother, their diagnostic interpretation still requires reasoning over shape characteristics (e.g., concavity, slope changes) and integration of quantitative measurements like FEV1 and FVC.

In this sense, the morphological variability is more subtle but still clinically meaningful, and effective interpretation depends on combining structural features with domain-specific diagnostic logic.

Prompt / instruction adaptation: Did you modify the knowledge‑guided instruction templates for spirograms?

Yes, adapting the knowledge-guided instruction templates to fit the spirogram modality is crucial, as it ensures that the prompts capture modality-specific features and align with the clinical reasoning process for diagnostics.

The overall prompting pipeline remains consistent:

(1) we first extract a morphological description from the spirogram curve (similar to ECG feature extraction).

(2) we incorporate relevant patient-level metadata such as pulmonary function test (PFT) values.

(3) we use the knowledge from the GOLD guidelines to guide the LLM in generating final, high-quality training data.

Below are two examples of the generated training data:

case 1 (non-copd):
    The patient is a 44-year-old female with no history of smoking, presenting for pulmonary function testing. Spirometry results reveal an FEV1 of 2.82 L (91.28% predicted) and an FVC of 3.62 L (94.62% predicted). The post-bronchodilator FEV1/FVC ratio is 0.779, which is above the diagnostic threshold of 0.70 for airflow obstruction as per GOLD criteria. Additionally, neither the FEV1 nor the FVC falls below the lower limit of normal (LLN), with FEV1 at -0.694 z-score and FVC at -0.422 z-score. The FEF25-75 is 2.95 L/s (94.51% predicted), also within normal limits. The spirometry graph description demonstrates a rapid initial rise, a rounded peak expiratory flow, and a gradually descending limb with mild concavity, consistent with normal expiratory flow patterns without significant obstruction. Given these findings, there is no evidence of fixed airflow obstruction. The absence of smoking history further reduces the likelihood of COPD. Therefore, the diagnostic criteria for COPD are not met based on the current pulmonary function testing and clinical context.

case 2 (copd):
    The patient is a 62-year-old male with a significant smoking history, presenting for pulmonary function testing. Spirometry reveals a post-bronchodilator FEV1/FVC ratio of 0.61, which is below the diagnostic threshold of 0.70 as established by GOLD criteria for airflow obstruction. The FEV1 is 74.866% of predicted, indicating moderate airflow limitation. The FVC is within normal limits at 94.411% of predicted. The FEF25-75, a marker of small airway function, is reduced at 55.38% of predicted, further supporting obstructive physiology. The spirometry curve demonstrates an obstructive pattern with a steep initial rise to peak flow followed by a concave descending limb, consistent with airflow limitation. Given the patient's age, smoking history, and spirometric evidence of fixed airflow obstruction not fully reversible with bronchodilators, the diagnosis of COPD is confirmed.

Thank you very much for your encouraging comments on our paper! We are glad that you found that the paper is presented and articulated very clearly. We provide detailed responses to your concerns as follows.

[Weakness 1, Question 1] GPT-4o cost, inference time numbers, scalability

• GPT-4o cost and inference time numbers

  Thank you for raising this important point. In our implementation, generating each sample required approximately 6,000 input tokens and 300 output tokens, resulting in a total cost of roughly USD $3,000 using the OpenAI GPT-4o API for the 300,000 samples in ECG-Grounding dataset. The average inference time per sample is 8s on a single A100 GPU.

• Substituting GPT-4o with an open-source alternative

  We test an open-source substitute model, the latest Deepseek-R1. We conduct a human evaluation on 400 ECG-Grounding data (200 from GPT-4o and 200 from Deepseek-R1).

  The human evaluation is conducted by eight board-certified cardiologists, each independently evaluating a subset using seven predefined criteria that assess both reliability and usefulness in clinical practice.

  As shown in the human evaluation results in Table 1, Deepseek-R1 is also capable of generating clinically acceptable, high-quality ECG interpretations with our methods. This demonstrates that our method is adaptable to alternative LLM backbones and remains applicable in settings without commercial API access. We will include these results in the revised version of the paper.

  Due to time and resource constraints, we have not yet completed GEM training and evaluation on Deepseek-R1 generated data. We plan to include this experiment in the revised paper to quantify the performance changes.

  Table 1: Human Evaluation on GPT-4o/Deepseek-R1 Generated Training Data

  Reliability metrics

  	Analytical Relevance	Analytical Accuracy	Analytical Completeness
  GPT-4o	4.7/5	4.6/5	4.7/5
  Deepseek-R1	4.8/5	4.7/5	4.9/5

  Usefulness metrics

  	Reasoning Quality	Findings Novelty	Clinical Value	Overall Satisfaction
  GPT-4o	4.7/5	4.4/5	4.7/5	4.5/5
  Deepseek-R1	4.8/5	4.5/5	4.6/5	4.7/5

  Scoring criteria (Note: Due to space limitations, we only present scores 4–5 here. The full scoring criteria (including 1–3) will be provided in the updated version):
  
  Analytical Relevance: Do the model’s analyses closely support the diagnosis, and is there corresponding ECG evidence?
  5: Every analysis point is highly relevant to the diagnosis, with clear supporting evidence.
  4: Most analyses are strongly relevant, with minor insufficiencies.
  
  Analytical Accuracy: Are there any medical factual errors in the model's output?
  5: Completely accurate
  4: Mostly accurate
  
  Analytical Completeness: Does the model comprehensively discuss key ECG components relevant to the diagnosis, including rhythm, intervals, and waveforms?
  5: All relevant ECG features (rhythm, PR, QRS, ST, T waves, intervals, etc.) are accurately discussed.
  4: Most key ECG features are covered, with minor omissions.
  
  Reasoning Quality: Does the model provide a clear, evidence-based reasoning process similar to that of a clinician, logically deriving the diagnosis from ECG features?
  5: Clear and coherent reasoning structure, explaining each step from ECG to diagnosis causally.
  4: Overall reasonable reasoning, but some steps lack detail.
  
  Findings Novelty: Does the model provide insights or findings not noticed by the clinician?
  5: Important new diagnoses or findings.
  4: Novel and somewhat insightful content.
  3: Some new findings, but of limited value.
  
  Clinical Value: Does the model output help in clinical decision-making?
  5: Direct and significant support for clinical judgment; content is clear and reliable.
  4: Most content is helpful and practically useful.
  
  Overall Satisfaction: Subjective rating of the overall quality of this analysis.
  5: Very satisfied.
  4: Satisfied
  

[Weakness 2] Discussion on JoLT

Thank you for bringing JoLT to our attention, we will include a detailed discussion of JoLT in the revised paper, highlighting similarities and differences in model architecture, supervision strategy, and output format.

[Weakness 3] Disagreement between GPT-4o and experts

We recognize that disagreement between GPT-4o-generated interpretations and expert opinions is particularly important in the healthcare domain, where model outputs may directly influence clinical reasoning. We interpret these disagreements not simply as model failures, but rather as reflections of the inherent variability in clinical ECG interpretation.

Overall, the diagnostic conclusions are often aligned between models and human experts. However, discrepancies may arise in finer-grained feature judgments, such as waveform morphology or interval thresholds. This is not unexpected, it mirrors real-world clinical practice, where interpretation differences at the detail level are common even among experienced cardiologists.

Importantly, GPT-4o generated explanations are conditioned on the ground-truth diagnosis from the original clinical reports, which represent a physician’s interpretation at the time of care. As a result, GPT-4o offers one valid expert perspective, while the reviewing cardiologist may offer another. This is consistent with known inter-observer variability, particularly in morphology-sensitive cases.

In addition, expert evaluation in Table 1 confirms that GPT-4o generated ECG-Grounding data satisfy clinical requirements in the majority of cases. The model consistently receives high scores across both reliability and usefulness, with particularly strong performance in analytical completeness and reasoning quality. These results indicate that, under our knowledge-guided instruction data generation framework, GPT-4o can produce ECG interpretations that are not only clinically relevant but also aligned with the expectations of cardiologists.

We also view these discrepancies as valuable opportunities: they highlight cases where model reasoning could be refined with human feedback, and they underscore the importance of modeling expert disagreement to better capture clinical uncertainty. In future work, we plan to explore multi-expert consensus frameworks, inspired by real-world clinical workflows for handling disagreements, such as multi-physician case reviews, to further improve model reliability and alignment with clinical standards.

[Question 2] Experiment with more LLM decoders

Thank you for the insightful question. We test GEM using Qwen2-7B as the LLM decoder, and the results on ECG-Bench are included in Table 2 below.

Unlike the performance gains observed in JoLT or BLIP-2 with decoder swapping, we do not observe significant improvements when replacing the Vicuna-7B used in LLavA with Qwen2-7B in our setting. We hypothesize two potential reasons:

• Qwen2-7B is not pretrained with vision-language alignment, unlike models such as LLaVA (based on Vicuna), which may limit its ability to process cross-modal input effectively.

• Qwen2-7B has a significantly larger vocabulary than Vicuna, which may reduce fine-tuning efficiency or generalization under constrained supervision, particularly in specialized domains like ECG interpretation.

In future work, we plan to explore other vision-language pretrained LLMs as decoders and study their impact on GEM’s performance across tasks.

Table 2: Experiment with Qwen2-7B decoders

Thank you very much for your constructive comments! We are delighted to see that you found the significance of the problem we studied and the value of our proposed methods. We provide detailed responses to your concerns as follows.

Human Evaluation Results

To fully address your concerns, we conduct a comprehensive quantitative human evaluation covering three sources of outputs: GPT-4o generated data, Deepseek-R1 generated data, and GEM generated interpretations.

In total, 400 ECG-Grounding data (200 from GPT-4o and 200 from Deepseek-R1) and 200 GEM's interpretation are independently reviewed by eight board-certified cardiologists, using seven predefined clinical criteria designed to assess both reliability and usefulness. This unified evaluation protocol allows us to (1) verify the quality of GPT-4o generated training data, (2) test the effectiveness of open-source substitutes, and (3) validate the clinical utility of the GEM model.

We report the evaluation results in Table 1 below and provide further discussion in the following responses.

Table 1: Human Evaluation on GPT-4o/Deepseek-R1 Generated Training Data and GEM's Interpretation

Reliability metrics

Usefulness metrics

Scoring criteria (Note: Due to space limitations, we only present scores 4–5 here. The full scoring criteria (including 1–3) will be provided in the updated version):

Analytical Relevance: Do the model’s analyses closely support the diagnosis, and is there corresponding ECG evidence?
5: Every analysis point is highly relevant to the diagnosis, with clear supporting evidence.
4: Most analyses are strongly relevant, with minor insufficiencies.

Analytical Accuracy: Are there any medical factual errors in the model's output?
5: Completely accurate
4: Mostly accurate

Analytical Completeness: Does the model comprehensively discuss key ECG components relevant to the diagnosis, including rhythm, intervals, and waveforms?
5: All relevant ECG features (rhythm, PR, QRS, ST, T waves, intervals, etc.) are accurately discussed.
4: Most key ECG features are covered, with minor omissions.

Reasoning Quality: Does the model provide a clear, evidence-based reasoning process similar to that of a clinician, logically deriving the diagnosis from ECG features?
5: Clear and coherent reasoning structure, explaining each step from ECG to diagnosis causally.
4: Overall reasonable reasoning, but some steps lack detail.

Findings Novelty: Does the model provide insights or findings not noticed by the clinician?
5: Important new diagnoses or findings.
4: Novel and somewhat insightful content.
3: Some new findings, but of limited value.

Clinical Value: Does the model output help in clinical decision-making?
5: Direct and significant support for clinical judgment; content is clear and reliable.
4: Most content is helpful and practically useful.

Overall Satisfaction: Subjective rating of the overall quality of this analysis.
5: Very satisfied.
4: Satisfied

[Weakness 1, Question 1, Limitation 1, Weakness 3, Question 3, Limitation 3] Concerns on GPT-4o generated training data

• Data quality control

  Thank you for raising this important point. We fully agree that quality control is essential when using LLM-generated data in clinical applications. In our data construction pipeline, we incorporated human expert review as a key quality control measure.

  The quantitative human expert evaluation results in Table 1 shows that GPT-4o consistently achieves high scores across both reliability and clinical usefulness, with particularly strong performance in analytical completeness and reasoning quality. These results demonstrate that, with our data generation pipeline, GPT-4o is capable of generating high-quality ECG interpretations that are both clinically reliable and practically valuable.

• GPT-4o cost, scalability, and reproducibility

  In our implementation, generating each sample required approximately 6,000 input tokens and 300 output tokens, resulting in a total cost of roughly USD $3,000 using the OpenAI GPT-4o API for the 300,000 samples in ECG-Grounding dataset.

  We recognize that GPT-4o, while powerful, is not open-sourced and introduces additional cost and access constraints, which may affect scalability and reproducibility.

  To address these concerns, we provide two solutions:

  • First, given the effectiveness of our method, we have scaled up the data generation to the full MIMIC-IV-ECG dataset, which includes 784,680 ECG records from 160,597 patients. These GPT-4o generated grounding data are publicly released and intended to benefit the community as a high-quality and large-scaled resource.

  • Second, to support researchers who wish to apply our method to their own datasets, we test an open-source substitute model, the latest Deepseek-R1. As shown in Table 1, Deepseek-R1 also shown strong capability of generating clinically acceptable, high-quality ECG interpretations with our methods. This demonstrates that our method is adaptable to alternative LLM backbones and remains applicable in settings without commercial API access.

  We will include these results and provide a discussion of the limitations you mentioned in the revised version of the paper.

[Weakness 2, Question 2, Limitation 2] “LLM-as-a-judge” may introduce potential bias

We understand the reviewer’s concern regarding potential stylistic bias and self-reinforcement in using GPT-4o as the evaluation model.

To mitigate these risks and ensure objective assessment, we carefully designed the evaluation prompt to include quantitative scoring criteria and mandatory justification for each score. This structured format helps constrain GPT-4o's evaluation behavior and reduces susceptibility to superficial stylistic similarity.

To further mitigate potential bias, we plan to include a comparative analysis using multiple independent LLMs (e.g., Deepseek-R1, LLaMA) to evaluate the same test outputs in our updated manuscript. This will enable a more comprehensive assessment of the model outputs and help reduce the risk of bias introduced by relying on a single model for judgment.

In addition, as in Table 1, our human evaluation on GEM generated interpretations demonstrate that GEM consistently achieves high scores across both reliability and usefulness dimensions, with most metrics rated above 4 out of 5.

These findings indicate that GEM is capable of producing clinically meaningful and accurate interpretations that align well with cardiologists’ expectations. Also, the human evaluations on GEM are consistent with GPT-4o’s assessment, and support GEM’s potential as a trustworthy assistant for real-world cardiology applications.

We will add the results and discussions in our revised paper.

[Question 4] Disagreements between cardiologists and models

Thank you for this thoughtful question. We interpret the observed disagreements between cardiologists and model outputs as a reflection of the inherent variability in clinical ECG interpretation.

Overall, the diagnostic conclusions are often aligned between models and human experts. However, discrepancies may arise in finer-grained feature judgments, such as waveform morphology or interval thresholds. This is not unexpected, it mirrors real-world clinical practice, where interpretation differences at the detail level are common even among experienced cardiologists.

Importantly, GPT-4o generated explanations are conditioned on the ground-truth diagnosis from the original clinical reports, which represent a physician’s interpretation at the time of care. As a result, GPT-4o offers one valid expert perspective, while the reviewing cardiologist may offer another. This is consistent with known inter-observer variability, particularly in morphology-sensitive cases.

Rather than treating these inconsistencies as errors, we believe they represent valuable opportunities. They highlight cases where model reasoning could be refined with human feedback, and they underscore the importance of modeling expert disagreement to better capture clinical uncertainty. In future work, we plan to explore multi-expert consensus frameworks, inspired by real-world clinical workflows for handling disagreements, such as multi-physician case reviews, to further improve model reliability and alignment with clinical standards.

[Question 5] Candle plot for different models

Thank you for the insightful suggestion. While we are unable to include figures during the response period, we plan to incorporate candle plots in the revised paper to illustrate the distribution and consistency of model performance, particularly under evaluation by multiple independent LLMs as previously discussed.