Identifying Hypertrophic or Dilated Cardiomyopathy: Development and Validation of a Fine-Tuned ResNet50 Model Based on Electrocardiogram Image

We sincerely appreciate your valuable feedback and have carefully considered your concerns regarding this manuscript. Thank you for highlighting these critical points. We believe these revisions will strengthen the manuscript’s transparency and relevance to clinical practice. Below is our point-by-point response.

Comments 1: The title presents the topic of the study. However, ResNet50 could be included in the title to clarify the research.

Response 1: Thank you for this suggestion. We agree that specifying the methodology enhances clarity. We have revised the title to include "ResNet50". This change is reflected in the revised manuscript (Title, Line 1).

Revised Title: " Identifying hypertrophic or dilated cardiomyopathy: Development and Validation of a fine-tuned ResNet50 Model Based on Electrocardiogram images"

Comments 2: Abstract represented the text body.

Response 2: We thank the reviewer for this observation. The abstract has been restructured to align precisely with the manuscript body.

“There is no established detecting tool for hypertrophic cardiomyopathy (HCM) and dilated cardiomyopathy (DCM). This study aimed to develop a deep learning-based model for identifying HCM and DCM using standard 12-lead electrocardiogram (ECG) images. We obtained a cohort of patients with HCM (171 ECG images) or DCM (364 ECG images) confirmed by cardiovascular magnetic resonance (CMR) examinations, underwent both ECG and CMR within 30 days at our institution. Age- and sex-matched healthy controls (2314 ECG images) were selected from our Health Check Center. A total of 2,849 ECG images were processed via a fine-tuned ResNet50 architecture, with stratified 5-fold cross-validation for model training, validation and testing. The proposed model demonstrated strong performance in distinguishing DCM, achieving an area under the receiver operating curve (AUROC) of 0.996 and an area under the precision-recall curve (AUPRC) of 0.940. For the detection of HCM, the model also achieved an AUROC of 0.980 and an AUPRC of 0.953, respectively. The model prospectively exhibited stability in temporal validation. Furthermore, representative images of the Gradient-weighted Class Activation Mapping (Grad-CAM) technique analysis showed the region corresponding to the anterior and anteroseptal leads were the most important areas for prediction of HCM or DCM. This temporally validated fine-tuned ResNet50 model shows promise to inexpensively detect individuals with HCM or DCM.”

Revised text is marked in red (Abstract, Lines 14-20).

Comments 3: Introduction section represents the scope of the problem. It is brief and can include more data on AI ECG-reading models. Moreover, the aim of the study could be more precise.

Response 3: We appreciate this feedback. The introduction now revised includes more data on existing AI-ECG models (Introduction, Lines 55-78; Page 2) and clarifies the study aim as Added text is highlighted in red:

"To develop a ResNet50-driven framework for differentiating HCM or DCM from standard 12-lead 10-s ECG images, addressing gaps in screening tools for cardiomyopathies."

Added text is highlighted in red (Introduction, Lines 81-82; Page 2).

Comments4.1: Study population. Flowchart is presented. Were patients with ischemic DCM excluded? Who were the controls? Were only patients with normal ECG included as controls?

Response 4.1: We thank the reviewer for highlighting this ambiguity. Flowchart also presented. The revised Methods section clarifies (2.1. Study Population, Lines 98-102):

1. Exclusion Criteria: "Patients with ischemic DCM were excluded."
2. Control Group: "Healthy controls were selected from individuals with no cardiovascular heart disease, normal ECG from our Health Check Center during the same time period. "Additions are marked in red.

Comments 4.2: Experiment design. The title has to be clarified. As well as the phrase “Each included ECG was clinically confirmed by three cardiologist within”. What was confirmed?

Response 4.2: We apologize for the lack of clarity. The revised Methods section includes (Section 2.2, Lines 109-116):

- Revised Subheading: " ECG Image Preprocessing and Validation:

" Changes are highlighted in red.

- Clarified Text: " Each ECG image in the HCM/DCM was reviewed by three cardiologists blinded to HCM/DCM vs. control in order to document the following ECG features using pre-defined criteria: normal vs. abnormal ECG, atrial fibrillation or flutter, Abnormal Q, ST-T change, right bundle branch block (RBBB). ECG images in all were standard, 12-lead 10-s ECGs acquired in the supine position at a sampling rate of 500 Hz using a GE-Marquette ECG machine. Standardized high-resolution electrocardiographic imaging protocols with clinically validated data integrity ensure acquisition of diagnostically reliable waveforms. Quality assurance metrics confirm complete image preservation."

Comments 6: Discussion section. Notably, Limitation section has to be incorporated. The study only includes three cohorts: DCM, HCM, and controls. These results could not be translated into daily practice with high prevalence of CAD with previous MI, LVH etc.

Response 6: We deeply appreciate this critical point. A dedicated "Limitations" subsection has been added to the Discussion, stating:

" This study has several limitations. Firstly, this model was to establish a robust framework for detecting two major forms of cardiomyopathy (DCM/HCM) before expanding to rarer subtypes, limiting generalizability to populations with CAD or secondary LV hypertrophy. Secondly, validation in real-world cohorts may be with overlap-ping comorbidities. The need for future studies to validate our results in cohorts with diverse cohorts, and generalization to Restrictive cardiomyopathy (RCM) or Arrhythmogenic Right Ventricular Cardiomyopathy (ARVC) re-quires further validation. Caution in extrapolating conclusions to patients with complex comorbidities until further multi-center collaborations evidence is available."

Marked in red (Limitations, Lines 368-376).

Comments 7: Conclusion section should be included.

Response 7: A standalone Conclusion section has been added Highlighted in red (Conclusion, Lines 378-387):

"Based on the availability of real 12-lead ECG images in clinics, this study developed and temporally validated an image-model that detects HCM or DCM. This image-model is demonstrating robust performance across different time periods in patients, provides an automated and accessible screening strategy for patients with HCM or DCM, which may inform the early timing of imaging/interventions and facilitate improved access to care.

The model can be effectively applied to real clinic care, reducing the cost of cardiomyopathy diagnosis for patients that are not diagnosed. In addition, ECG model interpretation related to HCM/DCM 12-lead ECG images can provide clinicians with auxiliary interpretation to improve the application in clinical practice.".

8. Figures and  tables are  ok.

Response 8: Thank you for your valuable feedback. I have made the necessary adjustments to the figures and tables as suggested (Figure 4, Figure 5).

9. References are updated.

Response 9: References are updated Highlighted in red.

We sincerely appreciate your valuable feedback and have carefully considered your concerns regarding this manuscript. Thank you for highlighting these critical points. We believe these revisions will strengthen the manuscript’s transparency and relevance to clinical practice. Below is our point-by-point response.

Comments and Suggestions for Authors

I find the manuscript interesting and would like to honestly express my congratulations and admiration to the team for their study.

However, there are certain aspects I would like to address:

Comments 1: Why is the cardiac ultrasound not discussed, since it is a widely available technique with a great utility in CMH and CMD?

Response 1: We sincerely appreciate the reviewer’s acknowledgment of echocardiography pivotal role in cardiomyopathy screening. Our decision to exclude echocardiographic data aligns with the study’s specific objectives and practical considerations:

1.Rationale for Prioritizing CMR Over Echocardiography:

While echocardiography remains a cornerstone for initial HCM/DCM evaluation due to its accessibility, this study utilized cardiac magnetic resonance (CMR) as the gold standard for inclusion. Accuracy Constraints: Echocardiographic measurements are prone to variability from operator-dependent techniques, anatomical complexities, and suboptimal acoustic windows in obese patients, leading to underdiagnosis rates.

2.Diagnostic Superiority of CMR: CMR provides myocardial fibrosis, precise wall thickness mapping, and differentiation of HCM mimics. This ensures standardized phenotyping, reducing misclassification risks inherent to echocardiography.

3.Detecting Feasibility: AI-ECG models are designed to replace echocardiography- or CMR-based population screening, particularly in low-resource settings where echocardiography expertise is scarce.

Comments 2: I personally consider ECG trace to be a better term than ECG image.

Response 2: We appreciate this feedback. The Introduction now includes expanded discussions on existing AI-ECG trace models and image model. Added text is highlighted in red (Introduction, Lines 68-78; Page 2).

However, the raw signals ECG data is difficult to be saved and does not conform to the actual clinical care. Based on clinician observation needs, some research attempt to use ECG images to detect heart disease.

Recent advances in deep learning have enhanced etiological differentiation of cardiomyopathies. Haimovich et al. (2023) developed a 12-lead ECG-based CNN achieving AUC 0.92 for distinguishing hypertrophic cardiomyopathy (HCM) from other left ventricular hypertrophy etiologies. Tayal et al. (2022) implemented a random forest model integrating clinical, genetic, and cardiac MRI data to subclassify dilated cardiomyopathy (DCM) into fibrosis-metabolic and non-fibrotic subtypes. Shrivastava et al. (2021) demonstrated a CNN model with 98.8% sensitivity and 100% negative predictive value for DCM detection (LVEF≤45%), vali-dated across multiple cohorts (AUC 0.955-0.98). Zhou et al. (2023) employed XGBoost with clinical/echocardiographic data to differentiate ischemic vs. non-ischemic DCM (AUC 0.934). Comparative ML analyses revealed random forest models with mini-mum redundancy maximum relevance achieved 91.2% accuracy in distinguishing DCM, HCM, and healthy subjects (average AUC 0.98).

Comments 3: Please formulate a conclusion to underline the practical aspects of your research.

Response 3: Based on the availability of real 12-lead ECG images in clinics, this study developed and temporally validated an image-model that detects HCM or DCM. This image-model is demonstrating robust performance across different time periods in patients, provides an automated and accessible screening strategy for patients with HCM or DCM, which may inform the early timing of imaging/interventions and facilitate improved access to care.

The model may be effectively applied to real clinic care, reducing the cost of cardiomyopathy diagnosis for patients that are not diagnosed. In addition, ECG model interpretation related to HCM/DCM 12-lead ECG images can provide clinicians with auxiliary interpretation to improve the application in clinical practice.

In the references you cited the 2014 ESC Guidelines on diagnosis and management of hypertrophic cardiomyopathy, however the European Society of Cardiology has released a newer edition in 2023 of the guidelines, which would be better to be referenced.