Author response:
The following is the authors’ response to the original reviews
Public Reviews:
Reviewer #1 (Public review):
Summary:
The electrocardiogram (ECG) is routinely used to diagnose and assess cardiovascular risk. However, its interpretation can be complicated by sex-based and anatomical variations in heart and torso structure. To quantify these relationships, Dr. Smith and colleagues developed computational tools to automatically reconstruct 3D heart and torso anatomies from UK Biobank data. Their regression analysis identified key sex differences in anatomical parameters and their associations with ECG features, particularly post-myocardial infarction (MI). This work provides valuable quantitative insights into how sex and anatomy influence ECG metrics, potentially improving future ECG interpretation protocols by accounting for these factors.
Strengths:
(1) The study introduces an automated pipeline to reconstruct heart and torso anatomies from a large cohort (1,476 subjects, including healthy and post-MI individuals).
(2) The 3-stage reconstruction achieved high accuracy (validated via Dice coefficient and error distances).
(3) Extracted anatomical features enabled novel analyses of disease-dependent relationships between sex, anatomy, and ECG metrics.
(4) Open-source code for the pipeline and analyses enhances reproducibility.
Weaknesses:
(1) The linear regression approach, while useful, may not fully address collinearity among parameters (e.g., cardiac size, torso volume, heart position). Although left ventricular mass or cavity volume was selected to mitigate collinearity, other parameters (e.g., heart center coordinates) could still introduce bias.
(2) The study attributes residual ECG differences to sex/MI status after controlling for anatomical variables. However, regression model errors could distort these estimates. A rigorous evaluation of potential deviations (e.g., variance inflation factors or alternative methods like ridge regression) would strengthen the conclusions.
(3) The manuscript's highly quantitative presentation may hinder readability. Simplifying technical descriptions and improving figure clarity (e.g., separating superimposed bar plots in Figures 2-4) would aid comprehension.
(4) Given established sex differences in QTc intervals, applying the same analytical framework to explore QTc's dependence on sex and anatomy could have provided additional clinically relevant insights.
We thank Reviewer 1 for their kind and constructive comments. While we have thoroughly addressed all specific recommendations below, in brief, we have added new analysis of the variance inflation factor in Supplementary Tables 2 and 3 to reassure readers that the chosen parameter sets exhibit low levels of collinearity, and provided more explanation for why the relative positional parameters were chosen to avoid this issue. We have added explanatory figures for all positional and orientational parameters to improve understanding of the technical details, and improved clarity of existing figures as detailed below. We welcome the suggestion to add QT interval to the manuscript – whilst this was only available in the UK Biobank for a single lead, we have included an analysis of both QT and QTc intervals in this lead to Page 10, and added some discussion of this to the second full paragraph of Page 14.
Recommendations for the authors:
Reviewer #1 (Recommendations for the authors):
Comment 1: “Collinearity and Regression Analysis: It would be valuable to assess the collinearity among the regressed parameters (e.g., cardiac size, torso volume, heart center positions [x, y, z], and cardiac orientation angles) and evaluate whether alternative regression methods (e.g., ridge regression) might improve robustness. Additionally, cardiac digital twinning with electrophysiological models could help isolate the exact contribution of electrophysiology while enabling sensitivity analysis. Nonlinear regression or machine learning approaches might also enhance the predictive power of the analysis.”
We thank the reviewer for drawing attention to the important issue of collinearity in the parameter sets used in the regression analysis. To address this, we have added Supplementary Tables 2 and 3, which detail the variance inflation factors for each of the parameter sets used. This was considered in the selection of anatomical parameters – e.g. using relative position not absolute distances between landmarks, which would be more collinear. As these are all below a value of 3.4, we believe that the effect of collinearity is limited, and thus to reduce subjectivity of parameter selection in more complex methods, and encourage interpretability, we have retained our linear regression analysis. In addition, we have added an explanation to the second full paragraph on Page 6 of how we calculated the relative, rather than absolute position of the cardiac centre partially to avoid the problem of collinearity when using multiple absolute distances. We concur that modelling and simulation techniques are well suited to explore the electrophysiological component further – as this is out of the scope of this work, we have addressed the role of these methods in future work in the final paragraph of Page 16.
Comment 2: “Figure Clarity (Bar Plots): The superimposed bar plots in Figures 2-4 are difficult to interpret; separating the bars for each coefficient would improve readability.”
We accept that the stacked bar plots could be improved in their clarity. Whilst plotting each anatomical parameter separately multiplies the number of plots by a factor of nine, and makes comparison between parameters more difficult, we have added clear horizontal grid lines in order to make values easier to read and interpret.
Comment 3: “Feature Extraction Visualization: A schematic figure illustrating the steps for measuring heart positional parameters (e.g., with example annotations) would help readers better understand the feature extraction methodology.”
We agree with the reviewer that the calculation of positional and orientational parameters is crucial to illustrate clearly. We have included additional Supplementary Figures 2 and 3 to better convey these parameters.
Reviewer #2 (Public review):
Summary:
Missed diagnosis of myocardial ischemia (MI) is more common in women, and treatment is typically less aggressive. This diagnosis stems from the fact that women's ECGs commonly exhibit 12 lead ECG biomarkers that are less likely to fall within the traditional diagnostic criteria. Namely, women have shorter QRS durations and lower ST junction and T wave amplitudes, but longer QT intervals, than men. To study the impact, this study aims to quantify sex differences in heart-torso anatomy and ECG biomarkers, as well as their relative associations, in both pre- and post-MI populations. A novel computational pipeline was constructed to generate torso-ventricular geometries from cardiac magnetic resonance imaging. The pipeline was used to build models for 425 post-myocardial infarction subjects and 1051 healthy controls from UK Biobank clinical images to generate the population.
Strengths:
This study has a strength in that it utilizes a large patient population from the UK Biobank (425 postMI and 1051 healthy controls) to analyze sex-based differences. The computational pipeline is stateof-the-art for constructing torso-ventricular geometries from cardiac MR and is clinically viable. It draws on novel machine learning techniques for segmentation, contour extraction, and shape modeling. This pipeline is publicly available and can help in the large-scale generation of anatomies for other studies. This allows computation of various anatomical factors (torso volume, cavity volume, etc), and subsequent regression analysis on how these factors are altered before and after MI from the 12-lead ECG.
Weaknesses:
Major weaknesses stem from the fact that, while electrophysiological factors appear to play a role across many leads, both post-MI and healthy, the electrophysiological factors are not stated or discussed. The computational modeling pipeline is validated for reconstructing torso contours; however, potential registration errors stemming from ventricular-torso construction are not addressed within the context of anatomical factors, such as the tilt and rotation of the heart. This should be discussed as the paper's claims are based on these results. Further analysis and explanation are needed to understand how these sex-specific results impact the ECG-based diagnosis of MI in men and women, as stated as the primary reason for the study at the beginning of the paper. This would provide a broader impact within the clinical community. Claims about demographics do not appear to be supported within the main manuscript but are provided in the supplements. Reformatting the paper's structure is required to efficiently and effectively present and support the findings and outcomes of this work.
We thank Reviewer 2 for their considered and detailed feedback. We greatly appreciate the invitation to elaborate on the electrophysiological factors, and we have added discussion of this matter to the second and third full paragraphs on Page 14, extending to Page 15 and first full paragraph on Page 15, and highlighted the role of modelling and simulation in future work on the third full paragraph of Page 16. We agree that registration errors are one reason behind remaining reconstruction errors and feel a strength of our study is that the large number of subjects used aided in reducing the effect of this noise, and have updated the second full paragraph of Page 16 to reflect this. We are wary of moving too many supplemental figures and tables describing demographic trends to the main manuscript for fear of diluting the specific answers to our research questions. We have however actioned the suggestions as detailed below to reformat the paper, including redressing the balance of supplemental versus main methodological sections, and thank the reviewer for their guidance in increasing our clarity.
Reviewer #2 (Recommendations for the authors):
(1) Please detail what "chosen to be representative of the underlying dataset" means in terms of a validation dataset.
We thank the reviewer for addressing the lack of clarity in this matter. We have added a reference in the third full paragraph on Page 6 to Supplementary Appendix 1.1, where we have included full details of the selection criteria.
(2) “Current guidelines ... further research [16]." The paragraph should begin with a broader statement that is relevant to the fact that the entire body of work focuses on ECG-based diagnosis differences in women, rather than LVEF through echocardiography.
We have revised the introduction to Paragraph 3 on Page 3 to clarify our motivation for focusing on the ECG in order to shape proposals for novel ECG-based risk stratification tools.
(3) The last paragraph of the introduction should more clearly state what was performed and how you aim to prove your hypothesis. There is no mention of the data, the regression model, or other key aspects important to the reader.
We have added methodological details to Paragraph 5 on Page 3 in order to clarify our approach in testing our hypothesis.
(4) An overview paragraph should be included in the Methods at the beginning.
We thank the reviewer for this valuable suggestion – we have added an overview paragraph to the start of the methodology section on Page 5.
(5) The computational pipeline portion of the methods should be written in full paragraphs instead of almost a bulleted list. In general, more details from the supplement should be provided in the methods.
We thank the reviewer for raising important points concerning the balance of methodological description in the main manuscript and the supplementary materials. We have added detailed description of the reconstruction pipeline to Pages 5 and 6. We feel that the ordered format of the methods section adds to the reproducibility and transparency of our methodology.
(6) The torso reconstruction method was already validated in Smith et al. [29]. What value does your additional validation bring to this methodology? Furthermore, how does the construction of the ventricular-torso reconstructions using the cardiac axes (not just the torso contours) influence ECG metrics?
We apologise that this was not clear – we have clarified in Paragraph 4 on Page 5 that while Smith et al. 2022 provided a detailed validation to the contour extraction networks, it did not validate the torso reconstruction pipeline, as it only presents the reconstruction of two cases as a proof of concept. We have also expanded the second full paragraph on Page 6 to explain that the sparse (but not dense) cardiac anatomies were constructed in order to calculate the cardiac size, which we found was a key factor moderating many ECG biomarkers. We also specified that the cardiac position and orientation were necessary in order to relate these to the torso axes and positions of the ECG electrodes.
(7) Include the details of the regression analysis in the main body of the methods for the readers. This is crucial to the claims and outcomes of the paper. Only a sentence is included in the results and one in the figure: "Each factor's contribution is calculated from the product of the regression coefficients and anatomical sex differences (Supplementary Appendix 1.5)." What specific contributions can I expect to see in the results figures? The results are filled with methodological aspects that should be in the results.
We thank the reviewer again for this important comment regarding the balance of the main text methodology and supplementary methodology sections. We have added detail to the statistical analysis section of the main text on Pages 7 and 8 in order for the reader to understand the following results section without consulting the supplemental methods. We have also removed these details from the results section.
(8) What is "the remaining estimated effect of electrophysiology". Did you do simulations on the electrophysiology, or how is this computed from the clinical data of patients? More explanation is needed, as without this, the paper is just focusing on anatomy.
We have clarified this important point by moving the explanation of the methodology underpinning our estimation of the electrophysiological contributions using the clinical ECGs from the supplementary methods to the main manuscript on the second full paragraph on Page 7, and continuing to Page 8. We have also specified the role of simulations studies in future work on the final paragraph on Page 16.
(9) Include an overview paragraph of the methods to create more structure.
We thank the reviewer again for the further attention to this issue – as previously, we have added an overview paragraph to the methodology section on Page 5.
(10) Only 19.8% of the patients were female, which is probably due to females having a more severe presentation of the disease. How does this impact, bias, or skew your results?
This comment raises a very interesting point, and while the origin of this imbalance is of course multifactorial – women likely do have lower rates of MI events due to the cardioprotective role of estrogen and different health promoting behaviours, and our sex imbalance was reflective of wider trends in MI diagnosis. However, as mentioned in Paragraph 2 Page 3 of the text, there are more missed MI diagnoses in women, and we agree that this may lead to a more severe presentation of female MI pathophysiology. We have expanded the first full paragraph on Page 16 to specify the ECG and demographic impacts that this has on our results, and that it is a strength of this work that we may contribute to future adjustment of the diagnostic criteria, such that future investigations do not have this bias, and that clinical outcomes are improved.
(11) A lot of extra information is provided in Tables 1 and 2. Include additional information in the supplements that is not directly relevant to your findings.
We agree that Table 2 is supplementary, rather than critical information, and have moved it accordingly to the Supplementary Materials on Page 38. We do believe that Table 1 is central for understanding the extracted dataset.
(12) Combine paragraphs 3 and 4 into a single paragraph. "Current guidelines..." and "T wave amplitude...". They are part of a single coherent concept.
We have removed the paragraph break on Page 3 Paragraph 3.
(13) Check all acronyms throughout the paper. The abbreviation for sudden cardiac death (SCD) is only used once in the same paragraph. Remove the acronym and type it out. T-wave amplitude (TWA) is introduced twice in a Figure caption and not introduced until the methods.
Many thanks for this suggestion – we have reviewed all acronyms in the manuscript.
(14) "Figure 1B showcases the capability of the computational pipeline to extract torso contours and reconstruct them into 3D meshes". Isn't this Figure 1A?
We apologise that this was unclear, and have updated the sentence on the first full paragraph of Page 8 to clarify the purpose of Figure 1B.
(15) No need to state: "Female y-axis limits have been adjusted by the difference in healthy QRS duration between sexes for ease of comparison" in the Figure 2 caption.
We have removed this statement on all relevant captions.
(16) The paragraph "For lead V6, 15.9% of healthy subjects..." can be combined with the previous section.
We have removed this paragraph break on Page 9 to improve readability.
(17) The only demographics I could find were age and BMI. State which demographics you used explicitly. This is especially true when the discussion makes claims like "Our findings suggest that corrected QRS duration taking into consideration demographics...". How did you take them into account?
We accept that our previous description of the demographic adjustment to QRS duration in the discussion did not adequately reflect the comprehensiveness of our approach, and have adjusted the second paragraph on Page 14 to rectify this.
(18) The results section is also almost a bulleted list that should be written and reformatted into paragraphs.
The ordered style of our results section was designed to compare how our obtained data answers our research question differently for ECG intervals, amplitudes, and axis angles. Whilst we have adjusted paragraph breaks and moved methodological details to more appropriate sections, we have retained this stylistic choice.
(19) The following sentence should be in the introduction: "Alterations to the polarity and amplitude of the T wave are used in the diagnosis of acute MI [42] and TWA affects proposed risk stratification tools, particularly markers of repolarization abnormalities [9, 43]."
We thank the reviewer for this suggestion. We have included the discussion of how TWA is separately used in proposed risk stratification and current diagnostic tools in Paragraph 3 of Page 3.
