A Screening Method for Determining Left Ventricular Systolic Function Based on Spectral Analysis of a Single-Channel Electrocardiogram Using Machine Learning Algorithms

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The study presents a machine learning–based screening method for detecting LVSD using spectral analysis of a single-lead ECG. The topic is clinically significant and aligns with the need for low-cost, scalable cardiovascular screening tools. Following should be addressed: 
1. The ECG feature extraction process and model hyperparameters are not disclosed. The authors mention algorithms but the concrete hyperparameters for each model are not listed. Without training configuration the study cannot be reproduced. The authors should list architecture of neural networks, activation functions and optimization methods that are being used.
2.  Presenting SHAP or coefficient plots would clarify which ECG parameters are most associated with LVSD. Feature importance should be visualized in order to enhance interpretability.

Author Response

The study presents a machine learning–based screening method for detecting LVSD using spectral analysis of a single-lead ECG. The topic is clinically significant and aligns with the need for low-cost, scalable cardiovascular screening tools. Following should be addressed:

1. Comments to the Author:

“The ECG feature extraction process and model hyperparameters are not disclosed. The authors mention algorithms but the concrete hyperparameters for each model are not listed. Without training configuration the study cannot be reproduced. The authors should list architecture of neural networks, activation functions and optimization methods that are being used.”.

Response:

The addition has been made to the text of the article.

2. Comments to the Author: “Presenting SHAP or coefficient plots would clarify which ECG parameters are most associated with LVSD. Feature importance should be visualized in order to enhance interpretability.”.

Response:

The addition has been made to the text of the article. Table 3.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors This paper presents a single-channel ECG–based machine learning method to screen for left ventricular systolic dysfunction (LVSD). The study reports strong performance metrics (AUC = 0.849–0.972), but several methodological and conceptual clarifications are required to strengthen its scientific validity. Comments

1. In the Introduction, please elaborate on why prediction using single-channel ECG data is clinically necessary. What specific gap or unmet need does it address compared with existing 12-lead AI-ECG approaches?
2. The definition of “normal” LVEF as 52% for men and 54% for women is based on older standards. Please justify its use and discuss how age, sex, and ethnicity may affect the validity of this threshold.
3. In Figure 1, please clarify why six and ten participants were excluded from each phase. Were these due to ECG quality, missing echocardiographic data, or other reasons?
4. The paper lacks a detailed explanation of the feature selection process. Describe explicitly how the 200+ ECG parameters were screened and which statistical or ML-based methods determined the final predictors.
5. The study’s purpose appears ambiguous: is the primary goal to identify the most informative ECG features or to compare multiple machine-learning algorithms? Please clarify this conceptual framework.
6. Provide more information on the external validation cohort: how were the 600 subjects recruited, what were their demographic/clinical characteristics, and how were reference echocardiograms obtained? Include PPV and NPV to complement sensitivity and specificity.
7. Since AI-ECG–based LVSD detection is a highly active research field, the reference list is insufficient. Incorporate additional contemporary studies, particularly from leading groups (e.g., Attia et al., Mayo Clinic; Vaid et al., Mount Sinai).
8. Discuss the trade-off between convenience and information loss when using a single-lead ECG. Cite literature showing that diagnostic accuracy tends to decline as lead number decreases.
9. Most single-lead applications assume portable or home-based settings, yet your study was hospital-based. Please discuss this discrepancy and address potential limitations regarding user proficiency and data quality in real-world use.
10. Drawing strong conclusions from external validation alone seems overstated. Please temper the Discussion accordingly and clarify that further multicenter validation is required.
11. The concluding phrase “modern signal processing and machine learning technologies” is too vague. Specify which particular techniques (e.g., wavelet transform, Lasso regression, Extra Trees) yielded the best performance.
12. The Limitations section should be rewritten narratively rather than as bullet points, with clear reasoning and acknowledgment of generalizability constraints.

Author Response

This paper presents a single-channel ECG–based machine learning method to screen for left ventricular systolic dysfunction (LVSD). The study reports strong performance metrics (AUC = 0.849–0.972), but several methodological and conceptual clarifications are required to strengthen its scientific validity.

1. Comments to the Author:

In the Introduction, please elaborate on why prediction using single-channel ECG data is clinically necessary. What specific gap or unmet need does it address compared with existing 12-lead AI-ECG approaches?

Response:

The addition has been made to the text of the article.

 

2. Comments to the Author:

The definition of “normal” LVEF as 52% for men and 54% for women is based on older standards. Please justify its use and discuss how age, sex, and ethnicity may affect the validity of this threshold.

Response:

According to clinical guidelines for echocardiography analysis at the time of patient examination and results processing (2024 - echocardiography guidelines in Russia doi: 10.15829/1560-4071-2025-6271 and guidelines of the European Society of Echocardiography), the left ventricular ejection fraction specified in our article was considered normal. In 2025, the American Society of Echocardiography guidelines stated that a decrease in ejection fraction below 53% should be considered significant. In our patient cohort, only two men had an ejection fraction of 53%. The remaining patients were either above or below the specified range. We consider the study results valid when using both previous and updated guidelines.

 

3. Comments to the Author:

In Figure 1, please clarify why six and ten participants were excluded from each phase. Were these due to ECG quality, missing echocardiographic data, or other reasons?

Response:

The addition has been made to the fig.1 of the article.

 

4. Comments to the Author:

The paper lacks a detailed explanation of the feature selection process. Describe explicitly how the 200+ ECG parameters were screened and which statistical or ML-based methods determined the final predictors.

Response:

First, parameters that had independent significant diagnostic value were selected. Table 3 has been added.

 

5. Comments to the Author:

The study’s purpose appears ambiguous: is the primary goal to identify the most informative ECG features or to compare multiple machine-learning algorithms? Please clarify this conceptual framework.

Response:

Using the identified diagnostically significant ECG parameters and patient data, we built various machine learning models, selecting the one with the optimal balance of sensitivity and specificity. A list of the models is included in the description of the method in the article.

 

6. Comments to the Author:

Provide more information on the external validation cohort: how were the 600 subjects recruited, what were their demographic/clinical characteristics, and how were reference echocardiograms obtained? Include PPV and NPV to complement sensitivity and specificity.

Response:

The addition has been made to the text of the article.

 

7. Comments to the Author:

Since AI-ECG–based LVSD detection is a highly active research field, the reference list is insufficient. Incorporate additional contemporary studies, particularly from leading groups (e.g., Attia et al., Mayo Clinic; Vaid et al., Mount Sinai).

Response:

We included Attia as well as Vaid works in the reference list. A review of the available literature revealed only few studies on the use of AI analysis specifically for single-channel ECG.

 

8. Comments to the Author:

Discuss the trade-off between convenience and information loss when using a single-lead ECG. Cite literature showing that diagnostic accuracy tends to decline as lead number decreases.

Response:

The addition has been made to the text of the article.

 

9. Comments to the Author:

Most single-lead applications assume portable or home-based settings, yet your study was hospital-based. Please discuss this discrepancy and address potential limitations regarding user proficiency and data quality in real-world use.

Response:

Portable devices available to consumers require only a short training period and do not require a doctor. Purchasers receive instructions from the seller. In our studies monitoring patients with heart failure, less than 5% of home recordings were unusable. Therefore, there are no significant limitations.

 

10. Comments to the Author:

Drawing strong conclusions from external validation alone seems overstated. Please temper the Discussion accordingly and clarify that further multicenter validation is required.

Response:

The addition has been made to the text of the article.

 

11. Comments to the Author:

The concluding phrase “modern signal processing and machine learning technologies” is too vague. Specify which particular techniques (e.g., wavelet transform, Lasso regression, Extra Trees) yielded the best performance.

Response:

The addition has been made to the text of the article.

 

12. Comments to the Author:

The Limitations section should be rewritten narratively rather than as bullet points, with clear reasoning and acknowledgment of generalizability constraints.

Response:

The addition has been made to the text of the article.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors fulfilled the requests and paper is ready for publication.