Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article discusses an innovative approach for non-invasive calculation of cardiac output (CO) through combined analysis of ECG and PPG signals combined with deep learning. The goal is to create a predictive model of Cardiac Index (CI) in patients undergoing cardiac surgery, using only wearable devices. The study is relevant in the context of minimally invasive monitoring especially in patients at increased risk.

Relevance and innovation - the use of combined ECG and PPG signals from wearable devices for CI is a new and practically useful approach.
Clinical validity - the model has been tested on real patients undergoing cardiac surgery which adds significant value.
Extensive methodological section - both the characteristics of the input signals and the neural network architecture used are described in detail.
Comparative analysis - different baseline models are used and it is shown that the proposed architecture outperforms existing approaches in accuracy.
Explainability techniques are included which is important for trust in medical models.

Suggestions and comments:

1. It is not entirely clear whether the model includes any temporal component (e.g. LSTM or TCN) given that PPG and ECG are temporal signals. This should be clarified. It is recommended that the model of the neural architecture used be presented in tabular form (main characteristics). It is recommended to include a figure with the overall model of the architecture used.
2. The formulas used should be numbered.
3. The font size of tables can be slightly reduced to match the font size in the main text. Also, the tables and figures can be aligned with main text where it is possible.

4. Generalizability - the test sample size is relatively small (27 patients) – which limits generalizability. It is recommended to discuss the implications. How exactly the training, validation and testing data are distributed – to be included as numbers and percentages in the text of the article itself. Why are they distributed this way? What distribution method was applied? There is no mention of an external validated dataset (external validation) which is especially important in medical applications.

5. Normalization and preprocessing - it is not entirely clear how the normalization of the signals before feeding them into the network is carried out – especially for PPG. This is key as different patients can have different amplitudes.
6. Assessment measures - the authors use MAE and R², which is appropriate, but do not provide confidence intervals or a test of statistical significance between models. The visualizations are useful, but it would be good to add Bland-Altman analyses, which are standard in evaluating cardiac output methods. Limits of agreement, mean difference, and standard deviation are missing. Including such analysis would improve the reliability of the model estimate against reference measurements.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,
Thank you for your submission. Please find my detailed comments and suggestions in the attached file. I hope you find them helpful for further improving your manuscript.

Kind regards

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript addresses the timely and clinically relevant problem of noninvasive and continuous estimation of cardiac output (CO) using a wearable ECG–PPG and deep learning. The prospective intraoperative design with thermodilution reference using PAC is appropriate, and the comparison between a direct approach (regression to CO) and an indirect approach (CI->CO using CI × BSA) is relevant. In their results, the indirect method clearly outperforms the direct method. However, there are significant issues of internal consistency, clarity, and statistical rigor that must be corrected before the work can be considered publishable.

* In section 3.2, the text reports CI errors in L/min/m² (e.g., MAE 0.241 L/min/m²; RMSE 0.317 L/min/m²), but Table 2 labels MAE/RMSE/bias as L/min. It is necessary to standardize the units throughout the manuscript.

* For CO, the text reports bias and limits of agreement in % (e.g., bias 4.21%, LoA −19.54% to 27.96%), while in other parts L/min is used for the same comparison. It is necessary to report consistently and indicate precisely how the % was calculated.

* Although bootstrap/permutation is used, it should be clarified whether the resampling was done by patient (cluster/hierarchical bootstrap) and not by segment.

* It is indicated that several thermodilutions are averaged over a period and that this average is assigned to the corresponding window, and that the segments are matched by timestamps. It is necessary to specify the exact averaging window for the reference, how the PAC value (at 1 min) was synchronized with the 60-second wearable segment.

* The references contains URLs with tracking parameters such as utm_source=chatgpt.com. These should be replaced with canonical links/DOIs and the format should be adjusted to the journal's style.

* This is not inherently problematic, but the manuscript should clarify the company’s role in study design, data collection, analysis, and manuscript preparation (and ensure wording in the COI section is fully transparent).

* When describing the dataset, also report the min–max range of the reference CO/CI values in the main text, not just the median and IQR;

* Justify more clearly the choice of 60 s windows and 50% overlap, discussing the balance between physiological relevance, signal stability, and sample dependence

* Briefly explain why Fourier resampling is appropriate in this context.

* available upon request is acceptable, but consider sharing at least derived/anonymous features or a reproducible pipeline for verification.

In general, I recommend making the previously mentioned suggested changes.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

I read with great interest this manuscript on non-invasive cardiac output estimations by ECG and PPG signals.

The manuscript is well-written and I have little comments. I only think that the limitations and conclusions need te be slightly adjusted: First of all, only patients undergoing aortic surgery were included. It would have been nice to have patients with reduced LV ejection fraction included in the sample. Secondly, the use for out-of-hospital environments is far away, patients not under anesthesie change their body positions from supine to standing etc. and movements can introduce signal changes. All of these situations have not been tested in the current study.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Please see the attached document for my reviewer comments.

Kind regards

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Dear authors:

After reviewing the modifications incorporated into the new version of the manuscript, I believe that the adjustments made adequately address the comments from the review. In particular, they corrected the consistency of units and metrics, made the operational (and therefore reproducible) form of the Bland–Altman calculation more explicit, clarified the bootstrap scheme used for confidence intervals, and detailed the synchronization/tagging procedure between the wearable signal and the PAC reference, including the justification for the window and overlap. Taken together, these changes make the manuscript methodologically clearer and report results consistently, within the limitations associated with the sample size.

Author Response

Please see the attachment

Author Response File: Author Response.pdf