Machine Learning Framework for Classifying and Predicting Depressive Behavior Based on PPG and ECG Feature Extraction
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors present a noteworthy contribution to the field of affective computing by developing a machine learning model capable of detecting depressive states using ECG and photoplethysmograph data. This research topic is of significant interest due to the potential for non-invasive, objective assessment of mental health. The methodology is clearly described, and the results are presented in a comprehensive and well-supported manner. The findings offer promising insights into the relationship between physiological signals and depressive symptoms, warranting further investigation and potential clinical applications.
Author Response
Thank you very much for your thoughtful review and for recognizing the work we have done, as well as for validating the methods and results we presented. Please note that some significant changes were made in various sections of the paper based on feedback from other reviewers, particularly in a new results section titled "Stratified k-fold results." I remain attentive to any further comments you may have regarding the revised submission of the paper.
Reviewer 2 Report
Comments and Suggestions for AuthorsIn this paper, the authors give a method to detect depressive states using electrocardiogram (ECG) and photoplethysmogram (PPG) signals. There are 59 participants in this study, split into two groups: those with and without depressive symptoms. By utilising machine learning models, particularly a Random Forest Model, to examine the physiological signals gathered, the authors were able to identify depressive states with 81% accuracy. It is observed that this approach could be a valuable tool in screening for depression, with potential applications in preventive diagnostics and psychiatric alarm systems.
I have few comments on the paper:
1. The authors could improve the robustness of the results by increasing the sample size, i.e., by including participants from diverse demographic backgrounds and clinical settings.
2. They could utilize advanced feature selection techniques like RFE or PCA to decrease dimensionality and enhance model performance.
3. If the authors could use deep learning models, such as CNNs or LSTMs, to better capture complicated patterns in ECG and PPG signals, this work might be more interesting.
4. The paper lacks clear information on whether the model was tested on a separate validation set or cross-validation, potentially causing overfitting where the model performs well on training data.
5. The paper discusses feature extraction from PPG and ECG signals, but lacks detailed information on selection methods. Understanding predictive features could enhance the study's depth.
6. Despite the 81% accuracy rate of the model, a more thorough explanation of the consequences of this accuracy in clinical situations would be beneficial for the paper. What effects, for example, might false positives or false negatives have when it comes to the diagnosis of depression?
7. The article makes suggestions for possible uses in psychiatric alarms and preventive diagnosis, but it skips over the difficulties in putting this into practice in actual situations. How might this technology be incorporated, for instance, into the current healthcare systems?
Author Response
Thank you for taking the time to review the paper and for validating the methods and findings. Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe study aimed at classifying depressive behavior using PPG and ECG features. The paper is generally good. I have several comments.
I think key information is missing that could significantly impact the interpretation of the results, such as the severity of depression, duration of depression, medication and treatment history. These factors could influence the physiological markers and would help readers gain deeper insights into the findings.
The authors mentioned other studies that use CNNs but do not explain why this approach wasn’t used in their own study. I think the authors should provide a rationale for the choice of machine learning methods.
The description of artifact removal is somewhat unclear. The authors mentioned filtering the signals, but is filtering the only step in the artifact removal process? It would be helpful to clarify
The authors recognized the overfitting in their model. I think the authors should provide a more thorough discussion on this.
Author Response
Thank you for taking the time to review the paper and for validating the methods and findings. Please see the attachment.
Author Response File: Author Response.pdf