Explainable Machine Learning for Predicting Dengue Recovery Duration: Insights from Multi-Center Clinical Data

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper addresses a relevant problem at the intersection of machine learning and public health, and I appreciate the authors' emphasis on interpretability over mere predictive accuracy. That said, several aspects warrant deeper consideration. The choice of models — Logistic Regression, Decision Tree, Random Forest, and a basic feed-forward Neural Network — is rather conservative. Given that the study positions itself within the explainable AI space, the omission of gradient boosting methods such as XGBoost or LightGBM, which pair naturally with SHAP-based explanations and routinely outperform the selected models on tabular clinical data, is a notable gap. Furthermore, SHAP itself is conspicuously absent from the XAI toolkit despite being the most widely adopted and theoretically grounded framework for model-agnostic explanation; relying solely on PDP, ICE, and LIME leaves the interpretability analysis incomplete, particularly at the global level.

The multi-centre dataset is a commendable aspect of this work, yet a sample of only 100 patients across multiple hospitals raises serious concerns about statistical power and the reliability of cross-validated performance estimates. With such a small cohort, ten-fold cross-validation folds contain very few samples, and the decision to forgo hyperparameter tuning — while understandable given the interpretability focus — makes it difficult for readers to judge whether the models have been given a fair chance before their explanations are scrutinised. The paper would benefit from reporting standard classification metrics such as accuracy, F1-score, and AUC alongside the explainability plots to allow the community to contextualise the explanations against actual model competence.

From an EEE-to-clinical translation standpoint, I would encourage the authors to discuss how this framework could be embedded within a real-time clinical decision support system — for instance, as an edge-computing module in hospital monitoring infrastructure. Addressing latency, computational constraints, and how clinicians would interact with LIME outputs at the bedside would significantly strengthen the paper's practical contribution and better bridge the engineering and healthcare domains.

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The work provide an explainable artificial intelligence framework for predicting dengue recovery outcomes using multicentre clinical data from Pakistan. It’s related to dengue fever continues to impose a major burden on healthcare systems in endemic regions, while interpretable AI approaches are increasingly important for trustworthy clinical decision-making. The integration of predictive modeling with explainability techniques is a good strength of the study. Based on my evaluation folioing aspects require refinement. Major revision is recommended.

1. The introduction should more clearly establish the clinical and public health significance of predicting dengue recovery trajectories, particularly in resource-constrained healthcare systems.
2. The manuscript should provide stronger theoretical rationale for selecting explainability techniques such as PDP, ICE, and LIME in the context of clinical decision support.
3. The manuscript should discuss whether differences in hospital infrastructure or treatment protocols may have influenced recovery outcomes across participating centers.
4. The preprocessing pipeline requires more comprehensive explanation, including normalization methods, feature encoding, class imbalance management, and validation strategy.
5. The selection of Logistic Regression, Decision Tree, Random Forest, and Neural Network models is appropriate; however, clearer justification for model comparison is needed.
6. The manuscript should provide more detailed performance metrics including sensitivity, specificity, F1-score, calibration, and AUC values for all models.
7. The identification of age, platelet count, hospital type, education level, and blood group as influential predictors is clinically relevant, but the biological or healthcare-system implications should be discussed more deeply.
8. The influence of socioeconomic variables such as education level on recovery outcomes should be interpreted carefully to avoid oversimplification or confounding bias.
9. The patient-level variability observed in recovery trajectories is highly important and should be linked more explicitly to personalized medicine and individualized risk stratification concepts.
10. The application of PDP and ICE plots is valuable, though the manuscript should discuss their limitations in highly correlated feature spaces.
11. The use of LIME improves local interpretability, but the manuscript should critically evaluate the stability and reproducibility of local explanations.
12. The manuscript should discuss whether the explainability outputs are understandable and actionable for clinicians in routine healthcare settings.
13. The introduction should supplement the current research and enrich the content of the paper. For example: https://doi.org/10.1016/j.compbiolchem.2020.107200, https://doi.org/10.1186/s12859-020-03682-4

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper is about a very important topic by applying xplainable machine learning to dengue recovery prediction using related data from Pakistan. The study is relevant because interpretability is important in healthcare and decision-making. Yet, the paper still needs some improvement before being accepted. The dataset is very small, the model evaluation is not sufficiently detailed, and several language problems, typos, and formatting issues. The authors should improve proofreading, strengthen the methodological explanation, report clearer performance results. They shall avoid strong clinical claims that are not fully supported by the available data.

Some points to focus on:

1. The paper has an interesting topic, but the writing needs careful proofreading. I found many typos and broken words, such as “invistigates,” “behavoir,” “speicific,” dis-ase and similar errors. This shows that the authors are not meticlious.
2. The dataset is very small, with only 100 patients. This weakens the strength of the machine learning results and should be discussed more clearly.
3. The paper says it uses multi-centre data, but the data description is not detailed enough. The authors should explain how many patients came from each hospital or region.
4. The model evaluation is incomplete. The paper discusses interpretability, but it does not clearly report enough performance metrics such as accuracy, precision, recall, F1-score, ROC-AUC, or confusion matrices.
5. The figures are useful, but some of them are hard to read. The authors should improve the quality, labels, and captions of the figures.
6. The conclusion is reasonable, but the claims should be softer. With this small dataset, the authors should avoid making strong clinical or public health claims.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have done a good job responding to the earlier feedback by engaging with each point. Reframing the problem as a regression task and reporting cross-validated RMSE, MAE, and R² was the right call — it gives readers a much clearer picture of where the models stand.
I also found the expanded limitations section refreshingly honest; acknowledging the small sample size and the absence of SHAP and gradient boosting upfront, rather than burying these issues, actually strengthens the paper's credibility.
The new clinical interpretation table is a nice touch that helps ground the statistical patterns in something a clinician can relate to. That said, I would still strongly encourage the authors to pursue SHAP-based analysis and external validation on a larger cohort in their next study — doing so would turn what is currently a solid exploratory piece into a genuinely impactful one.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

No more comments. Revised version should accepted. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

I believe the revision is done well enough when compared with the first version. The authors responded to the major methodological weakness by changing the modeling from vague recovery outcome classification toward recovery duration regression.

They also improved dataset transparency, performance reporting, and limitations. Yet, I want still ask for minor correction, mainly to clean the continuous versus binary outcome inconsistency and they shall correct the risk category explanation in Figure 11.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

We thank the reviewer for the positive assessment of the revised manuscript and for the helpful suggestion. In response, we carefully reviewed the manuscript to ensure consistency between the continuous recovery duration outcome and its interpretation throughout the paper. Additionally, we revised the explanation of the risk categories associated with Figure 11, updated the figure accordingly, and clarified its interpretation on Page 15 of the revised manuscript. We believe these changes have improved the clarity and consistency of the presentation.