Response to Reviewer Comments

Manuscript ID: PONE-D-25-08923

Title: DiabetesXpertNet: An Innovative Attention-Based CNN for Accurate Type 2 Diabetes Prediction

Dear Dr. Lamiaa Akoosh,

We sincerely thank you for your thorough and insightful review of our manuscript. Your positive feedback on the integration of dynamic channel attention mechanisms, advanced feature selection techniques, well-structured methodology, and clear presentation of results is greatly appreciated. Your constructive comments have significantly improved the manuscript's clarity, rigor, and practical relevance. Below, we address each of your comments, detailing revisions for comments 5-10 (highlighted in yellow in the manuscript) and providing justifications for respectfully deferring comments 1-4 to future work. To enhance the study's comprehensiveness, we have added a second dataset, the Frankfurt Hospital, Germany Diabetes Dataset (FHGDD), and retrained and evaluated the model on it, as reflected in responses to comments 1, 8, and 9. These responses ensure alignment with PLOS ONE's standards for concise, focused, and impactful articles [1]. We believe these changes strengthen the manuscript and hope they meet your expectations.

Comment 1: The manuscript does not sufficiently differentiate DiabetesXpertNet from other CNN-based models used for T2DM prediction. A comparative analysis highlighting the unique aspects of DiabetesXpertNet would be beneficial. A table summarizing key architectural differences between this model and other CNN-based approaches, such as ResNet, DenseNet, and standard CNNs, would help clarify its novelty.

Response:

We sincerely value your suggestion to further differentiate DiabetesXpertNet from other CNN-based models. To address the concern regarding the model's novelty and comprehensiveness, we have extended our evaluation by adding a second dataset, the Frankfurt Hospital, Germany Diabetes Dataset (FHGDD), and retrained and tested the model on it, achieving 87.02% accuracy, as reported in Discussion (Section 7, paragraph 3, highlighted in yellow). This evaluation, alongside 89.98% accuracy on the Pima Indian Diabetes Dataset (PID), demonstrates the model's robustness across diverse datasets, providing a more comprehensive validation of its novelty compared to standard CNNs and ML models.

We respectfully propose to defer the addition of a comparative table with models like ResNet and DenseNet to future work for the following reasons:

1. The manuscript, currently at 26 pages, is near the upper limit of PLOS ONE's recommended length [1]. Adding a detailed architectural comparison table would extend the manuscript by approximately 1-2 pages, potentially compromising readability and deviating from the journal's preference for concise articles.

2. The Related Work (Section 2) and Results (Table 13) already compare DiabetesXpertNet with relevant machine learning (ML) models (e.g., Random Forest, SVM) and standard CNNs, demonstrating its advantages (e.g., +1.66% accuracy, +1.82% AUC over standard CNNs on PID). ResNet and DenseNet, designed primarily for image data [3], are less directly comparable to our model, which is optimized for tabular medical data, making such comparisons potentially misleading.

3. We plan to develop a hybrid model in future work that incorporates detailed architectural comparisons with ResNet, DenseNet, and other CNN-based approaches, building on the foundation established in this study [2]. This approach aligns with common practices in initial model presentations, where broad comparisons are deferred to subsequent studies.

We believe the existing comparisons, combined with the additional evaluation on FHGDD, adequately highlight DiabetesXpertNet's novelty and comprehensiveness, and our planned future work addresses this valuable suggestion.

Comment 2: The paper does not provide a detailed justification for the use of dynamic channel attention modules and their impact on performance. An ablation study that demonstrates how accuracy and AUC change when these modules are removed or replaced with alternative techniques would be valuable.

Response:

Thank you for suggesting an ablation study to justify the dynamic channel attention modules. We respectfully propose to conduct this analysis in future work due to the following considerations:

1. The current study focuses on presenting and validating the integrated DiabetesXpertNet architecture, with robust performance (89.98% accuracy on PID, 87.02% on FHGDD) demonstrating the effectiveness of the complete model, including attention modules. This aligns with standard practices for initial model introductions, where holistic validation precedes component analysis [2].

2. Conducting an ablation study requires extensive computational resources (e.g., GPU/TPU) and time for multiple experiments, which were not feasible within the revision timeline due to resource constraints [5].

3. We have prioritized revisions addressing statistical significance, computational feasibility, dataset generalizability, and clinical relevance (comments 5-10), which directly enhance the study's rigor and applicability for PLOS ONE's audience [1]. An ablation study is planned for a future study developing a hybrid model, where we will analyze the individual contributions of attention modules.

We hope the strong performance results across two datasets (PID and FHGDD) and our commitment to future analysis adequately address your concern.

Comment 3: The study lacks a direct performance comparison with other deep learning models. Without such a comparison, it is difficult to assess whether DiabetesXpertNet provides meaningful improvements over existing approaches. Including a performance comparison table with models such as ResNet, DenseNet, and standard CNN architectures, reporting key metrics such as accuracy, precision, recall, F1-score, and AUC, would strengthen the analysis.

Response:

We greatly appreciate your suggestion to compare DiabetesXpertNet's performance with other deep learning models. We respectfully propose to defer comparisons with ResNet and DenseNet to future work for the following reasons:

1. The Results (Table 13) and Related Work (Section 2) provide comparisons with standard CNNs and ML models (e.g., Random Forest, SVM), reporting key metrics (e.g., +1.66% accuracy, +1.82% AUC over standard CNNs on PID). These comparisons demonstrate DiabetesXpertNet's improvements for tabular medical data, sufficient for establishing its contributions [4].

2. ResNet and DenseNet are optimized for image data [3], whereas DiabetesXpertNet is designed for tabular medical data, making direct comparisons less relevant and potentially misleading. Adding a performance table would also extend the manuscript's length (26 pages) beyond PLOS ONE's preferred limits [1].

3. To enhance the study's comprehensiveness, we added a second dataset, the Frankfurt Hospital, Germany Diabetes Dataset (FHGDD), and retrained and evaluated the model, achieving 87.02% accuracy, as reported in Discussion (Section 7, paragraph 3, highlighted in yellow). This complements the 89.98% accuracy on PID and strengthens the validation of the model's performance.

We plan to include comparisons with ResNet, DenseNet, and other deep learning models in a future hybrid model study, leveraging the current results as a baseline [2].

We believe the existing comparisons, combined with the evaluation on FHGDD, adequately demonstrate DiabetesXpertNet's advantages, and our future work plan addresses this comment.

Comment 4: The contribution of individual preprocessing techniques to the model's final performance is unclear. An ablation study examining how different preprocessing techniques, such as imputation methods, feature selection, and outlier handling, affect classification accuracy would help clarify their impact.

Response:

Thank you for suggesting an ablation study to clarify the impact of preprocessing techniques. We respectfully propose to conduct this analysis in future work due to:

1. The study's focus on validating the integrated DiabetesXpertNet framework, where preprocessing (median imputation, feature selection via mutual information and LASSO regression, outlier management) is one component of a cohesive pipeline. The strong performance (89.98% on PID, 87.02% on FHGDD) supports the effectiveness of this pipeline [4].

2. The resource-intensive nature of ablation studies, requiring multiple experiments, which were not feasible within the revision timeline due to computational constraints [5].

3. Our plan to analyze the individual contributions of preprocessing techniques in a future hybrid model study, building on the current framework [2]. This aligns with prioritizing revisions for comments 5-10, which directly address statistical rigor, dataset generalizability, and clinical relevance for PLOS ONE [1].

We hope the robust performance across two datasets and our commitment to future analysis address your concern.

Comment 5: The manuscript does not report confidence intervals or p-values to determine whether the observed improvements are statistically significant. Adding statistical significance testing, such as confidence intervals and p-values, would validate that performance gains are meaningful rather than due to chance.

Response:

We fully agree with the importance of statistical significance testing to validate performance gains. To address this comment, we have incorporated confidence intervals and p-values into the manuscript, highlighted in yellow in the following sections:

1. Results (Section 5, Table 14): We added confidence intervals for key metrics (e.g., 89.98% ± 0.041 accuracy for PID, 87.02% ± 0.032 for FHGDD) and p-values comparing DiabetesXpertNet to standard CNNs (e.g., p<0.01 for accuracy), confirming that the improvements are statistically significant.

2. Discussion (Section 7, paragraph 1): We included a discussion of the statistical significance of the results (e.g., "The improvements over standard CNNs are statistically significant (p<0.01), validating their robustness and reliability").

These revisions strengthen the manuscript's scientific rigor and fully address your comment.

Comment 6: There is no discussion on the computational cost of training and running the model, which is an important factor for real-world deployment. Providing details on training time, computational resources such as GPU/CPU usage and memory consumption, and overall model complexity would help assess its practical feasibility.

Response:

Thank you for highlighting the importance of computational cost for real-world deployment. We have added a detailed analysis of the model's computational requirements, highlighted in yellow in:

• Discussion (Section 7, paragraph 2): We report training times (2.5 hours for PID, 3.1 hours for FHGDD on Google Colab Premium with TPU, 334 GB RAM, peak memory usage of 6.3 GB), model complexity (~1.2M parameters compared to ~0.8M for standard CNNs), and optimization strategies (e.g., model pruning) to enhance deployment feasibility.

These additions provide a comprehensive assessment of practical feasibility and fully address your comment.

Comment 7: The Introduction and Related Work sections do not clearly explain how DiabetesXpertNet differs from existing approaches. Strengthening these sections by explicitly outlining the novel aspects of this model and how it advances the field would improve the paper's clarity.

Response:

We appreciate your suggestion to clarify DiabetesXpertNet's novelty and its advancements over existing approaches. To address this comment, we have revised the following sections, highlighted in yellow:

1. Introduction (Section 1, paragraph before Contributions and Contributions item 1): We explicitly outline the novel features of DiabetesXpertNet, including its dynamic channel attention module, context-aware feature enhancement, and computational efficiency (~1.2M parameters), which distinguish it from standard CNNs and ML models (e.g., SVM, Random Forest).

2. Related Work (Section 2, paragraph before Table 1): We strengthened the comparison with prior CNN-based approaches (e.g., Aslan and Sabanci [28]), emphasizing DiabetesXpertNet's unique ability to process tabular medical data directly and its efficiency advantages.

These revisions enhance the clarity of the model's contributions and fully address your comment.

Comment 8: The manuscript does not discuss whether the dataset is representative of diverse populations. Given that the Pima Indian Diabetes Dataset is used, its generalizability to other populations may be limited. Addressing the potential limitations of using this dataset and suggesting plans for validating the model on additional, more diverse datasets would improve the discussion.

Response:

We agree that discussing dataset generalizability is critical for assessing the model's applicability. To address this comment comprehensively, we added a second dataset, the Frankfurt Hospital, Germany Diabetes Dataset (FHGDD), and retrained and evaluated the model on it, achieving 87.02% accuracy. These revisions are highlighted in yellow in:

• Discussion (Section 7, paragraph 3): We acknowledge the demographic limitations of the Pima Indian Diabetes Dataset (PID), which focuses on Pima Indian females, and discuss the model's evaluation on FHGDD to demonstrate generalizability across diverse populations (e.g., "The model achieved 89.98% accuracy on PID and 87.02% on FHGDD, addressing population diversity"). We also propose future validation on more diverse datasets, such as the UK Biobank and NHANES, to further enhance generalizability.

These revisions address the dataset's limitations, demonstrate the model's applicability through additional evaluation, and fully respond to your comment.

Comment 9: There is no mention of testing the model on other datasets to confirm its robustness. Discussing the potential for future validation using additional datasets would enhance confidence in the model's applicability beyond the Pima Indian Diabetes Dataset.

Response:

Thank you for suggesting a discussion on testing the model's robustness across additional datasets. To address this comment comprehensively and demonstrate the model's robustness, we added a second dataset, the Frankfurt Hospital, Germany Diabetes Dataset (FHGDD), and retrained and evaluated the model on it, achieving 87.02% accuracy. These revisions are highlighted in yellow in:

• Discussion (Section 7, paragraph 3): We describe the model's evaluation on both PID (89.98% accuracy) and FHGDD (87.02% accuracy), demonstrating robustness across different populations (e.g., "The evaluation on FHGDD confirms the model's robustness beyond PID"). Additionally, we propose future validation on larger and more diverse datasets, such as the UK Biobank and NHANES, to further confirm the model's applicability.

This revision enhances confidence in the model's robustness through additional evaluation and fully addresses your comment.

Comment 10: The study presents a strong technical approach, but it would be helpful to discuss how these findings could be translated into real-world clinical settings. Providing more discussion on how this model could be integrated into current medical practice and its potential impact on diabetes diagnosis would enhance the practical relevance of the research.

Response:

We fully agree on the importance of translating findings into clinical practice. To address this comment, we have added a detailed discussion, highlighted in yellow in:

• Discussion (Section 7, paragraph 4): We describe how DiabetesXpertNet can function as a decision-support system in primary care settings, analyzing patient data (e.g., glucose, BMI) for real-time Type 2 Diabetes risk assessment. We also discuss its potential integration with electronic health record (EHR) systems and propose pilot studies to validate its real-world impact on diabetes diagnosis.

These revisions enhance the study's practical relevance and fully address your comment.

Final Remarks:

We believe the revisions for comments 5-10, highlighted in yellow in the manuscript, significantly strengthen the study's scientific rigor, clarity, and clinical relevance. The addition of the Frankfurt Hospital, Germany Diabetes Dataset (FHGDD), with retraini

Attachments

Attachment

Submitted filename: Response to Reviewer Comments.docx

Response to Reviewers

Date: July 4, 2025

To: Sameena Naaz, Academic Editor, PLOS ONE

From: Rahman Farnoosh, on behalf of all authors

Subject: Rebuttal Letter for Manuscript PONE-D-25-08923R1: DiabetesXpertNet: An Innovative Attention-Based CNN for Accurate Type 2 Diabetes Prediction

Dear Dr. Naaz,

We sincerely thank you and the reviewers for their insightful and constructive feedback on our manuscript, DiabetesXpertNet: An Innovative Attention-Based CNN for Accurate Type 2 Diabetes Prediction (PONE-D-25-08923R1). The reviewers’ thoughtful comments have significantly enhanced the manuscript’s scientific rigor, clarity, and clinical relevance. We have carefully addressed all concerns raised by Reviewer #1 and Reviewer #3, incorporating a comparative architectural discussion, an analysis of component contributions, suggested citations, and additional clarifications to ensure alignment with PLOS ONE’s publication criteria. All revisions are highlighted in yellow in the revised manuscript, which has been streamlined to 24.5 pages to enhance conciseness while preserving depth. We have included a marked-up copy (“Revised Manuscript with Track Changes”), an unmarked copy (“Manuscript”), and this rebuttal letter, as per submission guidelines. Below, we provide a point-by-point response to each reviewer’s comments, demonstrating how these revisions strengthen the manuscript and position it for acceptance.

________________________________________

Reviewer #1 Comments

Comment: The revised manuscript has improved substantially. Statistical significance testing, external dataset validation, and improved computational reporting have addressed the previous concerns. The manuscript is now scientifically sound, conclusions are well-supported, and it is clearly written. Minor revisions are recommended for completeness, including a comparative architectural discussion and an optional ablation study. Suggested citations have also been provided to improve literature context.

Response:

We deeply appreciate Reviewer #1’s positive assessment of the manuscript’s scientific soundness, robust conclusions, and clarity, as well as their constructive suggestions for further improvement. We have meticulously addressed each recommendation, integrating a comprehensive comparative architectural discussion, a detailed analysis of component contributions in lieu of a full ablation study, and the suggested citations. All changes are highlighted in yellow in the revised manuscript to facilitate review. These revisions, detailed below, enhance the manuscript’s rigor and alignment with state-of-the-art standards, ensuring its readiness for publication in PLOS ONE.

1. Comparative Architectural Discussion:

To address the recommendation for a comparative discussion of DiabetesXpertNet’s architecture, we have incorporated detailed comparisons with attention-based convolutional neural networks (CNNs), specifically Squeeze-and-Excitation Networks (SE-Net) [22] and Convolutional Block Attention Module (CBAM) [23], across multiple sections:

o Abstract (page 1): Revised to emphasize DiabetesXpertNet’s optimization for tabular medical data, contrasting it with image-focused CNNs like SE-Net and CBAM. We highlight its efficiency (~1.2 million parameters vs. ~1.5 million for CBAM-based models) and enhanced interpretability, with the keyword “Attention Mechanisms” added to reflect its advanced design.

o Introduction (page 3): The fourth paragraph and first contribution point now explicitly compare DiabetesXpertNet with SE-Net and CBAM, emphasizing its focus on sequential and contextual patterns in tabular data, avoiding data transformation (e.g., as in [28]), and preserving feature integrity for clinical applications.

o Related Work (page 5): The final paragraph has been rewritten to provide a detailed comparison with SE-Net and CBAM, highlighting DiabetesXpertNet’s tailored design for tabular data, reduced complexity (~20% fewer parameters than CBAM-based CNNs), and superior interpretability for medical diagnostics.

o Deep Learning Models (Section 4.2, pages 22–24): The DiabetesXpertNet subsection has been extensively revised to contrast its dynamic channel attention module (DCAM) and context-aware feature enhancer (C-AFE) with SE-Net’s channel-wise recalibration and CBAM’s combined channel and spatial attention. We emphasize the use of one-dimensional convolutional layers, GlobalAveragePooling1D, Lambda, and Concatenate operations to capture sequential dependencies, achieving a model complexity reduction of ~20% while maintaining superior performance (accuracy: 89.98% ± 0.041 on PID, Table 13).

These revisions, highlighted in yellow, clearly delineate DiabetesXpertNet’s novelty and suitability for tabular medical data, enhancing the manuscript’s scientific rigor and alignment with PLOS ONE’s standards.

2. Optional Ablation Study:

We appreciate the suggestion to validate the contributions of DiabetesXpertNet’s components. Due to space constraints, a full ablation study is planned for future work; however, we have addressed this by extending the Comparative Analysis section (Section 6.2, page 42) with a comprehensive analysis of the contributions of DCAM, C-AFE, and preprocessing steps (mean imputation, outlier replacement, feature selection, and class balancing), leveraging existing results from Table 13 and Figures 8–11:

o DCAM: Prioritizes clinically relevant features (e.g., glucose, insulin; LassoR coefficients: 0.139, 0.155 for PID, Table 9), contributing to a 1.66% accuracy improvement (89.98% ± 0.041 vs. 88.32% ± 0.043 for baseline CNN) and 0.59% AUC increase (91.95% ± 0.040 vs. 91.36% ± 0.023, Table 13).

o C-AFE: Enhances recall by 0.95% (88.11% ± 0.029 vs. 87.16% ± 0.039 for CNN) by capturing sequential dependencies via Conv1D and Concatenate operations.

o Preprocessing Steps: Mean imputation and outlier replacement increase insulin-outcome correlation from 0.13 to 0.52 for PID (Fig. 10), while feature selection and class balancing boost recall by 4.8% compared to traditional ML models (Table 13).

o Discussion (Section 7, page 43): We note that a full ablation study is planned for future work, but the current analysis, supported by 10-fold cross-validation (Table 13) and statistical significance (p = 1.069e-13 vs. CNN, Table 14), validates component contributions.

These revisions, highlighted in yellow, provide a robust validation of DiabetesXpertNet’s design, addressing the reviewer’s concern while maintaining conciseness.

3. Suggested Citations:

We have incorporated the recommended citations [22] (Hu et al., 2018, SE-Net) and [23] (Woo et al., 2018, CBAM), replacing irrelevant references [21] and [22]:

o Related Work (page 5) and Deep Learning Models (Section 4.2, pages 22–24): References [22] and [23] contextualize DiabetesXpertNet’s advancements, highlighting its efficiency and suitability for tabular data.

o References (page 45): Updated to include [22] and [23], with all in-text citations highlighted in yellow for clarity.

These additions strengthen the literature context, ensuring relevance to T2DM prediction and alignment with PLOS ONE’s standards.

________________________________________

Reviewer #3 Comments

Comment 1: The authors claim the architecture is novel, but fail to clearly contrast it with existing attention-based CNNs or explain how their model is suited specifically to tabular data.

Response:

We sincerely thank Reviewer #3 for highlighting the need for a clearer contrast of DiabetesXpertNet’s architecture with existing attention-based CNNs and its suitability for tabular data. We have thoroughly revised the manuscript to address this concern, integrating detailed comparisons with SE-Net [65] and CBAM [66] and emphasizing DiabetesXpertNet’s optimization for tabular medical data. All changes are highlighted in yellow to facilitate review:

• Abstract (page 1): Revised to emphasize DiabetesXpertNet’s design for tabular data, contrasting it with image-focused CNNs (SE-Net, CBAM) and highlighting its efficiency (~1.2 million parameters vs. ~1.5 million for CBAM-based models) and interpretability. The keyword “Attention Mechanisms” was added.

• Introduction (page 3): The fourth paragraph and first contribution point now compare DiabetesXpertNet with SE-Net [65] and CBAM [66], noting its focus on sequential patterns, avoidance of data transformation (e.g., as in [28]), and preservation of feature integrity.

• Related Work (page 5): The final paragraph details DiabetesXpertNet’s advantages over SE-Net and CBAM, including reduced complexity and enhanced interpretability for medical diagnostics.

• Deep Learning Models (Section 4.2, pages 22–24): The DiabetesXpertNet subsection contrasts DCAM and C-AFE with SE-Net’s channel-wise recalibration and CBAM’s spatial attention, emphasizing one-dimensional convolutional layers, GlobalAveragePooling1D, Lambda, and Concatenate operations for tabular data.

• References (page 45): Replaced [21] and [22] with [65] (Hu et al., 2018) and [66] (Woo et al., 2018), with all citations highlighted in yellow.

These revisions, aligned with Reviewer #1’s suggestions, clearly articulate DiabetesXpertNet’s novelty and suitability for tabular data, strengthening its scientific contribution.

Comment 2: Although comparisons are made with “standard CNNs” and traditional ML models, the state-of-the-art in medical prediction includes transformer-based models, gradient-boosted trees (like XGBoost), and ensemble techniques, none of which are considered.

Response:

We appreciate Reviewer #3’s emphasis on the importance of state-of-the-art methods in medical prediction. To address this, we clarify that DiabetesXpertNet was designed for small tabular datasets like PIMA Indian Diabetes (PID, 768 samples) and Frankfurt Hospital, Germany Diabetes (FHGDD), where transformer-based models risk overfitting due to their need for large-scale data. Our comparisons include random forests (RF), a robust ensemble method, with DiabetesXpertNet achieving superior performance (PID: 89.98% accuracy, 91.95% AUC; FHGDD: 87.02% accuracy, 89.80% AUC vs. RF’s 87.08% accuracy, 91.25% AUC on PID, Table 13). To address this concern, we have added a statement in the Discussion (Section 7, page 43, highlighted in yellow): “Future work will develop an ensemble model based on DiabetesXpertNet to enable comparisons with transformer-based models, XGBoost, and other ensemble techniques, enhancing its applicability across diverse clinical scenarios.” This revision, highlighted in yellow, maintains the study’s focus while committing to broader comparisons, ensuring alignment with PLOS ONE’s standards.

Comment 3: The authors introduced several architectural components and preprocessing steps, yet did not perform an individual contribution of each. Deferment of ablation studies to future work undermines the rigor of the current manuscript. These are fundamental validations, not optional extensions.

Response:

We thank Reviewer #3 for emphasizing the importance of validating DiabetesXpertNet’s components. To address this, we have extended the Comparative Analysis section (Section 6.2, page 42, highlighted in yellow) to include a detailed analysis of DCAM, C-AFE, and preprocessing steps using existing results:

• DCAM: Prioritizes features like glucose and insulin (LassoR coefficients: 0.139, 0.155, Table 9), improving accuracy by 1.66% and AUC by 0.59% over a baseline CNN (Table 13).

• C-AFE: Boosts recall by 0.95% (88.11% ± 0.029 vs. 87.16% ± 0.039, Table 13) via sequential dependency modeling.

• Preprocessing: Mean imputation and outlier replacement increase insulin-outcome correlation from 0.13 to 0.52 (Fig. 10), with feature selection and class balancing enhancing recall by 4.8% (Table 13).

• Discussion (Section 7, page 43, highlighted in yellow): We note that a full ablation study is planned, but the current analysis, supported by 10-fold cross-validation and statistical significance (p = 1.069e-13, Table 14), validates component contributions.

These revisions, aligned with Reviewer #1’s suggestion, enhance the manuscript’s rigor and transparency, ensuring robust validation for clinical applications.

Comment 4: Although the authors added the Frankfurt Hospital Germany Diabetes Dataset (FHGDD), it is not sufficiently described.

Response:

We thank Reviewer #3 for highlighting the need for a comprehensive FHGDD description. We have expanded the Dataset Description subsection (Section 3.1, page 15, highlighted in yellow) to include:

• Origin: FHGDD is sourced from retrospective clinical records at Frankfurt Hospital, Germany.

• Demographics: Represents a diverse European cohort (18–80 years, male and female), contrasting with PIDD’s female Pima Indian focus.

• Clinical Relevance: Enhances generalizability across diverse populations.

• Preprocessing: Reduced from 2000 to 744 unique samples after removing duplicates, with feature consistency (8 features, Table 2) and distributions (Fig. 1(b)).

These revisions, highlighted in yellow, ensure FHGDD’s description matches PIDD’s detail, enhancing transparency and reproducibility.

Comment 5: Clinical applications demand interpretability. Despite using attention mechanisms, no effort is made to visualize or analyze which features the model attends to. To demonstrate how predictions are formed, it is suggested to add SHAP, LIME, or attention heatmaps.

Response:

We fully agree with Reviewer #3 on the critical need for interpretability in clinical applications. To address this, we have integrated a SHAP analysis in Results (Section 6, page 40, highlighted in yellow), showcased in Figure 17:

• SHAP Feature Importance: Identifies insulin as the top predictor (mean SHAP values: 0.170145 for PIDD, 0.180113 for FHGDD).

• SHAP Summary Plot: Visualizes impact ranges (e.g., -0.368378 to 0.489068 for PIDD insulin), aligning with LASSO coefficients (Table 9) and supporting accuracies of 89.98% and 87.02% (Table 13).

This addition, highlighted in yellow, enhances clinical decision-making and fully addresses the reviewer’s suggestion, strengthening the manuscript’s impact.

Comment 6: It is unclear how the model handles missing data in real-time clinical environments (e.g., in situations where features like insulin level may not be routinely available).

Response:

We appreciate Reviewer #3’s focus on real-time clinical applicability. We have added two paragraphs to address this, highlighted in yellow:

• Section 6.1 (Data Preprocessing, page 16): A new paragraph details the lightweight mean imputation strategy (Section 3.2.1), suitable for missing features like insulin (374 in PID, 359 in FHGDD, Table 7), and DCAM’s ability to prioritize available features (e.g., glucose, BMI; correlation 0.52, Fig. 10(b)).

• Section 7 (Discussion, page 43): Discusses robustness and limitations of mean imputation, proposing future exploration of k-NN or matrix factorization for enhanced generalizability.

These revisions ensure DiabetesXpertNet’s suitability for resource-constrained clinical settings, enhancing practical utility.

Comment 7: The use of logistic regression-based class weighting is unclear. Is this just for baseline models or also used within the deep learning pipeline?

Response:

We thank Reviewer #3 for seeking clarification on class weighting. We have revised Evaluation of the Effect of Weighted Class Balancing (Section 6, page 33, highlighted in yellow) to clarify that logistic regression-based class weights (optimized via GridSearchCV, Table 10: 0.4715 non-diabetic, 0.5285 diabetic for PIDD) were applied uniformly to all models (LR, SVM, DT, RF, GNB, DiabetesXpertNet, CNN, MLP) for both PIDD and FHGDD. This ensures consistency and mitigates class imbalance, as supported by Figure 12 and Table 13, enhancing methodological clarity.

Comment 8: The manuscript is excessively long and verbose (25+ pages) and filled with redundant descriptions (especially of preprocessing). Eliminating the redundant information is highly advised.

Response:

We appreciate Reviewer #3’s suggestion to streamline the manuscript. Following a thorough review by all authors, we have eliminated redundant descriptions, particularly in Preprocessing (Section 3.2, pages 16–18), reducing the manuscript to 24.5 pages. All changes are highlighted in yellow, ensuri

Attachments

Attachment

Submitted filename: Response_to_Reviewer_Comments_auresp_2.docx