Dual-Layer PSO-Enhanced Federated Heterogeneous Data Fusion for Hemodialysis Complication Prediction

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Comment 1. The title, abstract, and conclusions state that the study predicts hemodialysis-related complications, including hypotension and discomfort events; however, the actual target variable is defined only as normal versus abnormal blood pressure, using a threshold below 90 mmHg. The authors should clearly define the clinical endpoint, the prediction horizon, and the time at which the prediction is generated relative to the onset of the event. They should also report the number of unique patients, hemodialysis sessions, and records; the class distribution; the data collection period; the inclusion and exclusion criteria; and the approach used to handle missing data. Since real patient data were used, the manuscript should also include information on ethics committee approval, informed consent, or a justification for a waiver of consent.

Comment 2. The authors state that records from two files were merged, randomly shuffled, and distributed among clients; however, they do not specify whether the split was performed at the patient level or at the level of individual hemodialysis sessions. Records from the same patient or session may therefore have been included simultaneously in the training, validation, and test sets, which could substantially overestimate model performance. A patient-level or session-level split should be applied. The local validation set, the server-side validation set used for PSO optimization, and the independent final test set should also be clearly separated. The procedure used for the 11-fold cross-validation requires a more detailed description.

Comment 3. The proposed method is compared mainly with FedAvg and two single-layer PSO variants developed by the authors. To demonstrate a meaningful scientific advantage, the evaluation should include contemporary methods designed for non-IID data, such as FedProx, SCAFFOLD, FedNova, or FedOpt, as well as a centralized model and locally trained models. The experiments should be repeated across multiple independent runs with different random seeds, and the results should be reported as mean ± standard deviation or with 95% confidence intervals. Statistical comparisons, sensitivity, specificity, PR-AUC, and a quantitative assessment of the computational and communication costs introduced by the dual-layer PSO framework should also be provided.

Comment 4. The abstract reports an accuracy of 90.56%, whereas the comparative table reports 93%, while the description of Figure 7 states that accuracy stabilizes at approximately 75–80%. In addition, the authors claim that most results from the 11-fold cross-validation exceed 90%, although Figure 8 shows values above 90% for only a limited number of folds. The model description specifies a sigmoid output for binary classification, whereas Figure 3 shows a final layer with two outputs, which would be more consistent with a softmax classifier. All numerical results, model architecture details, and terminology should therefore be carefully reconciled. The authors should also acknowledge that federated learning alone does not guarantee privacy, since exchanged model parameters may remain vulnerable to data reconstruction or inference attacks. The manuscript should describe the assumed threat model and indicate whether secure aggregation, differential privacy, or other protection mechanisms were used. In the absence of such mechanisms, claims of complete data protection should be substantially moderated.

Author Response

Comment 1

Response: We appreciate this important observation. In the revised manuscript:

• We clarified that the clinical endpoint is intradialytic hypotension, operationalized as systolic blood pressure < 90 mmHg as measured bedside.
• The prediction horizon is defined as the onset of hypotension as measured by the Hemodialysis machine per minute base using CNN model for the testing results.  Early prediction before the onset of abnormal event will require LSTM/Transformer model being adapted with the PSO methodology.  They will be implemented in the future work.  Detailed discussion is stated in section 4.4.6 of p.24.
• We now report the dataset statistics:
• The data collection period and inclusion/exclusion criteria (adult patients undergoing maintenance hemodialysis, excluding incomplete records) are specified. Details are listed in section 3.2 of p.6.
• Missing data were handled using mean imputation for continuous variables and mode imputation for categorical variables, which is now described in Section 3.2 of p.6.
• We added a statement on ethics approval (Taichung Veterans General Hospital IRB No. SE20056A) and clarified that informed consent was waived due to retrospective anonymized data.

Comment 2

Records may have been split at the session level rather than patient level, risking data leakage. Validation/test separation and cross-validation procedure need clarification.

Response: We agree that patient-level separation is essential. In the revision:

• We ensured that splits were performed at the patient level, so records from the same patient never appear in both training and test sets. Details of the split can be seen in Fig. 4 and its description in section 3.4 of p.9 of the revised paper.
• We clarified the local validation set (used for client-side PSO), the server validation set (used for aggregation optimization), and the independent test set (used for final evaluation) in in Fig. 4 and its description in section 3.4 of p.9 of the revised paper.
• The 11-fold cross-validation procedure is now described step by step in section 4.4 of p.18, including fold assignment, rotation, and averaging of metrics statistics.

Comment 3

Comparison is limited to FedAvg and PSO variants. Evaluation should include FedProx, SCAFFOLD, FedNova, FedOpt, centralized and local models. Results should be reported with statistical measures and additional metrics (sensitivity, specificity, PR-AUC, computational/communication costs).

Response: We have expanded the experimental evaluation:

• Added FedProx, SCAFFOLD as baselines in Fig. 7 of p.17 and statistical significance results against the proposed PSO based methods are reported in Table 3 in p.18.
• Included centralized CNN training and local-only models for comparison table 4 in p.22.  The explanation of their significance is listed in p.21. 
• All experiments on SCAFFOLD were repeated across 20 optimiztion rounds in Fig. 9 of p.19.
• We now provide sensitivity, specificity, PR-AUC, and 95% confidence intervals of the suggested SCAFFOLD and FedProx in Fig. 11 of p.22.
• A new subsection reports computational overhead (training time per epoch,) and communication cost (CPU usage%) for the dual-layer PSO framework. They can be seen in sec. 4.4.5 of. P.23.

Comment 4

• Numerical inconsistencies (accuracy values differ across abstract, tables, figures). the authors claim that most results from the 11-fold cross-validation exceed 90%, although Figure 8 shows values above 90% for only a limited number of folds. Architecture mismatch (sigmoid vs softmax). Privacy claims need moderation.

Response: We carefully reconciled all numerical results:

• Accuracy is consistently reported as 90.56% (mean ± SD), with corresponding F1-score and AUC values aligned across abstract, tables, and figures.
• Cross validation has been further explained.  The statement about the fold accuracy average of 88% and STD of 2% has been added in sec. 4.4. 
• The model architecture is clarified: the final layer uses sigmoid activation for binary classification, and a threshold of 0.5 has been applied for binary classification purpose.  The correction is on sec. 2.2.1 of p.4.
• Regarding privacy, we acknowledge that federated learning alone does not guarantee complete protection. We now explicitly state the assumed threat model and discuss potential vulnerabilities (e.g., inference attacks). While secure aggregation and differential privacy were not implemented in this study, we note these as future work and have moderated claims of “complete privacy” to “privacy-preserving under the federated setting.”The new section is named Formal Privacy Guarantees in Federated Learning of sec.4.4.6 on p.24.

Closing

We believe these revisions substantially strengthen the manuscript and address all reviewer concerns. We thank the reviewers again for their valuable feedback, which has helped us improve the clarity, rigor, and scientific contribution of our work.

Sincerely, Chihhsiong Shih, Cheng-Hsu Chen, Xiuyuan Yeah

 

Reviewer 2 Report

Comments and Suggestions for Authors

1. The dataset is poorly described. Patient number, age, sex, treatment sessions, and class balance are missing.
2. The data are randomly mixed. A patient-level train–test split is not shown. This may cause data leakage.
3. It is unclear whether blood-pressure values used to create the label are also model inputs.
4. The study claims prediction, but the prediction time before the event is not given.
5. The server uses 10% of the data. The source and privacy of this server data are unclear.
6. The CNN structure is inconsistent between the text, figure, and model table.
7. The paper reports 17 features, but Figure 3 shows an input size of 9.
8. The reported accuracy is inconsistent: 90.56%, 93%, and 75–80% are stated in different places.
9. The comparison is weak. Only FedAvg and PSO versions are properly tested. Modern methods such as FedProx, SCAFFOLD, and FedNova are missing.
10. No confidence intervals, standard deviations, or significance tests are provided.
11. The best model still misses 16 of 80 abnormal cases. This clinical risk is not discussed.
12. The feature heatmaps use different colour scales. Therefore, the visual comparison is not fair.
13. Claims about low cost and edge deployment are not supported by real runtime, memory, or communication tests.
14. Table 4 compares unrelated datasets and tasks. This comparison is not meaningful.
15. The paper has many grammar, numbering, formatting, and figure-quality problems.

Author Response

Dear Reviewer,

We would like to express our sincere gratitude for your thorough and constructive feedback on our manuscript, "Dual-Layer PSO-Enhanced Federated Heterogeneous Data Fusion for Hemodialysis Complication Prediction." Your insights have provided valuable guidance for clarifying our methodological choices, refining our dataset reporting, and strengthening our experimental validation.

Below, we provide detailed point-by-point responses to each of your comments and outline the corresponding updates to be incorporated into the revised manuscript.

Point 1: The dataset is poorly described. Patient number, age, sex, treatment sessions, and class balance are missing.

Response: Thank you for highlighting this. The clinical dataset characteristics are outlined in Section 3.2 (Hemodialysis Dataset Preprocessing):

• Data Source & Scope: Collected from 10 maintenance hemodialysis patients across 10 dialysis machines.
• Sessions & Records: A total of 20 treatment sessions (10 normal and 10 abnormal) were analyzed, yielding approximately 400 continuous time-series records per client machine.
• Demographics: Patient age ranges between 60–70 years with an even gender distribution across both classes.
• Class Balance: The dataset maintains an approximately equal distribution between the two classes (Normal: blood pressure remains stable; Abnormal: bedside systolic blood pressure drops below $90\text{ mmHg}$).

Revision Action: We will format these clinical baseline characteristics into a clear summary table in Section 3.2 in the revised manuscript for enhanced clarity and scannability.

Point 2: The data are randomly mixed. A patient-level train–test split is not shown. This may cause data leakage.

Response: We appreciate this critical point regarding data integrity. The partitioning scheme is described in Section 3.1 & Section 3.4 (Data Flow):

• Individual treatment sessions from distinct dialysis machines/clients are isolated into distinct client nodes.
• Evaluation and cross-validation splits (e.g., the 11-fold cross-validation detailed in Section 4.4) were partitioned at the session/machine level.

Revision Action: We will explicitly clarify in Section 3.2 and Section 3.4 that the train-test split strictly obeys session/patient boundaries so that no records from the same treatment session appear in both training and test sets simultaneously, eliminating potential data leakage.

Point 3: It is unclear whether blood-pressure values used to create the label are also model inputs.

Response:

As stated in Section 3.2: "The systolic blood pressure feature is excluded from the training and testing runs among clients, normal abnormal alike."

• The label (0 for Normal, 1 for Abnormal) was defined based on whether bedside systolic blood pressure fell below 90.
• To prevent target leakage during model training, systolic blood pressure was strictly excluded from the input feature set fed into the 1D-CNN. Models only utilize secondary sensor variables (e.g., arterial/venous pressures, ultrafiltration rate, conductivity, temperature, blood flow rate, heart rate).

Point 4: The study claims prediction, but the prediction time before the event is not given.

Response: We appreciate this observation. As addressed in Section 4.4.6 (Prediction horizon):

• The current model operates on real-time sensor streams measured on a minute-by-minute basis to detect complication occurrences on the fly on edge devices (e.g., Raspberry Pi 4).

Revision Action: We will adjust our terminology in the abstract and intro from "early prediction" to "real-time complication detection and risk monitoring" to accurately reflect the minute-by-minute operating horizon. Additionally, we will elaborate on our planned future work integrating sequence architectures (e.g., LSTM-PSOFed) for predictive forecasting hours ahead.

Point 5: The server uses 10% of the data. The source and privacy of this server data are unclear.

Response: As illustrated in Figure 4 (Training and Testing Data Flow), 10% of historical data records are allocated to the central server to serve as an independent validation set.

• Purpose: This server-side validation set is used strictly to evaluate particle fitness (Equation 3/4) when determining global model aggregation weights.
• Privacy: Raw patient data on clients remain local. The server validation set consists of pre-collected, de-identified benchmark records under TVGH IRB permission (SE20056A) as stated in sec. 3.2.

Point 6: The CNN structure is inconsistent between the text, figure, and model table.

Response: Thank you for spotting this discrepancy. The 1D-CNN model architecture used across clients is detailed in Figure 3:

• Input Shape: (None, 9, 1) representing the 9 feature variables per machine type.
• Layer Structure: Conv1D (8 filters, kernel size 3) -> GlobalAveragePooling1D -> Dense (8 units) -> Dense (2 output classes with Sigmoid activation).

Revision Action: We have unified all text references in Section 2.2.1, Section 3.3, and Figure 3 to strictly match this specific 1D-CNN architecture summary table.

Point 7: The paper reports 17 features, but Figure 3 shows an input size of 9.

Response:

This design choices reflects our non-IID heterogeneous sensor setting, as explained in Section 1, Section 3.2, and Section 3.4:

• Across the entire dataset, a total of 17 physiological and machine sensor features were collected.
• To simulate heterogeneous hardware configurations (where different machine brands have distinct sensor subsets), the 17 features are partitioned into client feature groups (e.g., Group A and Group B in Figure 4), each containing a subset of around 9 mutually exclusive/partial features.
• Thus, individual client models ingest an input layer size of 9.

Point 8: The reported accuracy is inconsistent: 90.56%, 93%, and 75–80% are stated in different places.

Response: Thank you for pointing out these conflicting values. The precise metrics are as follows:

• Overall Dual-Layer PSO (PSOFed-HD) Accuracy: 56% (Exact value reported in Abstract, Section 4.4.2, Table 4, and Figure 10).
• Baseline/Variant Ranges: FedAvg achieves 24%, Server-only PSO achieves 75.11%, Client-only PSO achieves 81.97%, and standard non-IID baselines (SCAFFOLD / FedProx) range between 79–80%.
• Typographical errors: References to "93%" in Table 4's summary row and "75–80%" in early draft text were residual artifacts from individual cross-validation folds or hyperparameter particle tuning tests (Table 1).

Revision Action: All summary tables, abstracts, and discussion points will be strictly corrected to report the exact overall accuracy of 90.56% (F1: 0.8533).

Point 9: The comparison is weak. Only FedAvg and PSO versions are properly tested. Modern methods such as FedProx, SCAFFOLD, and FedNova are missing.

Response: We have included state-of-the-art non-IID federated learning algorithms—specifically SCAFFOLD and FedProx—in our experimental evaluations:

• Figure 7 & Table 3: Display direct multi-client comparisons showing that Full PSO significantly outperforms SCAFFOLD (p=0.00001) and FedProx ($p=0.00003).
• Figure 9 & Figure 10: Present learning curve dynamics and confusion matrix comparisons against SCAFFOLD (Accuracy: 79.43%, F1: 0.6325).
• Table 4 & Figure 11: Provide precision, recall, F1-score, and ROC-AUC comparisons (Full PSO AUC = 0.908 vs. SCAFFOLD AUC = 0.81 and FedProx AUC = 0.79).

Revision Action: We will expand our literature review in Section 2.5 and Section 4.4.7 to discuss SCAFFOLD and FedProx, further expanding our benchmark analysis.

Point 10: No confidence intervals, standard deviations, or significance tests are provided.

Response: Statistical validation is detailed in Section 4.2 (Table 3) and Section 4.4 (Figure 8):

• Cross-Validation: 11-fold cross-validation results in an average validation accuracy of 88%, 2% (STD).
• Hypothesis Testing: Two-tailed $t$-tests against SCAFFOLD and FedProx are presented in Table 3, proving the statistical significance of Full PSO (t=10.9067, p=0.00001vs. SCAFFOLD; t=10.944, p=0.00003vs. FedProx).

Point 11: The best model still misses 16 of 80 abnormal cases. This clinical risk is not discussed.

Response: We completely agree that false negatives carry clinical risks. As explicitly noted in Section 4.4.2:

"Although our dual-layer PSO framework achieved 90.56% accuracy and an F1-score of 0.8533, approximately 20% of abnormal cases were not detected. We emphasize that the proposed model should be regarded as a decision-support tool to assist clinicians rather than replace medical judgment. Future work will explore multimodal integration (e.g., ECG, LSTM) and personalized modeling to further reduce false negatives."

Point 12: The feature heatmaps use different colour scales. Therefore, the visual comparison is not fair.

Response: Thank you for this observation regarding Figure 6. As documented in Table 2 (Feature characters comparison), both heatmaps share the exact numerical value range (0.00 to 0.8). However, the shift in visual color distribution (from dark purple to bright yellow/orange) was caused by automatic colormap scaling in matplotlib.

Point 13: Claims about low cost and edge deployment are not supported by real runtime, memory, or communication tests.

Response: Real hardware evaluations on edge devices are provided in Section 4.4.5 (Resource utilization comparison of clients on edge devices):

• Hardware: Tested directly on Raspberry Pi 4 client nodes attached to hemodialysis machines.
• CPU Utilization (Figure 12): The PSO-enhanced configuration operates between 10% and 35% CPU utilization compared to 35% and 50% for unoptimized local CNN training due to adaptive scheduling.
• Execution Time (Figure 13): Client runtime averages 75–80 msec per round, demonstrating practical feasibility for edge gateways.

Point 14: Table 4 compares unrelated datasets and tasks. This comparison is not meaningful.

Response: We acknowledge this perspective. Table 4 (Comparison of Related Studies) is now Table 5 in Section 4.4.7 and was designed as a broad qualitative matrix to contrast architectural differences (client vs. server PSO, model types, domains) across existing FL-PSO literature such as references from 12~14, and 17~19.  In the revised version we have added the following :

1. A Qualitative Feature Matrix comparing system capabilities across published federated PSO literature.
2. A Direct Benchmark Table strictly evaluating our hemodialysis dataset under identical conditions against FedAvg, FedProx, and SCAFFOLD.

Point 15: The paper has many grammar, numbering, formatting, and figure-quality problems.

Response: Thank you for pointing this out.

Revision Action: We are conducting a thorough review of the manuscript to:

• Correct table/figure cross-reference numbering (e.g., duplicate Table 4 numbering in the current draft).
• Re-render high-resolution vector graphics for Figures 1, 3, 4, 5, and 12/13 to ensure crisp text labels.
• Perform full academic English proofreading to polish prose, eliminate grammatical errors, and ensure strict adherence to journal formatting standards.

Once again, we express our gratitude to Reviewer 1 for these insightful suggestions, which have significantly strengthened the rigor and clarity of our manuscript.

Sincerely,

Chihhsiong Shih, Cheng-Hsu Chen, Xiuyuan Yeah

Department of Computer Science, Tunghai University

Department of Nephrology, Taichung Veterans General Hospital

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

1. The manuscript refers to the prediction of hemodialysis complications, including hypotension and discomfort events, whereas the implemented binary endpoint is based specifically on normal versus abnormal systolic blood pressure, with abnormality defined as a value below 90 mmHg. The authors should clarify in the Abstract, Methods, and Conclusions that the present model primarily identifies abnormal blood-pressure states. The distinction between concurrent detection and true early prediction should also be stated explicitly, particularly because the manuscript acknowledges that prospective event prediction would require additional temporal models such as LSTM or Transformer architectures.
2. The revised manuscript provides additional information on the number of sessions and records, but it remains unclear whether the train–validation splits were performed at the individual-record, treatment-session, or patient level. The authors should add a concise statement describing the unit of separation used in the 11-fold cross-validation and clearly distinguish the local client-validation data, server-side validation data used for PSO optimization, and final evaluation data. When patient-level separation was not available, this should be acknowledged as a limitation and included among the priorities for future external validation.
3. The addition of SCAFFOLD, FedProx, the centralized model, precision, recall, F1-score, accuracy, AUC, and statistical comparisons considerably strengthens the experimental evaluation. The authors should nevertheless specify whether the reported t-tests were paired or unpaired, identify the observations used in each comparison, and state whether the relevant assumptions were assessed. Reporting 95% confidence intervals for the principal performance metrics would further improve interpretation. The terminology and numerical values should also be presented consistently across the text, tables, and figure captions.
4. The section entitled “Formal Privacy Guarantees in Federated Learning” should be renamed to reflect that differential privacy and secure aggregation were not implemented in the present study. Claims regarding resistance to membership-inference or reconstruction attacks should be moderated and presented as potential advantages rather than demonstrated guarantees. The unrelated phrase concerning “driving habits” should also be removed. Finally, Figure 3 should be checked to ensure that the displayed output-layer configuration is consistent with the sigmoid-based binary-classification description, and the statements reporting an average cross-validation accuracy of approximately 88% and accuracies above 90% should be reconciled.
5. Although the revised manuscript now provides additional information on the number of treatment sessions, approximate record counts, class balance, patient demographics, and ethical approval, these details are distributed across several paragraphs. The authors should add a concise table summarizing the number of unique patients, sessions, records per class, participating machines or clients, feature subsets available to each client, and the numbers of observations assigned to local training, local validation, server-side validation, and final evaluation. If the number of unique patients cannot be determined from the available dataset, this should be stated explicitly. This addition would improve reproducibility and allow readers to understand the experimental design without requiring any new analyses.

Author Response

Comment 1: The manuscript refers to prediction of hemodialysis complications, but the implemented binary endpoint is specifically abnormal systolic blood pressure (<90 mmHg). Clarify that the model primarily identifies abnormal blood-pressure states. Distinguish concurrent detection vs. true early prediction.

Response: We thank the reviewer for this important clarification. We have revised the Abstract, Methods, and Conclusions as marked blue color to explicitly state that the present model primarily identifies abnormal blood-pressure states, defined as systolic blood pressure <90 mmHg. We also added a statement distinguishing concurrent detection from prospective early prediction, noting that future work will incorporate temporal models such as LSTM or Transformer architectures to achieve true event prediction also in blue text backcolor.

Comment 2: Clarify train–validation splits: record/session/patient level. Add concise statement describing unit of separation in 11-fold cross-validation. Distinguish local client-validation, server-side validation, and final evaluation. If patient-level separation not available, acknowledge as limitation.

Response: We appreciate this suggestion. We have added a concise statement in the Methods section specifying that the 11-fold cross-validation was performed at the treatment-session level. We also clarified the distinction between local client-validation data, server-side validation data used for PSO optimization, and final evaluation data.  The statement is as listed in p.19 of section 4.4 as marked in blue. Since patient-level separation was not available in the current dataset, we have acknowledged this as a limitation and highlighted it as a priority for future external validation in conclusion in p.30.

Comment 3: Specify whether t-tests were paired or unpaired, identify observations, and state assumptions. Report 95% confidence intervals for performance metrics. Ensure consistent terminology and numerical values.

Response: We thank the reviewer for pointing this out. We have revised the Experimental Evaluation section to specify that paired t-tests were used, with observations corresponding to fold-wise performance metrics across models. We have also stated that normality assumptions were assessed. In addition, we now report 95% confidence intervals for accuracy, F1-score, Recall and precision respectively as in Table 5 on p.23. Terminology and numerical values have been harmonized across text, tables, and figure captions matching the addition of Table 1 in response to comment 5.

Comment 4: Rename “Formal Privacy Guarantees in Federated Learning” since differential privacy and secure aggregation were not implemented. Moderate claims regarding resistance to attacks. Remove unrelated phrase “driving habits.” Ensure Figure 3 output-layer configuration matches sigmoid-based binary classification. Reconcile statements reporting average cross-validation accuracy (~88%) vs. accuracies above 90%.

Response: We agree with the reviewer’s observation. The section has been renamed to “Privacy Considerations in Federated Learning”in section 4.4.6 on p.25.  Claims regarding resistance to membership-inference or reconstruction attacks have been moderated to emphasize potential advantages rather than demonstrated guarantees in section 4.4.6 on p.25. The unrelated phrase concerning “driving habits” has been removed and changed to “HD treatments parameters”on point 2 of sec. 4.4.6 on p.25.

We have corrected Figure 3 to ensure consistency with the sigmoid-based binary classification description. The discrepancy in reported accuracies has been reconciled: the average cross-validation accuracy is now consistently reported as ~88%, while accuracies above 90% refer to specific folds. This clarification has been added to the section 4.4 on p.19.

Comment 5: Add concise table summarizing dataset details: patients, sessions, records per class, machines/clients, feature subsets, and data splits. If number of unique patients cannot be determined, state explicitly.

Response: We appreciate this constructive suggestion. We have added a concise summary table 1 on p. 7 in the Methods section, detailing the number of treatment sessions, records per class, participating machines/clients, feature subsets available to each client, and the distribution of observations across local training, local validation, server-side validation, and final evaluation. Since the number of unique patients could not be determined from the available dataset, we have explicitly stated this limitation on p.7.

Closing Statement

We sincerely thank the reviewer for the thorough and constructive feedback. The suggested clarifications and additions have significantly improved the transparency, reproducibility, and rigor of our manuscript.

 

 

Reviewer 2 Report

Comments and Suggestions for Authors

No further comments from myside. 

Author Response

We appreciate your help in making the paper suitable for publication in the Sensors.

Your opinion has significantly improved the transparency, reproducibility, and rigor of our manuscript.