Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This article deals with development and evaluation of an ANFIS framework, which combines ANN architecture and interpretable fuzzy rules for wrist fractures and non-fracture injuries classification based on IRT images. Please, do the following to improve the submission:

- Avoid numerical values of metrics in the Abstract and clearly and shortly point out your own contribution

- In Introduction you started with the problem formulation and your approach to improve classification performance. Additionally, it is necessary to present/consider the context of data, methods and improvements and to add short description of the work by chapters. Also, fracture classification and prediction need to be diversified

- Related work is sufficiently considered, but you need to use it for more clear description of your contribution in relation to this existing approaches, as well as for comparison of your results with results of similar approaches based on ML and fuzzy approach

- Lines 155-159 need to be shortly extended with details what ANFIS can solve in the combination with   interpretable fuzzy 14 rules

- Number of collected and used data requires additional explanation – is this number of collected data sufficient for your model development? Additionally, compare your architecture (Fig. 1) with similar ANN approaches in Related Work section. In general, section 3 requires more details about your model functioning. It includes selection of K-means and fuzzy C-means for clustering

- Results structure must be clarified as well as idea or aim of your experiments in relation with similar approaches. Shortly, you need to explain the flow of your experiments and how obtained results (especially epochs) evaluate your framework. Figure 11/Table 5 also requires additional discussion how you consider these results for the framework improvements.

- Accuracy comparison with existing literature in Fig. 12 is superficial and requires more discussion, but not with your previous results. You need to compare it with similar existing solutions used for similar problems

- You can also adjust your Conclusion with other improvements of your submission

Author Response

Dear esteemed reviewer,

Thank you for taking the time to review our work. We have reviewed our manuscripts with your suggestions. Please see the attached document.

Regards,

Authors

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Please respond to each comment point-by-point (1:1 response), indicating how you have revised the manuscript accordingly.

Major Comments:
1. The Abstract says “Fourteen children (19 fractures, 21 sprains)”, which is impossible and conflicts with Methods (“forty participants … 19 fracture, 21 sprain”). Please reconcile counts consistently across Abstract, Methods (Sec. 3.1), Results, tables, and figures, and explain case-level accounting (per-subject vs per-case). 
2. You report five-fold CV to select σ and rule count, and an 80/20 split for training/validation, plus early stopping at training MSE < 0.08. The pipeline (CV scope vs. final hold-out) is unclear and risks leakage. Please provide a clear schematic: (i) outer hold-out test set separated once; (ii) inner CV on training only for hyperparameter selection; (iii) single final evaluation on untouched test. Remove any CV-on-all-data that leaks information. 
3. In Sec. 3.2.5 you describe two continuous outputs (o1, o2) “likelihood of fracture versus sprain”, yet later you apply a single threshold=0.5. Please specify whether the model outputs a single logit/score or a two-node TSK output. If two outputs, how is the decision made (argmax vs. o1−o2)? Align figures, ROC, and confusion matrices with the exact decision rule. 
4. Std, IQR, kurtosis are said to be “statistically discriminative”, while other candidates (mean/median/max/min, mode/skewness) were discarded. Please report the formal selection procedure (statistical tests, effect sizes, correction for multiplicity) and whether features are (a) absolute injured-hand values or (b) contralateral differences per subject—text mentions both concepts. Provide summary statistics per class and effect sizes. 
5. Clarify whether multiple images/ROIs per subject appear and how within-subject correlation is handled. If the unit is per-frame/sequence, adopt grouped CV (subject-wise split) to avoid optimistic bias. Report exact numbers per split and per experiment. 
6. Report confidence intervals for AUC/accuracy/recall/precision using bootstrap or exact intervals, given the tiny test set (e.g., 8 cases in some confusion matrices). The AUC=1.00 in Experiment 1 is suspicious for such a small sample; quantify uncertainty and repeat with repeated stratified CV or nested CV. 
7. You compare random, K-means, and FCM initializations. Ensure identical data splits and early-stopping criteria across experiments; state random seeds; show learning curves with shaded variability across ≥5 runs per setting (means ± SD). Also explain why σ was fixed at 0.1 for Experiment 1 but adaptively set from centroids for K-means/FCM (touching at ~0.61 peak). Use a uniform policy or justify differences. 
8. For emergency triage, sensitivity to fractures is critical. Present operating points prioritizing sensitivity (e.g., ≥95%) and show trade-offs (specificity) and potential reduction of x-rays. Provide decision-curve analysis or cost-sensitive metrics relevant to missed fracture risk. 
9. Expand IRT acquisition: camera model, emissivity settings, ambient temperature range, stabilization time, distance/angle, ROI definition, motion handling, and analgesic effects (30 participants used paracetamol/ibuprofen). Discuss how these factors could alter thermal statistics and how you controlled/adjusted for them. Provide a reproducible protocol. 

Minor Comments:
1. Improve English style and fix artifacts (e.g., “Paedi- atric”).
2. Fill in missing metadata (Received/Revised/Accepted, editor).
3. Ensure figures/tables are cited sequentially; captions must include axes, units, thresholds.
4. Use consistent class labels (“fracture” vs. “no fracture/sprain”).
5. Add FPR, specificity, PR curves, and F1 scores.
6. Specify optimizer, batch size, learning rate, and patience clearly.
7. Clarify hyperparameter ranges (rule counts, σ) and justify choices.
8. Add collection dates and inclusion/exclusion criteria in Ethics section.
9. Enrich Related Work with recent IRT and XAI triage studies.
10. Share code or rule parameters for reproducibility.
11. State normalization ranges and original units.
12. Add sample sizes to confusion matrices and report CIs.
13. Expand Limitations section (small sample size, overfitting risk).
14. Explain intended integration of ANFIS output into clinical workflow.

Author Response

Dear esteemed reviewer,

Thank you for taking the time to review our work. We have reviewed our manuscripts with your suggestions. Please see the attached document.

Regards,

Authors.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

It will be necessary to add in the text of the submission direct explanation of the Comment 7 as provided in the Response to this comment. Other comments are solved and elaborated in revised version of the submission.

Author Response

Dear esteemed reviewer,

Thanks for spending the time to review our work. We have reviewed our manuscripts with your suggestions. Please see the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for your thorough revision of the manuscript. I have carefully reviewed your response letter and the revised version of the manuscript. Most of the major comments, including clarification of sample size, prevention of data leakage, model output structure, confidence interval reporting, and experimental setup, have been addressed satisfactorily and are now clearly reflected in the revised manuscript. However, a few points remain partially addressed:

1. The feature selection procedure (Comment 4) is still described only by referring to previous work, without a direct statistical analysis or summary statistics included in the manuscript.

2. Sensitivity-focused analysis for emergency triage (Comment 8) was not implemented, though you mentioned it as a future direction.

3. Detailed information on thermal camera settings and potential external factors affecting infrared thermal acquisition (Comment 9) is still limited.

Overall, the revision demonstrates substantial improvement and a sincere effort, but minor enhancements would further strengthen the study. Thank you again for your work and your responsiveness to my feedback.

Author Response

Dear esteemed reviewer,

Thank you for taking the time to review our work. We have reviewed our manuscripts in light of your suggestions. Please see the attached document.

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for your careful and thorough revision of the manuscript.
I greatly appreciate the significant improvements you have made in response to the previous review comments.
In particular, the addition of a detailed feature selection description and comprehensive information on the thermal camera settings and calibration procedures has substantially strengthened the manuscript. These revisions have addressed the earlier concerns very well.
Regarding the sensitivity-focused analysis for emergency triage, I understand the limitations posed by the small sample size in this pilot study. While a high-sensitivity threshold analysis (e.g., ≥95%) was not included, your explanation and inclusion of confusion matrices and ROC curves provide adequate context for the current stage of research.
Based on these revisions, I believe the manuscript is now suitable for publication in MTI and recommend acceptance.