FMEA-Guided Selective Multi-Fidelity Modeling for Computationally Efficient Digital Twin-Based Fault Detection

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposed a fault diagnosis framework for autonomous ship electric propulsion systems that used FMEA-derived risk priority numbers (RPNs) to guide digital twin model fidelity selection, combined with a sliding window-based algebraic aggregation method. Overall, the topic is interesting, and the research idea shows certain practical value. Followings are my comments for its further improvements.

(i). Abstract and text claim the method ‘can reduce computational overhead’, but no quantitative computation time comparison data are provided.

(ii). Section 2.2 introduces a fidelity selection strategy based on RPN, but the thresholds (e.g., 1-5 low, 5-15 medium, >15 high) are presented without justification. How do different threshold settings affect results?

(iii). FMEA is the core component of the framework, authors should present a complete FMEA analysis table. Fig. 4 shows only two RPN values (12.38 and 3.862) without providing the specific scores and scoring rationale for severity (S), occurrence (O), and detection (D).

(iv). Authors acknowledge in lines 605-606 that ‘statistical optimization of the threshold settings was not performed’, which is a major limitation. If thresholds are improperly selected, the performance of the entire diagnostic framework may not be guaranteed.

(v). Real-world operations involve more diverse fault types, such as sensor failures, communication loss, and concurrent faults. Can the proposed method effectively handle these complex situations?

(vi). The accuracy of the digital twin model is the foundation of the entire fault diagnosis framework, but authors provide no quantitative model validation results.

Author Response

REVIEW 1

This paper proposed a fault diagnosis framework for autonomous ship electric propulsion systems that used FMEA-derived risk priority numbers (RPNs) to guide digital twin model fidelity selection, combined with a sliding window-based algebraic aggregation method. Overall, the topic is interesting, and the research idea shows certain practical value. Followings are my comments for its further improvements.

(i). Abstract and text claim the method ‘can reduce computational overhead’, but no quantitative computation time comparison data are provided.

Answer

The computational efficiency of the proposed framework is fundamentally rooted in the mathematical and structural differences between the models. While the High-Fidelity Model (HFM) necessitates an extremely small sampling time to accurately simulate the 12 kHz switching frequency, the Low-Fidelity Model (LFM) inherently eliminates this computational overhead by utilizing an averaging technique, as represented in Equation (2) .

The computational gains provided by these averaged models are a well-established theoretical fact in the field of power electronics. Therefore, this research focuses on proposing a structure optimized for the real-time fault diagnosis of autonomous ships by integrating these models with Risk Priority Numbers (RPN)

 

(ii). Section 2.2 introduces a fidelity selection strategy based on RPN, but the thresholds (e.g., 1-5 low, 5-15 medium, >15 high) are presented without justification. How do different threshold settings affect results?

Answer

We appreciate the reviewer’s insightful comment regarding the justification of the threshold settings. The rationale for the thresholds established in this study is based on the following logical and engineering backgrounds:

• Safety-Critical Approach : Autonomous ships operate in isolated maritime environments where immediate external assistance is unavailable. In particular, the 100-kW electric propulsion system used in this study features a single-line configuration, meaning that any failure can lead to a total loss of propulsion and pose critical risks. Consequently, we proactively applied High-Fidelity Models (HFM) starting from the ‘Moderate Risk’ level to ensure maximum safety and reliability.
• Conceptual Sensitivity Analysis : The selection of thresholds involves a critical trade-off between accuracy and efficiency. If the thresholds are set too high, computational efficiency improves; however, the system may miss opportunities for precise diagnosis of phenomena such as motor torque ripples or current harmonics, which can only be captured through high-fidelity modeling. Conversely, if the thresholds are set too low, the computational burden may exceed available resources, potentially compromising the stability and real-time responsiveness of the monitoring system.
• Characteristics of Weighted RPN : The RPN in this study is calculated as a weighted value rather than a simple arithmetic product, reflecting the relative importance of severity, occurrence, and detection in a maritime context . Therefore, the established thresholds represent an informed engineering judgment intended to balance system availability with diagnostic precision.

 

(iii). FMEA is the core component of the framework, authors should present a complete FMEA analysis table. Fig. 4 shows only two RPN values (12.38 and 3.862) without providing the specific scores and scoring rationale for severity (S), occurrence (O), and detection (D).

Answer

We sincerely appreciate your valuable comments and suggestions.

As you correctly pointed out, Failure Mode and Effects Analysis (FMEA) is a central methodological element of our framework, and we fully agree that providing more detailed information is essential for readers to clearly understand the evaluation rationale and results for each failure mode.

Accordingly, we have revised the manuscript as follows:

• Addition of Appendix: The complete FMEA analysis results have been added as a new Appendix.
• Systematic Failure Mode Classification: The revised table systematically presents all failure modes identified within the system and specifies detailed scores for the three evaluation categories: Severity (S), Occurrence (O), and Detection (D).
• Intuitive Risk Comparison: Additionally, the calculated Risk Priority Numbers (RPN = S x O x D) for each failure mode are included to allow readers to intuitively compare relative risk levels.

The FMEA evaluation in this study was conducted through the expert judgment of professionals with a comprehensive understanding of the design, operation, and maintenance of the target systems. To ensure consistency and objectivity, a multi-expert review process was utilized, and final scores for items with conflicting opinions were derived through a consensus-based approach. Through these measures, we have endeavored to maximize the reliability of each evaluation criterion.

We have revised the manuscript to incorporate your suggestions and look forward to your further review.

 

(iv). Authors acknowledge in lines 605-606 that ‘statistical optimization of the threshold settings was not performed’, which is a major limitation. If thresholds are improperly selected, the performance of the entire diagnostic framework may not be guaranteed.

Answer

We fully agree with the reviewer’s comment regarding the critical importance of threshold optimization for ensuring system reliability. We would like to clarify our approach and the measures taken to mitigate the risks associated with threshold selection as follows:

• Framework-Centric Approach : The primary objective of this study is to demonstrate the structural validity of a novel diagnostic framework that links FMEA-derived risk levels with digital twin model fidelity . Our focus remains on the architectural integration of these elements rather than the statistical optimization of individual parameters.
• Risk Mitigation via Sliding Window : To minimize the risk of misdiagnosis arising from threshold settings, the framework employs sliding-window-based algebraic aggregation and repetition criteria (k) instead of relying on instantaneous data comparisons. This design provides multiple safety mechanisms that prevent malfunctions caused by transient noise and ensure diagnostic stability, even if a specific threshold is defined conservatively or loosely.
• Heuristic Selection Based on Domain Expertise : The proposed 10% threshold was not arbitrarily selected; rather, it was derived from engineering judgment considering the rated operating range and transient state characteristics of the actual 100-kW electric propulsion system. By accounting for these physical constraints, we ensured a practical and reliable baseline for fault detection.
• Future Work for Statistical Optimization : We have explicitly identified the use of Receiver Operating Characteristic (ROC) and Precision-Recall (PR) curves for statistical optimization as a critical future research task . We plan to automate this process across diverse operational scenarios to elevate the proposed framework to a commercially deployable level.

(v). Real-world operations involve more diverse fault types, such as sensor failures, communication loss, and concurrent faults. Can the proposed method effectively handle these complex situations?

Answer

We appreciate the reviewer’s insightful comment regarding the complexities of real-world maritime operations. We have addressed the framework's capability to handle sensor failures, communication losses, and concurrent faults as follows:

• Response to Communication Loss and Delay :
  • The framework utilizes a network stream protocol to establish a fixed one-to-one communication channel between the ship and the shore-based center .
  • This protocol ensures rapid data transfer and reliable communication without data loss, which is critical for real-time monitoring.
  • Furthermore, the sliding-window method is inherently robust against real-time data delays or noise, maintaining diagnostic stability even during temporary data drops .
• Response to Sensor Noise and Failures :
  • Unlike conventional methods that rely on single data points, this framework performs algebraic aggregation (e.g., Mean and Count) within the sliding window.
  • This approach effectively filters out transient noise or spikes that could otherwise lead to false alarms.
  • By measuring the frequency of exceedances over a specified duration, the system prevents misclassifying normal operational transients as faults .
• Response to Concurrent Faults :
  • The framework is built upon Failure Mode and Effects Analysis (FMEA), which is designed to systematically evaluate potential failures at the component level .
  • While this study focused on propulsion motor failures, the framework features a modular structure that allows for the integration of sensor or communication failures as independent failure modes within the FMEA table .
  • This scalability enables the system to analyze the propagation of risks in complex, concurrent fault scenarios.
• Future Enhancements :
  • We agree with the necessity for precision validation regarding multiple-fault scenarios.
  • Consequently, we have clarified in the Conclusion that future research will expand the diagnostic scope to include batteries, cooling systems, and power converters to address more diverse operational profiles .

 

(vi). The accuracy of the digital twin model is the foundation of the entire fault diagnosis framework, but authors provide no quantitative model validation results.

We fully agree with the reviewer’s point that the accuracy of the digital twin (DT) model is fundamental to the reliability of the entire fault diagnosis framework. To address this, we have elaborated on the quantitative and qualitative grounds for our model validation as follows:

• Precision Tuning Based on Actual Ship Specifications (Model Parameterization) :
  • The digital twin in this study is not merely a theoretical construct; it strictly reflects the physical specifications of the 100-kW electric propulsion test vessel used in the experiments.
  • Physical validity was ensured by directly inputting detailed equipment specifications—including battery internal resistance and capacitance parameters , as well as the stator resistance, moment of inertia, and friction coefficient of the propulsion motor —into the MATLAB/Simulink model.
• Model Consistency under Nominal Conditions :
  • Figures 10, 11, and 12 present the results of comparing the behaviors of the High-Fidelity Model (HFM) and the Low-Fidelity Model (LFM) based on actual operational data.
  • The analysis confirms that the output values of both models are nearly identical across all steady-state regions, excluding the initial transient response period. This visually corroborates that the digital twin possesses sufficient accuracy to serve as a 'Nominal Baseline' representing the ship's normal operating state.
• Deviation-Based Diagnosis Logic :
  • Rather than focusing on absolute numerical values, this framework concentrates on detecting the 'relative deviation between the steady-state DT model and real-time data'.
  • Since the inherent error between the two models is maintained at an extremely low level, stable fault detection was achieved using a conservative 10% threshold without triggering false alarms.
• In conclusion, the digital twin in this study accurately replicates the physical characteristics of the actual vessel. This has been indirectly validated through the consistency of the simulation results and the stability of the observed diagnostic delay times (1.2 s and 2.7 s).

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Title:

- “Fidelity selection” is ambiguous

- Core novelty is not explicitly declared 

- Lacks emphasis on key advantage (e.g., real-time, efficiency)

- “Application to …” may suggest a limited generalizability

2. Abstract: 

- Novelty not clearly defined

- No quantitative results and missing evaluation metrics

- Not a clear comparison to AI methods or baselines 

- Claims not supported with numbers

- Key concepts not clearly defined (e.g., fidelity selection)

- Insufficient experimental details (dataset, fault types unclear)

3. Keywords: 

- Missing novel term, such as fidelity selection

4. Introduction: 

- It is preferred to separate the Introduction section from the Literature Review section. Also, transfer the study's contributions to the end of the Introduction section as a list. Try to include the most recent studies, e.g., 2021 to 2026, in the Literature Review section.  With at least 10 such references, complete Table 1. Not with long paragraphs, but with the main points of the methods. Avoid including multiple references for one sentence.  

5. Proposed Method: 
- In the overview of the proposed fault diagnosis method, the fault detection process should be technically discussed, considering the thresholds specifically.  

- In mathematical formulas (1–11), all symbols must be explained within the formula or immediately alongside it; or you can add a table for parameters like Table 3, before starting the definitions of formulas. 

- Try to include the "Binary Classification" discussion more technically for fault detection, both in the methodology and in the literature review, as you do the Normal/Abnormal classification task. 

- Figure 9 must be corrected as it is not a readable UI page for the reviewer. 

6. Results: 

- Explain "entanglement fault condition", including effective parameters for it in detail. 

- All results are currently presented in graphical form. It is recommended to also include key results in tabular format to improve clarity and facilitate quantitative comparison.  

- Moreover, statistical analysis should be performed to verify the significance of the results. 

7.  Conclusion and Discussion: 

- Separate the "Conclusion and Related Works" section from the "Discussion" section; in "Discussion", the proposed methodology must be compared with other state-of-the-art techniques in a table by including achieved results and improvements in various evaluation metrics. In "Conclusion and Related Works", authors must conclude the study and present future avenues for it.  No explicit comparisons are given to prove the outperformance of the proposed method. 

Author Response

REVIEW 2

1. Title:

- “Fidelity selection” is ambiguous

- Core novelty is not explicitly declared 

- Lacks emphasis on key advantage (e.g., real-time, efficiency)

- “Application to …” may suggest a limited generalizability

Answer

We sincerely appreciate your insightful suggestions regarding the title of our manuscript.

We fully agree that the term "fidelity selection" in the original title was somewhat ambiguous and failed to clearly convey the core novelty and methodological features of our research to the readers. Accordingly, we have revised the title to:

"FMEA-Guided Selective Multi-Fidelity Modeling for Computationally Efficient Digital Twin-Based Fault Detection".

The revised title reflects the following three key improvements:

• FMEA-Guided: By explicitly mentioning "FMEA-Guided," we emphasize that the selection between high-fidelity (HF) and low-fidelity (LF) models is a systematic process grounded in the Failure Mode and Effects Analysis rather than an arbitrary choice.
• Selective Multi-Fidelity Modeling: This phrase clarifies the vague "fidelity selection" term and more precisely describes our core approach of selectively combining multiple fidelity levels based on risk priority .
• Computationally Efficient: This addition directly communicates the primary practical advantage of our framework—computational cost optimization—at the title level.

Furthermore, we removed the phrase "Application to" to underscore the potential of the proposed framework as a general methodology with broad scalability, rather than one limited to a specific case study.

We hope these revisions effectively address your concerns, and we welcome any further comments you may have.

 

2. Abstract: 

- Novelty not clearly defined

- No quantitative results and missing evaluation metrics

- Not a clear comparison to AI methods or baselines 

- Claims not supported with numbers

- Key concepts not clearly defined (e.g., fidelity selection)

- Insufficient experimental details (dataset, fault types unclear)

Answer

We sincerely appreciate your valuable comments. As you pointed out, the original abstract lacked sufficient explanations regarding the study’s novelty, definitions of key concepts, quantitative results, and experimental details .

Accordingly, we have comprehensively revised the abstract to clearly highlight the core contributions of this research. The specific improvements are as follows:

• Clarification of Core Contribution : We have explicitly described the proposed "Selective Multi-Fidelity Digital Twin Framework," which quantifies risk through FMEA and selectively operates between low-fidelity and high-fidelity models based on the calculated risk priority .
• Definition of Key Concept s: The meaning of "fidelity selection" has been clarified as the selective operational strategy between two distinct fidelity models to balance computational efficiency and diagnostic precision.
• Methodological Refinemen t: We have elaborated on the methodological features, specifically the lightweight real-time diagnostic structure utilizing sliding-window-based algebraic aggregation.
• Specification of Experimental Conditions : The experimental setup has been detailed by specifying the use of real operational data from a 100-kW electric propulsion ship and the inclusion of diverse fault scenarios such as power and signal anomalies.
• Emphasis on Computational Efficiency : The advantages of the proposed method in terms of computational cost are now clearly described through a comparison with conventional data-driven AI techniques, underscoring its practical applicability to autonomous vessels.

These revisions are fully reflected in the abstract of the revised manuscript. We hope these changes adequately address your concerns.

3. Keywords: 

- Missing novel term, such as fidelity selection

Answer

Thank you for your valuable feedback.

We agree with the reviewer’s suggestion, and considering that "Fidelity selection" is one of the core concepts of this research, we have added this term to the list of keywords. We have revised the manuscript to reflect the points raised, and we look forward to your further review.

 

4. Introduction: 

- It is preferred to separate the Introduction section from the Literature Review section. Also, transfer the study's contributions to the end of the Introduction section as a list. Try to include the most recent studies, e.g., 2021 to 2026, in the Literature Review section.  With at least 10 such references, complete Table 1. Not with long paragraphs, but with the main points of the methods. Avoid including multiple references for one sentence.  

Answer

We sincerely appreciate your insightful comments regarding the structural weaknesses of our manuscript. In response to your suggestions, we have comprehensively revised the relevant sections as follows:

• Section Separation for Enhanced Readability : Following your recommendation, the Introduction and Literature Review have been separated into two independent sections. This restructuring significantly improves the logical flow and readability of the paper. The revised Introduction now focuses on the research background, problem definition, and the necessity of this study, while the Literature Review systematically organizes related works .
• Restructuring of Research Contributions : We have restructured the research contributions into a numbered list at the end of the Introduction. This allow readers to more clearly identify the primary contributions and the distinct novelty of our study .
• Update with Recent Literature (2021–2026) : The Literature Review has been updated to include 10 recent studies published between 2021 and 2026. Accordingly, Table 1 has been revised to provide concise summaries of key methodologies and features, replacing lengthy descriptions .
• Improved Citation Style : We have also moved away from the previous method of citing multiple references collectively within a single sentence. Instead, each reference is now discussed independently to ensure that its specific contribution is clearly presented.

Through these revisions, we have significantly improved the structural integrity, recency, and overall readability of the manuscript. We are grateful for your time and expertise in reviewing our work and look forward to your further consideration.

 

5. Proposed Method: 

We would like to express our sincere gratitude for your insightful comments and for pointing out the shortcomings of our manuscript. We have carefully revised the relevant sections to incorporate your suggestions, and the specific details are as follows

- In the overview of the proposed fault diagnosis method, the fault detection process should be technically discussed, considering the thresholds specifically.  

Answer

Following the reviewer's constructive suggestion, the fault detection process in the overview of the proposed methodology has been technically reinforced.

Rather than merely stating that a threshold is exceeded, we have elaborated on the physical significance of the residual magnitude falling outside the system's permissible tolerance range. Furthermore, we have provided a detailed description of how these residuals are statistically processed within the sliding window-utilizing algebraic aggregation to arrive at the final fault determination.

These revisions clarify the logical transition from raw data anomalies to definitive fault decisions, ensuring the technical rigor of the proposed diagnostic framework.

- In mathematical formulas (1–11), all symbols must be explained within the formula or immediately alongside it; or you can add a table for parameters like Table 3, before starting the definitions of formulas. 

Answer

We have supplemented the variable definitions within the main text to ensure that all symbols used in the mathematical formulas are immediately understandable in their proximity.

Specifically, additional explanations have been provided for symbols that were previously insufficiently defined or omitted, ensuring that both the physical significance and the mathematical role of each symbol are clearly conveyed to the reader. These revisions have been applied to Equations (1) through (11) to enhance the technical clarity and readability of the manuscript.

- Try to include the "Binary Classification" discussion more technically for fault detection, both in the methodology and in the literature review, as you do the Normal/Abnormal classification task. 

Answer

We have redefined the fault detection process in this study as a Binary Classification task (Normal vs. Abnormal) and substantially reinforced both the Methodology and Literature Review sections accordingly.

In particular, we provided a technical discussion comparing the proposed sliding-window-based algebraic aggregation method with complex AI-driven binary classification models. Our discussion emphasizes that while AI models often act as "black boxes," our proposed approach offers significantly lower computational costs and higher interpretability. These attributes are critical for the real-time reliability and transparent decision-making required in autonomous ship propulsion systems.

- Figure 9 must be corrected as it is not a readable UI page for the reviewer. 

Answer

In response to the reviewer’s feedback, the resolution of Figure 9 has been increased, and its visual clarity has been significantly enhanced to ensure that the user interface (UI) components are clearly readable for the readers.

We sincerely appreciate the time and effort you have dedicated to reviewing our manuscript. We hope the updated figure meets your expectations and look forward to your further consideration.

 

6. Results: 

- Explain "entanglement fault condition", including effective parameters for it in detail. 

Answer

The physical significance of the 'Entanglement fault condition' presented in this study and the validity of the parameters used for its diagnosis are as follows :

• Physical Mapping : Case 1 simulates a sudden increase in load caused by foreign substances (such as abandoned fishing nets or ropes) becoming entangled in the propeller in a maritime environment. Mathematically, this manifests as a 'step-fault', which exerts a direct and continuous impact on the torque and current of the propulsion system.
• Role of Diagnostic Parameters : The threshold for torque deviation (set at 10%) is a key parameter determining diagnostic sensitivity. It was established at a level that significantly exceeds the inherent model error in steady-state conditions, thereby minimizing the false alarm rate. Furthermore, the window size(W) and the aggregation function(f) are specifically designed to filter out transient sensor noise and extract only 'sustained performance degradation.'
• Rationale for Detection Delay : The observed detection delay of 1.2 seconds represents the minimum time required for the sliding window to accumulate sufficient fault data to ensure statistical significance. This performance level is fully compatible with the requirements for real-time responsiveness in autonomous vessels, allowing for stable decision-making without overreacting to instantaneous spikes.

- All results are currently presented in graphical form. It is recommended to also include key results in tabular format to improve clarity and facilitate quantitative comparison.  

Answer

We appreciate the reviewer’s suggestion to organize key figures in a tabular format to improve clarity. Regarding the presentation of our results, we would like to provide the following clarification:

• Necessity of Graphical Representation : The core of the fault diagnosis addressed in this study lies in detecting dynamic changes within time-series data. In particular, presenting data in graphical form was essential to demonstrate the prevention of malfunctions during transient states and to illustrate the continuous flow of residual changes following a fault occurrence.
• Focus on Diagnostic Paradigm : The primary contribution of this research is not merely proving the numerical superiority of specific data, but rather proposing a novel diagnostic paradigm: the "Risk-based Multi-fidelity Digital Twin." The theoretical efficiency of the proposed models is firmly grounded in the Averaged Model technique, which is already widely accepted in the academic community.
• Validity of Quantitative Evidence : We believe that the presented time-series graphs serve as the most explicit quantitative data demonstrating the real-time responsiveness of the algorithm. Therefore, the validity of the framework is sufficiently explained through the detailed descriptions within the text, even without additional data reconfiguration into tables.

We hope this explanation clarifies the rationale behind our data presentation and underscores the structural contributions of our work.

 

- Moreover, statistical analysis should be performed to verify the significance of the results. 

Answer

We appreciate the reviewer’s suggestion regarding the verification of statistical significance. To ensure the reliability of real-time diagnosis, the proposed framework incorporates the following mechanisms for statistical robustness within its algorithmic structure:

• Instead of processing raw data directly, the algorithm performs algebraic aggregation (e.g., Mean, Sum) within a sliding window. This acts as a statistical filtering process that suppresses stochastic noise inherent in real-time data, thereby enhancing the reliability of the diagnostic results.
• The repetition criteria introduced in the fault detection function (Equation 16) serves as a critical threshold mechanism. It is designed to prevent misdiagnosis caused by isolated outliers and to capture only those fault signatures that are statistically significant and persistent over time.
• We have confirmed that the error rate between the nominal state data generated by the digital twin and the actual data remains extremely low under normal operating conditions. Consequently, the observed deviations of 10% or more during fault injection can be regarded as physically and statistically significant changes rather than random fluctuations.

In conclusion, the proposed methodology ensures the significance of diagnostic results in real-time through internal aggregation and repetition logic, precluding the necessity for separate post-hoc statistical tests.

 

7.  Conclusion and Discussion: 

Answer

- Separate the "Conclusion and Related Works" section from the "Discussion" section; in "Discussion", the proposed methodology must be compared with other state-of-the-art techniques in a table by including achieved results and improvements in various evaluation metrics. In "Conclusion and Related Works", authors must conclude the study and present future avenues for it.  No explicit comparisons are given to prove the outperformance of the proposed method. 

We sincerely appreciate your valuable feedback. As you correctly pointed out, the Discussion and Conclusion were previously combined into a single section, which limited the clarity of each component.

Following your suggestion, we have separated the original section into two distinct sections: "Discussion" and "Conclusion," clearly defining the purpose and content of each.

• In the revised "Discussion" section, we now focus on the interpretation of the research results and their technical implications.
• The "Conclusion" section has been restructured to provide a concise summary of the study’s primary findings and contributions.

Through this restructuring, the logical flow and the distinction between the analysis of findings and the final summary have been significantly improved.

Existing AI-based diagnostic models have faced limitations in real-time responsiveness due to their high computational costs. In contrast, this study maximizes computational efficiency through RPN-based adaptive fidelity selection while demonstrating superior performance in practical operational environments by ensuring rapid detection within approximately 1.2 seconds.

In the Conclusion section, we have addressed the optimization of threshold settings as a current limitation of this study. Furthermore, we presented a specific roadmap for future research, which includes expanding the diagnostic scope to components such as batteries and power conversion systems and implementing automated techniques based on Receiver Operating Characteristic (ROC) curves.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Authors have addressed my raised comments in this version.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for considering all the comments. The manuscript is acceptable in the current format from the reviewer's standpoint.