All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
Dear authors, we are pleased to verify that you have met the reviewer's valuable feedback to improve your research.
Thank you for considering PeerJ Computer Science and submitting your work.
Kind regards
PCoelho
[# PeerJ Staff Note - this decision was reviewed and approved by Shawn Gomez, a PeerJ Section Editor covering this Section #]
I am satisfied with the revised version of the manuscript. The authors have adequately addressed all previous concerns, and no further comments are necessary from my side.
N/A
N/A
N/A
Dear authors,
Thanks a lot for your efforts to improve the manuscript.
Nevertheless, some concerns are still remaining that need to be addressed.
Like before, you are advised to critically respond to the remaining comments point by point when preparing a new version of the manuscript and while preparing for the rebuttal letter.
The reviewer has suggested that you cite specific references. You are welcome to add it/them if you believe they are relevant. However, you are not required to include these citations, and if you do not include them, this will not influence my decision.
Kind regards,
PCoelho
**PeerJ Staff Note:** It is PeerJ policy that additional references suggested during the peer-review process should only be included if the authors agree that they are relevant and useful.
Thank you for the thorough revision. The revised manuscript has addressed the majority of the previous concerns effectively. The structure, methodological clarity, and presentation have notably improved. The integration of continuous wavelet transform with a hybrid deep learning model (MHSA, Bi-LSTM, 1D Conv ResNet) is well-justified and experimentally validated with strong results across multiple datasets.
No comments
Only a few minor comments remain before final consideration:
• Consider briefly explaining how MHSA integrates with Bi-LSTM in the fusion strategy (currently this is only implied but not clearly described in this section).
• The description of the signal strength ranges is helpful, but you might also indicate which sensor axis or activity mode is the most discriminative based on the CWT maps.
• It is recommended while discussing the pretrained models in the introduction section of the manuscript, discuss latest papers like https://doi.org/10.3390/machines12120905.
No additional comments.
Dear authors,
Thanks a lot for your efforts to improve the manuscript.
Nevertheless, some concerns are still remaining that need to be addressed.
Like before, you are advised to critically respond to the remaining comments point by point when preparing a new version of the manuscript and while preparing for the rebuttal letter.
Kind regards,
PCoelho
The manuscript presents a hybrid deep learning approach for lower limb joint torque estimation, integrating continuous wavelet transform (CWT), multi-head self-attention (MHSA), Bi-LSTM, and 1D convolutional residual networks to enhance feature extraction, noise suppression, and temporal dependency modeling. Experimental validation on public datasets demonstrates high accuracy and real-time applicability, outperforming existing models in computational efficiency and prediction performance.
How does the CWT-based preprocessing handle high-frequency noise, and how does it compare to traditional denoising methods like adaptive filtering or empirical mode decomposition?
The framework integrates 1D Conv ResNet, Bi-LSTM, and MHSA—what is the specific contribution of each module in improving torque estimation, and how were their hyperparameters optimized?
Given the computational complexity of CWT and MHSA, how does the proposed approach balance accuracy with real-time applicability on edge devices?
The ablation study indicates performance improvements with additional modules—can you provide a deeper explanation of how feature fusion across different network components enhances accuracy?
How does the proposed framework compare with other transformer-based or CNN-RNN hybrid architectures for time-series biomechanical analysis?
The model is trained and tested on public datasets—have you assessed its performance on real-world noisy data, and how does it handle sensor misalignment or calibration errors?
The t-SNE visualizations indicate strong feature separation—how does the model handle class imbalance in gait cycles, and do you employ any sampling strategies to mitigate bias?
Given the potential for wearable real-time monitoring, what are the key hardware limitations of deploying this framework, and do you plan to explore lightweight versions for embedded systems?
In the introduction section of the manuscript, it is recommended to cite the latest papers published
The quality of some of the figures needs to be enhanced for the final version of the manuscript, especially of the results and discussions.
The authors have addressed all the concerns raised by the reviewer and the manuscript may now be accepted.
The authors have addressed all the concerns raised by the reviewer and the manuscript may now be accepted.
The authors have addressed all the concerns raised by the reviewer and the manuscript may now be accepted.
The authors have addressed all the concerns raised by the reviewer and the manuscript may now be accepted.
Dear authors,
You are advised to critically respond to all comments point by point when preparing an updated version of the manuscript and while preparing for the rebuttal letter. Please address all comments/suggestions provided by reviewers, considering that these should be added to the new version of the manuscript.
Kind regards,
PCoelho
[# PeerJ Staff Note: It is PeerJ policy that additional references suggested during the peer-review process should *only* be included if the authors are in agreement that they are relevant and useful #]
1. Authors must show explain the novel contribution of the work with proper justification of the outcomes. What novelty is established in this work compared to existing works?
2. Literature survey need to be improved and updated based on current state of art methods. Some more paper based on Estimation of Lower Limb Torque via Deep Learning.
3. Organization of the paper can be added at the end of introductions.
4. Explaining the problem and the gaps in existing literature in a concise but self-contained way (although readers might wish to consult references, they should not be forced to do so)
5. The computational complexity in terms of time and space must be discussed. Also, compare the proposed method in terms of computational complexity?
6. Comparative analysis of various performance parameters with respect to sate of art methods must be discussed. More recent state-of-the-art approaches should be compared; the experiments should use more sizable real-world data sets from public repositories (if any);
7. Add industrial significance of the proposed approach.
8. Results must be verified based on some other data sets also. Describing in detail the data set used and what are the expected outcomes- widening the experimental comparison including other data and methods.
9. Precision vs. recall curves of the proposed algorithms with respect to different data sets must be included.
10. How much data should be considered for training and testing for architecture implementation? Details of training and testing data sets must be tabulated.
11. Comparative analysis with respect to real-time time analysis is missing?
12. Limitations of the proposed work must be included.
The manuscript titled "Estimation of Lower Limb Torque: A Novel Hybrid Method Based on Continuous Wavelet Transform and Deep Learning Approach" introduces an innovative method that integrates IMU sensors with deep learning architectures to estimate lower limb joint torques. The approach combines Continuous Wavelet Transform (CWT) and a Bi-LSTM neural network for precise real-time analysis, achieving high accuracy and robustness when compared to existing methods, particularly for the hip, knee, and ankle joint torque estimation. The following are a few questions and/or suggestions before considering the manuscript for any acceptance from my side.
1) The introduction outlines the need for lower limb torque estimation but lacks a deeper analysis of the limitations of existing methods. Can the authors elaborate on the current challenges faced in real-world applications of these methods?
2) How does the choice of Morlet wavelet as the mother wavelet compare to other wavelet types? Were any other types considered during experimentation?
3) The model uses Bi-LSTM for long-term temporal dependencies. How was the size of the LSTM units determined, and what effect does increasing or decreasing the number of units have on model performance?
4) The study mentions using a 64-filter convolutional layer in the 1D Conv ResNet. How was this number chosen, and could different filter sizes affect the results?
5) The preprocessing step uses the Acc/Gyro ratio and Dynamic Acceleration. What was the rationale behind introducing these metrics, and how do they improve torque estimation compared to standard preprocessing methods?
6) Did the model's performance differ significantly across different types of activities (e.g., stair descent vs. treadmill walking)? If so, how might this be addressed in future iterations of the model?
7) The RMSE values for certain scenarios, such as flat ground, are higher than others. Are there any data characteristics or sensor positioning factors that may have contributed to this?
8) Could the model’s generalization across different individuals be impacted by varying sensor placement or individual gait differences? If so, how might these issues be mitigated?
9) To enhance the quality of the manuscript, in the introduction section, in the line CWT is mentioned, it is recommended that you include the following recent paper as, Siddique, M.F.; Ahmad, Z.; Ullah, N.; Kim, J. A Hybrid Deep Learning Approach: Integrating Short-Time Fourier Transform and Continuous Wavelet Transform for Improved Pipeline Leak Detection. Sensors 2023, 23, 8079. https://doi.org/10.3390/s23198079.
10) While the manuscript discusses future directions, it would be helpful to outline potential real-world applications where this method could outperform existing techniques. Could the authors suggest specific fields or industries where this technology would be most beneficial?
11) The linguistic quality of this manuscript is not satisfactory. Many grammatical errors, personal pronouns, and poorly constructed sentences are found. Please proofread and improve the quality.
The manuscript introduces a method combining Continuous Wavelet Transform (CWT) with deep learning techniques like Bi-LSTM and convolutional residual networks to estimate lower limb joint torques using Inertial Measurement Units (IMUs) data. The system demonstrates high accuracy, with notable performance metrics, and aims to overcome the limitations of traditional gait analysis systems. The study validates its model using a public dataset, reporting superior results to baseline models. Despite the exciting and significant contributions, a few primary concerns must be addressed before publication.
1. The manuscript does not sufficiently address recent advancements in deep learning applications for biomechanical data analysis. Studies using cutting-edge models like transformers and more modern variations of recurrent neural networks should be included to provide better context.
2. The manuscript lacks detailed derivations and explanations of the mathematical functions, especially regarding optimization processes and hyperparameter tuning for neural networks.
1. The study is based purely on laboratory-controlled datasets, which limits the generalizability of the results. The authors should expand the validation to real-world scenarios or noisy environments, considering the inherent noise and variability in IMU data during uncontrolled movements.
2. The results indicate that the proposed model outperforms some baseline methods, but the paper lacks a detailed comparative analysis. It would benefit from deeper discussions on why specific models, such as Bi-LSTM, perform better in specific scenarios and how they address the limitations of alternative models.
3. The impact of using CWT for feature extraction is mentioned but not sufficiently analyzed. More details should be provided on how CWT improves performance over other time-frequency analysis methods, and the paper should evaluate the trade-offs in terms of computational complexity.
1. While the manuscript presents performance metrics like RMSE, R², and MAE, the results are scattered across various sections, making it difficult to follow the narrative. A more organized presentation of results, such as summarizing critical findings in comprehensive tables, would improve clarity.
2. The discussion section focuses on the strengths of the proposed model but lacks a critical analysis of its limitations. The authors should address the challenges of working with IMU data, such as sensor misalignment or drift, and the model's limitations in estimating joint torques under different gait patterns.
1. How was the wavelet frequency (ω₀) in the Morlet wavelet chosen, and were different values tested for optimization during the training process?
2. How were the time-series dependencies modeled in the Bi-LSTM model, and what specific activation functions were employed to improve the learning process?
3. Can the authors provide more details on the gradient optimization technique used in the 1D Conv ResNet model? Were any regularization techniques applied to prevent overfitting?
4. How was the Acc/Gyro Ratio metric mathematically validated regarding its contribution to the model’s accuracy?
5. How was the learning rate of the Adam optimizer adjusted during the training process, and what criteria were used for early stopping?
6. Were any specific statistical tests used to evaluate the significance of the model's performance improvements over the baseline models?
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.