Motion Pattern Recognition Based on Surface Electromyography Data and Machine Learning Classifiers: Preliminary Study
Round 1
Reviewer 1 Report
Comments and Suggestions for Authorsattached are the comments
Comments for author File: 
Comments.pdf
Author Response
We would like to thank the Reviewer for the questions and remarks that allowed us to revise and update our manuscript. Considering all questions/remarks of the Reviewer, we improved our manuscript and put these updates in red. Please find the detailed responses below and the corresponding revisions along with corrections highlighted in-red. We hope that this revised version meets all needed requirements.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper presents a preliminary study on recognizing upper limb motion patterns by classifying time-series features extracted from surface EMG data using various machine learning classifiers. Among the 23 models tested, K-NN consistently achieved the highest performance in distinguishing muscle activation states under different arm positions. While the findings have potential to contribute to the field, I believe substantial revision is needed.
While the Introduction provides a broad overview of existing literature on EMG-based motion classification, the logical structure is difficult to follow, and the core research question remains vague. Although the authors state that this is a “preliminary study,” the manuscript does not clearly define what specific gap in the literature is being addressed, nor does it sufficiently justify the novelty or significance of the work. The objective (recognizing motion patterns using time-series features from EMG data) is quite general and has already been widely explored in prior studies. I believe a more focused and well-motivated explanation of the problem and hypothesis is needed, along with a clearer rationale and specific contribution of this work.
The rationale behind selecting specific machine learning models, the composition of features (i.e., concatenation of EMG signals), and the choice of time windows or normalization methods is not clearly explained. These methodological decisions need to be explicitly justified in relation to the study's goals and relevant prior work.
I believe the preprocessing pipeline is not sufficiently reproducible. The segmentation step relies partly on visual inspection, but no objective criteria or examples are provided to define how onset and offset points were determined. Also, key parameters for filtering, rectification, and smoothing (e.g., filter type, cutoff frequencies, RMS windowing specifics) are missing. Without access to the exact processing code or a clear algorithmic description, it is difficult to replicate the time-series features used in classification.
Although the paper compares many classifiers using standard metrics and cross-validation, the evaluation lacks depth in several aspects. There is no statistical analysis to compare model performance, no baseline model for reference, and no ablation analysis to assess the contribution of design choices. In addition, all evaluation is conducted within the same subject pool without testing for generalization across subjects, which limits the practical implications of the findings.
The figure and table labeling is inconsistent with their actual presentation. Items like "Figure 6A" and "Figure 6B" or "Table 1A" through "1D" are not subpanels of a single grouped figure or table, but entirely separate elements. This naming convention is misleading and creates confusion in navigating the results.
The Discussion is underdeveloped and primarily repeats detailed results already presented in the previous section. It lacks deeper interpretation of findings, critical comparison with past studies beyond surface-level accuracy values, and any insight into why certain classifiers performed better than others. I suggest revising this section to focus less on restating metrics and more on drawing conclusions, exploring implications, and critically reflecting on the study’s scope and limitations.
The study uses a single RMS-based time-domain feature representation, without discussing widely used alternatives such as frequency-domain or time-frequency features, or higher-level representations like median frequency or muscle synergies. Even if the authors chose to use RMS features exclusively, they should cite and discuss prior work that employed more conventional or physiologically meaningful features. This omission weakens the methodological foundation of the paper. Regardless of whether such features are used directly for classification, they are widely adopted in upper-limb motor control research and offer critical insights into muscle coordination and function, as demonstrated in [1]. The absence of this discussion suggests an incomplete engagement with the broader scientific literature. I encourage the authors to consult the example and expand their review and discussion accordingly.
[1] https://doi.org/10.1007/s12541-019-00251-5
Author Response
We would like to thank the Reviewer for the questions and remarks that allowed us to revise and update our manuscript. Considering all questions/remarks of the Reviewer, we improved our manuscript and put these updates in red. Please find the detailed responses below and the corresponding revisions along with corrections highlighted in-red. We hope that this revised version meets all needed requirements.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe main objective of the paper is a study on the application of classical machine learning (ML) methods to recognize patterns and classify electromyography (EMG) signals. The contribution of the paper is focused on experimental results, as the methods employed are known. The paper includes extensive experiments that might be interesting from a practical standpoint. The literal presentation of the paper should be improved. The evaluation of results has room for improvement and the discussion on related methods should be extended. In summary, I consider the contents of the paper are potentially publishable, but the following issues should be addressed in a revised version of the paper.
- The literal presentation of the paper has room for improvement. For instance, (i) page 3, line 93, “the paper [36]” change to “in [36]”, "paper" is redundant and it should be removed. Check this in several parts of the paper. (ii) all acronyms should be defines the first time they are used, except in the abstract, e.g., “ACC”, page 4, line 178. Check for confusion between acronym for accelerometer and accuracy. (iii) There are some pages difficult to read, e.g., page 8. Please try to make those parts of the paper more readable; e.g., explanations that repeat data already presented in tables are redundant, and summary figure(s) could help.
Therefore, an English proofreading of the paper is required.
- Several lightweight ML methods were implemented for classification of EMG signals: linear and quadratic discriminant (LDA and QDA), k-nearest neighbors (kNN), decision trees, support vector machines (SVM), and logistic regression. An approach to improve the results in both accuracy and stability would be the fusion of the results of the results of the six single classifiers. Recently, alpha integration has been proposed for optimal late fusion of scores from multiple classifiers. Please discuss this, theoretically and/or practically. I suggest the following reference: https://doi.org/10.1109/ACCESS.2023.3344776.
- The authors refers the EMG signals as time series. Please discuss on classification methods that consider temporal dependencies (e.g., SICAMM, Sequential Independent Component Analysis Mixture Model) and their possible implementation for the posed problem.
- An analysis of the computational cost of the methods implemented should be included in comparisons.
- Receiver operating characteristic (ROC) curves and precision-recall (PR) curves are powerful tools to evaluate detection methods. Please implement those analyses and discuss the performance of the detection methods, particularly in the regimen of low or very low false alarm.
- Please include a statistical significance analysis of the results. In addition, the variance of the results should be also analyzed. The standard deviation of the indices employed for a number of Montecarlo experiments could be implemented, changing randomly the training and testing datasets
Comments on the Quality of English Language
Please see "Comments and Suggestions for Authors"
Author Response
We would like to thank the Reviewer for the questions and remarks that allowed us to revise and update our manuscript. Considering all questions/remarks of the Reviewer, we improved our manuscript and put these updates in red. Please find the detailed responses below and the corresponding revisions along with corrections highlighted in-red. We hope that this revised version meets all needed requirements.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsI think the revision was done well.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe quality of the paper has been improved significantly. All my concerns have been adequately addressed in the revised version of the paper, including the following: improvement of the literal presentation of the paper; extension of the discussion on related methods; and improvement of evaluation of results including analysis of the receiver operating characteristics (ROC) curves and statistical significance of the results. Therefore, I consider, the contents of the paper should be ready for publication.
