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.
Thank you for your very valuable contribution.
[# PeerJ Staff Note - this decision was reviewed and approved by Mehmet Cunkas, a PeerJ Section Editor covering this Section #]
no comment
no comment
no comment
The authors have addressed all reviewer comments, and the manuscript is suitable for publication. However, the figures are not very clear and should be improved.
I have read the author’s responses in blue text and the manuscript changes in green text.
Regarding question 1 on the use of multi-head attention Transformer architectures instead of LSTM architectures, the authors provided a good justification by explaining how, over short localized sequential pattern intervals of 1-10 seconds of accelerometry data, LSTM architectures work well in classification of falls. For deployment on wearable edge devices, the authors justify the use of embedded LSTM architectures over Transformer architectures due to the quadratic computational complexity of the self-attention mechanism in time and space, with respect to the sequence length. The authors updated their manuscript in the future work section that they will integrate self-attention mechanisms into their LSTM-based framework to enable their model to focus on critical temporal segments.
Regarding question 2 on my recommendation to train their model on the MobiFall and MobiAct datasets, the authors provided a good justification by explaining how different sampling rates were used in MobiFall and MobiAct, which would require resampling and other preprocessing actions to be compatible with the author's model.
Regarding question 3 on the use of a window stride to include an overlap between successive segments, the authors explain the inclusion of N/2 samples from the preceding window in each subsequent window. The authors updated their manuscript to clarify this.
Regarding question 4 on the use of accelerometry datasets derived from self-imposed falls conducted by young persons that may not represent the true biomechanical and behavioral of falls by older persons, the author’s have trained their model on the San Diego State University HealthLINK_Falls dataset which is comprised of accelerometry data from induced falls rather than self-imposed falls. This is an excellent addition to the manuscript.
Based on the author’s improvements made to the manuscript, I recommend acceptance.
The main contribution of the author's work lies in the specific set of features derived using signal processing techniques to train an LSTM/RNN model for classifying falls and ADLs. In the manuscript, there is a scheme to preprocess raw IMU (accelerometry and orientation rate) signals from wearable devices and extract relevant features from these signals for model training. Please follow the requests and suggestions of the reviewers assiduously.
**PeerJ Staff Note:** Please ensure that all review, editorial, and staff comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
Before using abbreviations, provide full forms like PSO, LSTM.
Take care of all abbreviations
Related work needs improvement
A table to add advantages and limitations used from existing articles.
Feature Extraction
Can you prepare a table of comparison of all techniques discussed
Discuss limitations, advantages, time complexity, space complexity, and whatever is applicable.
ASSESSMENT METRICS
All metrics mentioned are well-known to ML users
Add more information on why those are suitable for your research
-
-
Brief Summary:
The proposed fall detection technique utilizes a sophisticated pipeline involving sixth-order Butterworth filtering and Hamming window segmentation for preprocessing. It extracts a rich set of features, including State Space Correlation Entropy (SSCE), Mel Frequency Cepstral Coefficients (MFCC), Linear Predictive Cepstral Coefficients (LPCC), Parseval's energy, and Auto-Regressive (AR) coefficients. These features are then optimized using Particle Swarm Optimization (PSO) before being fed into LSTM networks for classification. Evaluated on the UP-Fall and UR-Fall public datasets, the method demonstrates a significant improvement in classification accuracy compared to traditional approaches, highlighting its exceptional performance for fall action detection systems.
Strengths of the paper :
1. Strong Novelty: The approach combines several advanced techniques (Butterworth filter, Hamming window, SSCE, MFCC, LPCC, Parseval's energy, AR coefficients, PSO optimization, and LSTM networks) in a unique way for fall detection. This multi-faceted approach suggests a sophisticated methodology.
2. Comprehensive Feature Engineering: The use of a diverse set of feature extraction methods (SSCE, MFCC, LPCC, Parseval's energy, and AR coefficients) indicates a thorough attempt to capture various relevant characteristics of fall actions.
Effective Feature Selection: Employing Particle Swarm Optimization (PSO) for feature selection is a strength, as it likely helps in identifying the most discriminative features and reducing dimensionality, potentially improving model efficiency and generalization.
3. Robust Classification: The choice of Long Short-Term Memory (LSTM) networks is well-justified for temporal sequence classification tasks like fall detection, as LSTMs can effectively model the temporal dependencies in the data.
3. Significant Performance Improvement: The abstract explicitly states a "substantial improvement in classification results compared to traditional approaches" and "enhanced accuracy outcomes." This is a key positive finding that supports the effectiveness of the proposed method.
3. Validation on Public Datasets: The experimental assessment using two publicly available datasets (UP-Fall and UR-Fall) adds rigor and allows for potential comparison with other existing methods.
4. Clear Application Focus: The abstract clearly highlights the methodology's exceptional performance specifically for "fall action detection systems," indicating a focused and relevant contribution to the field.
5. Well-Structured and Concise: The abstract provides a clear and concise overview of the proposed technique, its evaluation, and its key findings.
Ways to improve the paper
1. Before using abbreviations, provide full forms like PSO, LSTM
Take care of all abbreviations.
.
2. Are you planning any explainability in your model to make black box to a transparent model, at least add it in the future scope as 1 or 2 sentences about how to make your model transparent
Enhancing Transparency in Smart Farming: Local Explanations for Crop Recommendations Using LIME
Role of Explainable AI in Crop Recommendation Technique of Smart Farming
3. Related work needs improvement
A table to add advantages and limitations used from existing articles.
4. Feature Extraction
Can you prepare a table of comparison of all techniques discussed
Discuss limitations, advantages, time complexity, space complexity, and whatever is applicable
5. ASSESSMENT METRICS
All metrics mentioned are well-known to ML users
Add more information on why those are suitable for your research
6. Is there a possibility to consider the self-attention mechanism in your research? Add or propose a future scope on how the self-attention mechanism can be used in your research, consider the below paper or any other paper. Analysis of an Intellectual Mechanism of a Novel Crop Recommendation System using Improved Heuristic Algorithm-based Attention and Cascaded Deep Learning Network
7. Future scope is very limited
Add more future scope
As a reader will be more interested in the future scope for further research
The presented approach initiates by applying a sixth-order Butterworth filter to preprocess the raw data, then segmenting the signal using a Hamming window. After segmentation, crucial patterns are identified through a comprehensive feature extraction process. This process incorporates various descriptors, including State Space Correlation Entropy (SSCE) coefficients, Mel Frequency Cepstral Coefficients (MFCC), Linear Predictive Cepstral Coefficients (LPCC), Parseval's energy, and Auto-Regressive (AR) coefficients. The extracted features are refined through Particle Swarm Optimization (PSO) to identify the most relevant subset. Classification is performed using Long Short-Term Memory (LSTM) networks. The effectiveness of this methodology was evaluated experimentally on three publicly available datasets, namely UP-Fall and UR-Fall. The results indicate a notable enhancement in classification accuracy compared to conventional techniques. The superior performance of this framework in detecting fall incidents is attributed to the integration of sophisticated preprocessing steps, robust feature extraction strategies, and LSTM-based classification optimized via PSO.
- The introduction details the technical approach (Butterworth filtering, Hamming window, feature extraction, PSO, LSTM). Still, it does not adequately explain why this combination is superior or suitable for fall detection specifically. A rationale for each stage—particularly why LSTM is ideal for this task or how PSO improves classification—should be supported with either citations or logical argumentation.
- The introduction cites only a single reference (Yu et al., 2020) and does not adequately engage with or position the proposed study within the context of existing literature. A deeper review of related works—highlighting gaps or limitations in current methods—is necessary to justify the novelty and necessity of the proposed approach.
- The research states that three datasets were used in the abstract; however, Figure 1 in the methodology section shows only one dataset. Moreover, the research process itself involves only two datasets. Please explain this discrepancy in detail.
- A disconnect exists between the performance claims in Table 5 and the Limitations section. While the former suggests outstanding classification results, the latter admits to several fundamental shortcomings, particularly in generalization, robustness, and environmental dependency. However, these two elements are not reconciled through empirical validation or critical discussion, leaving the reader uncertain about the proposed system's true practical utility.
- The conclusion largely repeats the methodology (Butterworth filter, Hamming window, feature extraction, PSO, LSTM) rather than synthesizing insights or critically evaluating the findings.
It does not highlight what was learned, what challenges were faced, or why certain decisions were effective or limiting.
- Despite being labeled "Conclusion and Future Work," this section does not specify any future directions. There is no mention of exploring new datasets with greater variability, addressing real-world deployment challenges like noise, energy efficiency, or hardware constraints, and enhancing model generalization across populations or activity types.
The author's manuscript is clear, unambiguous, and uses professional English language throughout the document. The introduction and related work sections are adequately referenced with 13 relevant citations. The author's manuscript structure conforms to PeerJ standards and discipline norms. The introduction adequately introduces the subject and makes it clear that the motivation of this work is to implement a novel IMU sensor preprocessing scheme, followed by an LSTM/RNN classification model, to detect human falls and non-fall ADLs.
The author's content is within the Aims and Scope of the PeerJ journal.
The main contribution of the author's work is in the novelty of the specific set of features derived using signal processing techniques to train an LSTM/RNN model to classify falls and ADLs.
The authors have proposed a scheme to preprocess raw IMU (accelerometry and orientation rate) signals from wearable devices and extract relevant features from these signals for model training.
The author's novel scheme involves the use of a sixth-order Butterworth pass-band filter to remove unwanted high-frequency noise from raw IMU signals and then the use of a Hamming window to reduce spectral leakage. Then, digital signal processing methods are used to extract features from windowed signals. These features include (1) State Space Correlation Entropy (SSCE) coefficients to capture amplitude information, (2) Mel Frequency Cepstral Coefficients (MFCC) to capture the short-term power spectrum of a signal, (3) Linear Predictive Cepstral Coefficients (LPCC) to capture the spectral envelop of a signal, (4) Parseval's energy to captures the total energy of a signal, and (5) Auto-Regressive (AR) coefficients to express the current values of a signal as a linear combination of past values. The features are selected through PSO optimization. The authors extracted these features from the well-known UP-Fall and UR-Fall datasets.
Question 1:
With the growth in popularity of Transformer models in NLP applications, why did the authors choose to train an LSTM model instead of a multi-head attention Transformer model, with LSTM's inherent weakness in being sequential and thus not well-suited for embedded fall detection devices?
Question 2:
Why did the authors neglect to consider training their model on the well-established MobiFall and MobiAct datasets
https://bmi.hmu.gr/the-mobifall-and-mobiact-datasets-2/
and the SisFall dataset
https://pmc.ncbi.nlm.nih.gov/articles/PMC5298771/ ?
Question 3:
Do the authors employ a window stride in their windowing scheme?
Question 4:
The UP-Fall and UR-Fall datasets are derived from self-imposed falls conducted by young persons. Do the kinematics of self-initiated, intentional falls by young persons accurately model the kinematics of unintentional falls by elderly (65+) persons?
Question 5:
In the UP-Fall dataset, IMUs were placed on the left wrist and the right pocket. In the UR-Fall dataset, the IMU was placed at the pelvic region. Previous research suggests the shank (calf) is the optimal region to place an IMU for detecting falls (ref https://ieeexplore.ieee.org/document/9014212). What effect did sensor placement in the dataset have on the author's accuracy and F1 metrics? A dataset of externally induced falls with the IMU placed at the shank position is available via URL https://osf.io/wvea3/files/osfstorage
Figures should be enlarged to full text width. Axis labels on the figure graphs are difficult to read because of the small font size.
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.