Evaluating the Effectiveness of Plantar Pressure Sensors for Fall Detection in Sloped Surfaces

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper introduces plantar pressure sensors into the detection of people falling on the roof slope, and achieves good detection results by combining IMU units. The work is good, but there are still some issues that need further attention.

Main comments:

1. Please provide information about the participants, such as the number of participants, their height, weight, and age distribution, so as to have a clear understanding of the target population of the study.
2. How were the parameters chosen by the authors in each machine learning method, and how can it be ensured that they are the optimal choices to allow the best performance of different methods to be compared? This needs to be discussed by the authors.
3. The information about the dataset used for training and evaluation should be detailed, such as the number normal/fall situations in different postures, and the allocation of data for training, testing, and validation.
4. How would the performance of fall detection change if the number of cooperating IMUs were further reduced? It is suggested that the authors conduct further experimental verification.

Minor comments:

1. For reference 1, attention should be paid to ensuring that the information about WHO is complete and correct.
2. In Figure 1, the schematic diagram and the actual human body diagram indicating the IMU location are repetitive. It is suggested to retain the actual human body installation diagram.

Author Response

Thank you for your constructive and insightful comments, which have greatly helped us improve our manuscript. We have addressed each comment and also highlighted our changes in blue in our annotated manuscript for easy reference.

 

Comment 1: Please provide information about the participants, such as the number of participants, their height, weight, and age distribution, so as to have a clear understanding of the target population of the study.

Response 1: We thank the reviewer for this important observation. We have added participant demographic details (age, height, weight, and gender) to Section 3.2.1 to clarify the characteristics of the study population.

 

Comment 2: How were the parameters chosen by the authors in each machine learning method, and how can it be ensured that they are the optimal choices to allow the best performance of different methods to be compared? This needs to be discussed by the authors.

Response 2: We thank the reviewer for pointing out the need to clarify our hyperparameter selection process. We have added a new subsection (Section 4.10: Hyperparameter Tuning and Optimization) that explains the tuning strategies used for both traditional and deep learning models. For classical models, we used grid search with cross-validation to identify optimal parameters, and for deep learning models, we performed empirical tuning based on validation performance. This update ensures transparency and supports the comparability of model performance results reported in the study.

 

Comment 3: The information about the dataset used for training and evaluation should be detailed, such as the number of non-fall/fall situations in different postures, and the allocation of data for training, testing, and validation.

Response 3: We appreciate the reviewer’s suggestion to add more dataset details. We have added a new table (Table 2 in Section 3.4) that summarizes the distribution of fall and non-fall samples across activities and slope conditions. We also updated Section 5.1 to clearly explain how the dataset was split for training, validation, and testing using stratified 5-fold cross-validation. These changes enhance the reproducibility and clarity of our experimental design.

 

Comment 4: How would the performance of fall detection change if the number of cooperating IMUs were further reduced? It is suggested that the authors conduct further experimental verification.

Response 4: We thank the reviewer for this valuable suggestion. We extended our experiments to include 2-IMU and 1-IMU configurations combined with plantar pressure data. The results show that while performance declines slightly, the hybrid setup continues to outperform full and best 3 IMU-only configurations. For example, CNN: from 0.88 to 0.86 for both one and two IMUs with plantar pressure sensors; XGBoost: from 0.88 to 0.85 for two IMUs and 0.84 for one IMU with plantar pressure sensors. We have added these results and updated the Discussion section accordingly.

 

Minor Comments:
For reference 1, attention should be paid to ensuring that the information about WHO is complete and correct.

In Figure 1, the schematic diagram and the actual human body diagram indicating the IMU location are repetitive. It is suggested to retain the actual human body installation diagram.

Response: Both concerns are addressed in the revised version of the paper.

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript evaluates the effectiveness of the plantar pressure sensors for fall detection on inclined surfaces. Through machine learning-based data processing methods, the integration with plantar pressure sensor techniques has been proven to effectively enhance the accuracy of fall detection. However, this paper lacks critical data and results presentation, such as performance testing and characterization of sensors (plantar pressure sensors and inertial measurement units). Therefore, it is necessary to provide relevant key data, including sensor performance characterization, test results under different positions and postures, etc. Additionally, the data processing and machine learning algorithms should also include corresponding procedures and detailed explanations/demonstrations.

Author Response

Thank you for your constructive and insightful comments, which have greatly helped us improve our manuscript. We have addressed each comment and also highlighted our changes in blue in our annotated manuscript for easy reference.

 

Comment 1: This manuscript evaluates the effectiveness of the plantar pressure sensors for fall detection on inclined surfaces. Through machine learning-based data processing methods, the integration with plantar pressure sensor techniques has been proven to effectively enhance the accuracy of fall detection. However, this paper lacks critical data and results presentation, such as performance testing and characterization of sensors (plantar pressure sensors and inertial measurement units). Therefore, it is necessary to provide relevant key data, including sensor performance characterization, test results under different positions and postures, etc. Additionally, the data processing and machine learning algorithms should also include corresponding procedures and detailed explanations/demonstrations.

Response 1: We thank the reviewer for their detailed and constructive feedback. In response:

 

• We have added a sensor specification table (Table 1 in Section 3.1) describing the sampling rate, resolution, and other characteristics of the plantar pressure and IMU sensors used in this study.
• We acknowledge the importance of evaluating model performance across specific postures and slope conditions. However, due to the limited number of simulated falls per posture and slope combination, we were unable to reliably compute posture-specific performance metrics. We have added a note in the Discussion section to acknowledge this as a limitation and identified it as an important direction for future work.
• To address the lack of clarity regarding the analysis workflow, we have added a new diagram (Figure 4) illustrating the complete data processing and modeling pipeline.
• Additionally, we have added a new subsection (Section 4.10: Hyperparameter Tuning and Optimization) that explains the tuning strategies used for both traditional and deep learning models.

These revisions address the concerns regarding sensor characterization, clarity of data processing, and model interpretability, while also recognizing the need for deeper stratified performance analyses in future studies.

Reviewer 3 Report

Comments and Suggestions for Authors

Dear authors,

This study is a very interesting topic. However, there are some questions to be clarified before publication.

1. In 2. Related works, it is easy to understand because it organises the previous studies and      their papers. However, this is a matter to be discussed in the Introduction of Chapter 1 and is unnecessary for Chapter 2. After organising the research background, previous research and clarification, please change to a statement of purpose to test the hypotheses in this thesis.

 

2. In this study, four models are measured and analysed：IMU (17) IMU (3) Plantar Pressure (2) IMU (3) + Plantar Pressure (2)．However, the placement of wearable sensors such as IMUs is only explicitly stated for 17 models and not for the others. In Figure 1, the placement of the foot pressure sensor is not clearly indicated in the diagrams in this paper. Please add three diagrams (MU (3) Plantar Pressure (2) IMU (3) + Plantar Pressure (2)) and make it clear.

 

3. What is the relationship between the local coordinates of the IMU, the local coordinates of the foot pressure sensor and the global coordinates of the entire measurement system? Please specify this in the paper or add it to figures showing the mounting arrangement, such as Fig. 1, or diagrams of the walking conditions, etc. Does the result output by the IMU perform a coordinate transformation? If so, describe it, if not, state that it is based on local coordinates.

 

4. In Figure 2, we understood the appearance of the various movements we wanted to detect in this paper. However, although the inclination is mentioned in the text as 15 and 30 degrees, the details could not be read from this image. Please add a more clear diagram.

 

5. In Chapter 3, are there any kinematic algorithms, conditions, etc. for fall-turnover in this paper? It is important to present kinematic hypotheses when falling detection in machine learning, but they did not read well in this paper. If so, please describe them. If not, please provide an explanation in Chapter 3, or in the discussion of the limitations, or in the conclusion.
6. In Chapter 3,4 , there is a lack of explanation of the flow of the analysis methods, in particular the flow of conditional branching in the various types of machine learning in fall detection. Please add diagrams to clarify this.
   
   
7. Please add an example of the time-series data of the fall detection for each gait condition, or an explanation and diagram showing how the fall detection in one gait cycle was detected from the acceleration and gyro sensor data. Without this figure, the reader will not be able to understand which movements you are trying to detect correctly by machine learning.
   
   
8. In 7. Discussion, the content seems to be lacking; 2. As with related work, a scientific comparison of relevance, similarity, qualitative or quantitative validity with previous studies and a little more consideration and description of the validity of the results in this study is needed. Please cite and reconsider the literature supporting the results of this paper as it is easy to understand with many references. If you have described them in chapter 6 (results), please reconsider them, including their structure.

Author Response

Thank you for your constructive and insightful comments, which have greatly helped us improve our manuscript. We have addressed each comment and also highlighted our changes in blue in our annotated manuscript for easy reference.

 

Comment 1: In 2. Related works, it is easy to understand because it organises the previous studies and their papers. However, this is a matter to be discussed in the Introduction of Chapter 1 and is unnecessary for Chapter 2. After organising the research background, previous research and clarification, please change to a statement of purpose to test the hypotheses in this thesis.

Response 1: We thank the reviewer for the insightful suggestion regarding the organization of prior work. In response, we have revised the Introduction to integrate and summarize key elements from the Related Work section that directly support the research context and motivation. Specifically, we now provide a more coherent narrative covering the limitations of IMU-based fall detection, the evolution of machine learning approaches, and the underexplored role of plantar pressure sensors, culminating in a clear articulation of the research gap. The Related Work section has been retained to present detailed technical categorizations and citations of previous models and sensor strategies. These changes improve the logical flow of the manuscript and ensure the study’s purpose and hypothesis are introduced with appropriate background and clarity.

 

Comment 2: In this study, four models are measured and analysed：IMU (17) IMU (3) Plantar Pressure (2) IMU (3) + Plantar Pressure (2)．However, the placement of wearable sensors such as IMUs is only explicitly stated for 17 models and not for the others. In Figure 1, the placement of the foot pressure sensor is not clearly indicated in the diagrams in this paper. Please add three diagrams (MU (3) Plantar Pressure (2) IMU (3) + Plantar Pressure (2)) and make it clear.

Response 2: We thank the reviewer for highlighting the need to clarify sensor placement across different configurations. For the reduced IMU (3) configuration, we have now explicitly stated that the sensors were placed on the pelvis, left hand, and right hand, selected based on their high correlation with fall events. Also, the plantar pressure sensor placement on both feet.

 

Comment 3: What is the relationship between the local coordinates of the IMU, the local coordinates of the foot pressure sensor and the global coordinates of the entire measurement system? Please specify this in the paper or add it to figures showing the mounting arrangement, such as Fig. 1, or diagrams of the walking conditions, etc. Does the result output by the IMU perform a coordinate transformation? If so, describe it, if not, state that it is based on local coordinates.

Response 3: We thank the reviewer for raising this important point regarding coordinate systems. In response, we have clarified in Section 3.2 that IMU acceleration data were used in each sensor’s local coordinate frame, and no transformation to a global coordinate system was applied. Specifically, we computed the magnitude of acceleration across the local x, y, and z axes, which is orientation-invariant and suitable for fall detection without global alignment. Similarly, the plantar pressure sensors operate in a fixed 2D grid aligned with the foot and were used in their native local coordinate space.

While we considered adding illustrations to show the IMU coordinate axes, we ultimately decided not to include additional figures, as we found it challenging to accurately represent 3D coordinate frames in a 2D image without causing potential confusion. Instead, we have provided clear textual descriptions in the manuscript to ensure transparency regarding coordinate handling.

Comment 4: In Figure 2, we understood the appearance of the various movements we wanted to detect in this paper. However, although the inclination is mentioned in the text as 15 and 30 degrees, the details could not be read from this image. Please add a more clear diagram.

Response 4: We thank the reviewer for pointing out the lack of visual clarity regarding the slope conditions in the original posture figure. In response, we have added a new image (now Figure 1) that clearly depicts the roof setup used in our experiments, showing both the 30° (left) and 15° (right) inclined surfaces. 

 

Comment 5: In Chapter 3, are there any kinematic algorithms, conditions, etc. for fall-turnover in this paper? It is important to present kinematic hypotheses when falling detection in machine learning, but they did not read well in this paper. If so, please describe them. If not, please provide an explanation in Chapter 3, or in the discussion of the limitations, or in the conclusion.

Response 5: We thank the reviewer for this valuable comment. We acknowledge that our current study does not incorporate explicit kinematic hypotheses for fall turnover events. Our approach is fully data-driven, relying on the learning capacity of machine learning and deep learning models to extract discriminative features from raw sensor inputs. We have now added a clarification in the Discussion section to reflect this and recognize it as a limitation. In future work, we aim to explore hybrid strategies that combine interpretable kinematic features with learned models to enhance both performance and transparency.

 

Comment 6: In Chapter 3,4 , there is a lack of explanation of the flow of the analysis methods, in particular the flow of conditional branching in the various types of machine learning in fall detection. Please add diagrams to clarify this.

Response 6: We thank the reviewer for this helpful suggestion. To address the lack of clarity regarding the analysis workflow, we have added a new diagram (Figure 4) illustrating the complete data processing and modeling pipeline. This includes sensor input, feature extraction, training usingmachine learning models (traditional and. deep learning) and model evaluation . We believe this visual representation significantly improves the reader's understanding of our methodological approach.

 

Comment 7: Please add an example of the time-series data of the fall detection for each gait condition, or an explanation and diagram showing how the fall detection in one gait cycle was detected from the acceleration and gyro sensor data. Without this figure, the reader will not be able to understand which movements you are trying to detect correctly by machine learning.

Response 7: We thank the reviewer for this thoughtful suggestion. We agree that visualizing sensor patterns during fall events can help readers understand the dynamics underlying fall detection. However, given the large number of variables across 17 IMU sensors and 2 plantar pressure insoles, generating representative time-series plots would require either showing dozens of individual sensor streams or selecting only a small subset, which may not accurately capture the broader patterns. Since meaningful trends may only appear in specific sensors depending on the fall condition and sensor placement, we concluded that including such figures would risk being either overwhelming or unrepresentative. Therefore, we have opted to avoid this form of visual representation in the revised version of the paper.

 

Comment 8: In 7. Discussion, the content seems to be lacking; 2. As with related work, a scientific comparison of relevance, similarity, qualitative or quantitative validity with previous studies and a little more consideration and description of the validity of the results in this study is needed. Please cite and reconsider the literature supporting the results of this paper as it is easy to understand with many references. If you have described them in chapter 6 (results), please reconsider them, including their structure.

Response 8: We thank the reviewer for this constructive comment. In response, we have expanded Section 7 (Discussion) to include a more thorough comparison with prior studies. We now reference several relevant works that used IMUs or plantar pressure sensors for fall detection, and we critically examine how our findings relate to theirs in terms of performance, sensor configuration, and experimental settings. We also discuss the scientific validity of our results, particularly the observed benefit of integrating plantar pressure data with IMUs. These additions provide stronger context and justification for our study’s contributions.

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript presents a study evaluating the efficacy of plantar pressure sensors, both independently and in combination with IMUs, to enhance the fall detection capabilities on sloped surfaces. The primary objective of this study was to demonstrate that the integration of plantar pressure data can significantly improve the accuracy and efficiency of wearable fall detection systems.

The experimental approach involved controlled laboratory trials of a custom-built roof mockup to simulate realistic occupational conditions. A diverse set of machine learning models, including traditional classifiers and deep learning architectures, were evaluated for various sensor configurations. The finding that a reduced sensor set (three IMUs and two plantar pressure sensors) outperforms the full 17-IMU setup has significant practical implications. This provides empirical evidence that integrating plantar pressure sensors significantly enhances fall detection performance. Models incorporating plantar pressure data consistently outperformed those relying solely on IMU inputs across all the tested configurations.

I have made several suggestions and remarks to improve the quality of the manuscript.

Given the importance of synchronization between the IMU and plantar pressure data, the authors could provide more technical details on the synchronization method used.

Although the F1 score is presented as the primary metric, could the authors also provide a breakdown of precision, recall, and accuracy for the top-performing models?

I would appreciate a more detailed discussion regarding the computational complexity and feasibility of the proposed models for real-time processing of resource-constrained wearable devices (even theoretical considerations or estimations of computational load would enhance the practical implications).

This study focuses on fall detection. Could the authors expand on how this system could potentially be used for fall prevention, perhaps by providing real-time feedback on instability or risky postures identified by plantar pressure data?

Overall, this manuscript offers a substantial contribution to the fields of wearable health technologies and occupational safety. In my opinion, this manuscript is technically sound, of great interest to the scientific community, and can be accepted after minor revisions.

Author Response

Thank you for your constructive and insightful comments, which have greatly helped us improve our manuscript. We have addressed each comment and also highlighted our changes in blue in our annotated manuscript for easy reference.

 

Comment 1: Given the importance of synchronization between the IMU and plantar pressure data, the authors could provide more technical details on the synchronization method used.

Response 1: We thank the reviewer for pointing out the need to clarify the synchronization procedure between the IMU and plantar pressure sensors. In response, we have added technical details in Section 3.3.1 describing our synchronization strategy, which involved timestamp-based alignment, manual trigger events, and resampling to a common time base. We also quantified the temporal alignment precision and confirmed that it falls within an acceptable margin for fall detection analysis.

 

Comment 2: Although the F1 score is presented as the primary metric, could the authors also provide a breakdown of precision, recall, and accuracy for the top-performing models?

Response 2: We thank the reviewer for this suggestion. In response, we have added a breakdown of performance metrics: accuracy, precision, and recall for the top-performing model (CNN and MLP with IMU (3) + Pressure (2)). This information is now included in Section 6 along with the F1-score to give a more complete picture of the model's classification performance.

 

Comment 3: I would appreciate a more detailed discussion regarding the computational complexity and feasibility of the proposed models for real-time processing of resource-constrained wearable devices (even theoretical considerations or estimations of computational load would enhance the practical implications).

Response 3: We thank the reviewer for raising this important point. In response, we have added a new paragraph in the Discussion section addressing the computational complexity and potential for real-time deployment. While our models were trained and evaluated offline, we estimate that the CNN used in our study is sufficiently lightweight for real-time inference on wearable platforms. We have also outlined potential strategies for future optimization and deployment on embedded systems.

 

Comment 4: This study focuses on fall detection. Could the authors expand on how this system could potentially be used for fall prevention, perhaps by providing real-time feedback on instability or risky postures identified by plantar pressure data?

Response 4: We thank the reviewer for this thoughtful comment. While the current study is focused on fall detection, we agree that the system has potential applications in fall prevention. In future work, we plan to investigate fall forecasting by analyzing pre-fall movement patterns and instability cues extracted from IMU and plantar pressure data. This could enable early identification of high-risk postures or behaviors, forming the basis for real-time feedback and fall prevention strategies.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revision is good, and all my concerns have been addressed.

A little suggestion about  Figure 3,  its quality should be improved before publication. 

Author Response

Comment 1: A little suggestion about  Figure 3,  its quality should be improved before publication. 

Response: We have updated the figure.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This revision and response have addressed some of my concerns, but still lack and should include the presentation of sensor data, including testing and characterization of the sensors under different conditions.

Author Response

Comment 1: This revision and response have addressed some of my concerns, but still lack and should include the presentation of sensor data, including testing and characterization of the sensors under different conditions.

Response 1: We have included Figure 6 and Figure 7, which illustrate representative IMU acceleration and plantar pressure signals across different activities. These figures demonstrate the distinct signal patterns captured by each sensing modality, highlighting their responsiveness to each activity.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you very much for your prompt and courteous reply. I generally understand your answer. However, there is one matter that I would like to clarify, and I would appreciate your response again.

 

2: We thank the reviewer for highlighting the need to clarify sensor placement across different configurations. For the reduced IMU (3) configuration, we have now explicitly stated that the sensors were placed on the pelvis, left hand, and right hand, selected based on their high correlation with fall events. Also, the plantar pressure sensor placement on both feet.

 

Response: I have confirmed the details. I thought it was good that it was easier to understand. However, the arrangement of the foot pressure sensor could not be read from Figure 1. Please indicate in the image whether they are outside or inside the shoe, and what kind of placement they are. If it is difficult to do so in Fig. 1, please add a new image.

Author Response

Comment 1: I have confirmed the details. I thought it was good that it was easier to understand. However, the arrangement of the foot pressure sensor could not be read from Figure 1. Please indicate in the image whether they are outside or inside the shoe, and what kind of placement they are. If it is difficult to do so in Fig. 1, please add a new image.

Response: We have added a figure showing the placements of the plantar pressure sensors.

Author Response File: Author Response.pdf