Human Activity Recognition for Elderly People Using Machine and Deep Learning Approaches

Round 1

Reviewer 1 Report

interesting paper and the work is useful. suggest to incorporate more datasets for the validation - only the UCI dataset is insufficient. Suggest Major revision.

Author Response

Point 1: interesting paper and the work is useful. suggest to incorporate more datasets for the validation - only the UCI dataset is insufficient. Suggest Major revision.

  

Response 1: Thank you very much for the review. We think this is an excellent suggestion. We have added one more dataset “Activity recognition with healthy older people using a batteryless wearable sensor Data Set” [1] to validate our results.

Satisfying results were achieved by the proposed methods for the new dataset also. A detailed explanation is presented in the paper [Section 4, Table 5].

 

References

[1] Shinmoto Torres, et al., “Sensor enabled wearable RFID technology for mitigating the risk of falls near beds”, In 2013 IEEE International Conference on RFID (pp. 191-198), IEEE.

 

 

Author Response File: Author Response.docx

Reviewer 2 Report

The study focuses on the important problem of recognizing the activity of the elderly. Gyroscope and accelerometer data collected from a smart phone are used with this model. Activities such as Sitting, Walking, Going Upstairs, Going Downstairs, Standing, Laying are studied and well-known Machine Learning and Deep Learning methods are used to recognize them.

Remarks.

1. The title of the article is about Elderly People and in Abstract: There are more than 962 million people aged 60 and up globally (line 1). But in section 3.1 to solve the problem of recognizing the activity of older people used a dataset from UCI (University of California Irvine). Data was collected from a variety of participants aged between 19 to 48 who had a smartphone (lines 141, 142). It is not clear who the authors consider to be elder people aged 60 and over or between 19 to 48. And whether it is possible to draw conclusions about people aged 60+ from people aged 19 to 48? It is necessary to clearly describe and prove the correctness of their conclusions. The recommendation of the article for publication depends on it.
2. Along the way. Almost half of the literature sources are outdated (older than 5 years). You may need to link to more recent research. By the way, the base itself [30] is quite old.
3. Section 3 describes the well-known methods. It can be reduced.
4. I do not think that on modern hardware development processing time is a limitation. Moreover, this is not a real-time problem.
5. The authors' conclusion that the elder people usually follow the same routine activity day (line 250) is questionable.

Author Response

The study focuses on the important problem of recognizing the activity of the elderly. Gyroscope and accelerometer data collected from a smart phone are used with this model. Activities such as Sitting, Walking, Going Upstairs, Going Downstairs, Standing, Laying are studied and well-known Machine Learning and Deep Learning methods are used to recognize them.

 

Author response: Thank you very much for the detailed review.

 

Point 1: The title of the article is about Elderly People and in Abstract: There are more than 962 million people aged 60 and up globally (line 1). But in section 3.1 to solve the problem of recognizing the activity of older people used a dataset from UCI (University of California Irvine). Data was collected from a variety of participants aged between 19 to 48 who had a smartphone (lines 141, 142). It is not clear who the authors consider to be elder people aged 60 and over or between 19 to 48. And whether it is possible to draw conclusions about people aged 60+ from people aged 19 to 48? It is necessary to clearly describe and prove the correctness of their conclusions. The recommendation of the article for publication depends on it.

 

Response 1: Thank you for pointing this out. We have added one more dataset for cross-validation which contains 14 participants aged between 66 to 86 years to check the performance of the system [1]. HAR using a smartphones dataset is used to train the networks, which is a general human activity dataset, but it contains more complex features that will help model in training the network to achieve better generalization. Performance assessment with the new dataset is also added in the paper for validation which will further improve the robustness of the system [Section 4, Table 5].

 

Point 2: Along the way. Almost half of the literature sources are outdated (older than 5 years). You may need to link to more recent research. By the way, the base itself [30] is quite old.

 

Response 2: Thank you for the suggestion. We have improved the literature review [Section 2] by adding more recent studies. About the base dataset [30], yes, it is an old dataset but there are limited public datasets available among which only this dataset offers competent diversification in terms of sophisticated features and a wide range of participants for better generalizability.

 

Point 3: Section 3 describes the well-known methods. It can be reduced.

 

Response 3: We are grateful to the reviewer for pointing out this important concern about the description of state-of-the-art machine and deep learning methods. We also think that these methods should be presented in an abstract way, but since it is a comparative analysis of machine and deep learning methods, so it is important to discuss the methods used for this analysis. However, as per the reviewer’s suggestion, we have presented all the existing methods in the best possible abstract manner.

 

Point 4: I do not think that on modern hardware development processing time is a limitation. Moreover, this is not a real-time problem.

 

Response 4: We agree with the reviewer that with modern hardware processing time may not be the limitation of our dataset. But in the future, it is interesting to see the performance of the proposed models with respect to processing time also, as the data is coming from the sensors, and it is growing continuously. Especially with elder people, it is very important to detect the activities as soon as possible, as if some unusual activity has been seen, appropriate action would be taken as quickly as possible.

 

Point 5: The authors' conclusion that the elder people usually follow the same routine activity day (line 250) is questionable.

 

Response 5: We agree with the reviewer that with modern hardware processing time may not be the limitation of our dataset. But in the future, it is interesting to see the performance of the proposed models with respect to processing time also, as the data is coming from the sensors, and it is growing continuously. Especially with elder people, it is very important to detect the activities as soon as possible, as if some unusual activity has been seen, appropriate action would be taken as quickly as possible.

 

References

[1] Shinmoto Torres, et al., “Sensor enabled wearable RFID technology for mitigating the risk of falls near beds”, In 2013 IEEE International Conference on RFID (pp. 191-198), IEEE.

[2] Chifu, V.R., et al., “Identifying and Monitoring the Daily Routine of Seniors Living at Home,” Sensors 2022, 22, 992. https://doi.org/10.3390/s22030992

 

Author Response File: Author Response.docx

Reviewer 3 Report

This paper applies several existing machine/deep learning methods to HAR. The dataset is a public one. The research questions are not novel. There are a bunch of existing works that have applied these methods to HAR, such as summarized in the paper "Deep Learning for Sensor-based Human Activity Recognition: Overview, Challenges, and Opportunities" CSUR 2021. The authors claim that they focus on elder people, but there seems no specific design for this purpose. 

Author Response

Point 1: This paper applies several existing machine/deep learning methods to HAR. The dataset is a public one. The research questions are not novel. There are a bunch of existing works that have applied these methods to HAR, such as summarized in the paper "Deep Learning for Sensor-based Human Activity Recognition: Overview, Challenges, and Opportunities" CSUR 2021. The authors claim that they focus on elder people, but there seems no specific design for this purpose.

 

Response 1: Thank you very much for the review. We agreed with the reviewer. We have added one more dataset “Activity recognition with healthy older people using a batteryless wearable sensor Data Set” [1] to validate our results. This dataset was performed on 14 participants aged between 66-86 years old. This dataset was used for cross-validation. Proposed methods were shown excellent results in this dataset too. So, it is possible to say that the proposed methods can be used for recognizing the activity of elder people. A detailed explanation of the results is presented in the paper. 

 

References

[1] Shinmoto Torres, et al., “Sensor enabled wearable RFID technology for mitigating the risk of falls near beds”, In 2013 IEEE International Conference on RFID (pp. 191-198), IEEE.

Author Response File: Author Response.docx

Reviewer 4 Report

The document presents research results of comparing several machine learning algorithms for HAR for elderly people using smart phones. It is well organized.

The document can be published if the following issues are clearly specified:

(1) The HAR system does not seem to be tested to recognize 'real' activities of elderly people. If the system was not really applied and tested in recognizing activities of elderly people, the results of the document could not be regarded as trustful and useful to real people.

(2) The proposed LSTM network was tested with the dataset from UCI which was collected from participants aged between 19 and 48. Can it be sure if the dataset is working good to recognize activities of elderly people who are  aged more than 60 or 70?

(3) In the 'Conclusion' part, the HAR system can be used to detect 'a sudden change'. How can the proposed HAR system detect the change and do the 'right' action in time? 

Author Response

The document presents research results of comparing several machine learning algorithms for HAR for elderly people using smart phones. It is well organized.

 

Author response: Thank you very much for the detailed review.

 

Point 1: The HAR system does not seem to be tested to recognize 'real' activities of elderly people. If the system was not really applied and tested in recognizing activities of elderly people, the results of the document could not be regarded as trustful and useful to real people.

 

Response 1: Thank you for your concern, we also agree with the reviewer’s comment.  Collecting a data for elder people would require ethical permission which can take more than a year as per the author’s current country of residence. To avoid delaying this study we have utilized the publicly available dataset and proposed deep learning based methods for HAR. Furthermore, we have also added one more dataset “Activity recognition with healthy older people using a batteryless wearable sensor Data Set” [1] to test our results for elder people. This dataset was performed on 14 participants aged between 66-86 years old.

 

Point 2: The proposed LSTM network was tested with the dataset from UCI which was collected from participants aged between 19 and 48. Can it be sure if the dataset is working good to recognize activities of elderly people who are aged more than 60 or 70?

 

Response 2: As suggested by the reviewer, the proposed methods were tested on the other dataset which contains general activities of elder people. The proposed LSTM method was able to achieve an accuracy of almost 95% [Section 4, Table 5]. So, by the results, it is possible to say that the proposed architecture is robust and generalized and can be used for recognizing elder people’s activities.

 

Point 3: In the 'Conclusion' part, the HAR system can be used to detect 'a sudden change'. How can the proposed HAR system detect the change and do the 'right' action in time?

 

Response 3: Generally elder people whose age is above 60, use the same pattern of activities everyday [2]. So, if any sudden change can be seen in their regular activities e.g., standing or walking to laying down, appropriate action could be taken within time. However, due to the insufficient availability of related data, this problem was not addressed in this research.     

 

References

[1] Shinmoto Torres, et al., “Sensor enabled wearable RFID technology for mitigating the risk of falls near beds”, In 2013 IEEE International Conference on RFID (pp. 191-198), IEEE.

[2] Chifu, V.R., et al., “Identifying and Monitoring the Daily Routine of Seniors Living at Home,” Sensors 2022, 22, 992. https://doi.org/10.3390/s22030992

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

the authors added more validations - no more comments from this reviewer.

Author Response

Thank you very much for the review. We are grateful to you for your comments
and suggestions.

Reviewer 2 Report

The authors of the article took into account basically all my comments. 

Author Response

Thank you very much for the review and suggestions.

Reviewer 3 Report

While this paper proposes to use simple ML algorithms for HAR, there are many existing complex DL methods, such as "multi-agent attentional activity recognition" IJCAI'19. Please review these methods in your introduction and summarize their drawbacks, and motivate your work. 

Author Response

We are grateful to the reviewer for pointing out this important concern. This study is doing a comparative analysis of machine learning and deep learning methods and found out deep learning- based LSTM algorithm was outperforming other methods.
We have agreed with the reviewer and added new articles [1]–[4] which use complex DL methods in the introduction and found out that despite excellent results those methods are not validated for the elder people dataset.

Text Added in the manuscript:
[“Kaixuan Chen et al. [13] in their study gave an overview of challenges and opportunities in the area of human activity recognition using deep learning techniques. Due to the diversity of sensor data, it is really important to have multimodal features which could help in maximizing the performance of the system [14]. The attention-based mechanism could be used in HAR to highlight
the most important and differentiable modalities (Zeng et al. [15], and Chen et al. [16]). Chen et al. [16] used multi agents to focus on modalities that are related to sub-motions. Although they have outperformed all the state-of-the-art methods, still they are not validated the elder people dataset.”]

References
[1] K. Chen, D. Zhang, L. Yao, B. Guo, Z. Yu, and Y. Liu, “Deep Learning for Sensor-based Human Activity Recognition,” ACM Computing Surveys (CSUR), vol. 54, no. 4, May 2021, doi: 10.1145/3447744.
[2] K. Chen, L. Yao, D. Zhang, X. Wang, X. Chang, and F. Nie, “A Semisupervised Recurrent Convolutional Attention Model for Human Activity Recognition,” IEEE Transactions on Neural Networks and Learning Systems, vol. 31, no. 5, pp. 1747–1756, May 2020, doi: 10.1109/TNNLS.2019.2927224.
[3] M. Zeng et al., “Understanding and improving recurrent networks for human activity recognition by continuous attention,” Proceedings - International Symposium on Wearable Computers, ISWC, pp. 56–63, Oct. 2018, doi: 10.1145/3267242.3267286.
[4] K. Chen, L. Yao, D. Zhang, B. Guo, and Z. Yu, “Multi-agent Attentional Activity Recognition,” IJCAI International Joint Conference on Artificial Intelligence, vol. 2019-August, pp. 1344–1350, May 2019, doi: 10.48550/arxiv.1905.08948.

Round 3

Reviewer 3 Report

The authors have addressed all my concerns.