Fall Detection System Based on Simple Threshold Method and Long Short-Term Memory: Comparison with Hidden Markov Model and Extraction of Optimal Parameters

Round 1

Reviewer 1 Report

The authors proposed the STM-LSTM-based fall detection system. The parameters, such as regularization, are well studied to overcome overfitting. The paper is well-written overall, but I have a few comments:

1. In the introduction, can you emphasize the advantages of LSTM and why you chose LSTM instead of LSTM-CNN or GRU in this study?

2. Figure 6(b) doesn't seem right to me. It would also be good to show the confusion matrix's user and producer accuracy.

3. Can you explain why HMM works better than LSTM at P_theta in Figure 7 in the paper？

Author Response

Please find an attached file.

Yun Seop Yu

Author Response File: Author Response.pdf

Reviewer 2 Report

This research suggested an LSTM-based fall detection algorithm considering the threshold method. The fall detection system based on the pendant-type sensor was interesting.From my point of view, the authors are preferred to provide more information to support the claims .

1. Many previous studies have suggested LSTM-based fall detection algorithms. You should to provide more details regarding the differences between this study and the previous studies.

2. Compared with the previous studies, few gestures were selected as ADLs and Falls, and the gestures seem too simple. These experimental method may degrade the practicality of the fall detection algorithm. If the selection of the gestures had some reason, an additional explanation would be necessary.

3. It is necessary to describe why the experimental equipment was designed in the form of a pendant. In the future work, this design may become a critical problem in validating algorithms using public datasets due to structural differences of experimental equipment.

4. The suggestion of an efficient structure considering various factors, such as normalization, regularization rate, and sampling interval, in the development of fall detection algorithm was interesting. However, it looks like this structure will need to be optimized again when new gestures are added or the dataset changes. If this is a limitation of this study, it seems necessary to describe it with more detail.

5. The self-developed embedded edge device was used for this study. However, a reliability of sensor has to verified through comparison with commercial sensors.

6. The explanation of feature parameters and annotations are so complicated. We recommend using the table to explain them.

7. The components of the x-axis are not related to each other in Figure 7,8,9. Thus, the line cannot use in the graph.

8. Some previous studies used normalization and reported good performance. It is necessary to discuss whether the performance degradation is only a problem of normalization or whether generalization is possible.

9.The comparison between the STM-HMM and STM-LSTM is lacked. More detailed analysis was needed. For example, false alarms, confusion matrix etc.

10. Is there any reason about the number of subjects is six? The balance of gender was also unbalanced. It is necessary to discuss them.

Author Response

Please find an attached file.

Yun Seop Yu

Author Response File: Author Response.pdf

Reviewer 3 Report

The article presents a proposal that addresses an STM-LSTM-based system that detects falls. The proposal is based on single and multiple parameters calculated from 3-axis acceleration data. In turn, the authors use a training mode where the parameters are normalized to use LSTM.

In general, the paper is well organized, and section 5 writes the work's methodology well. However, I do not find a section where the results obtained from the study are discussed. Moreover, I do not see what impact the results obtained have from a clinical point of view. There is no doubt that the authors have made a great effort to compare LSTM, but I do not find the impact or consequences of the study for the fall detection community.

On the other hand, with respect to the manuscript, specific paragraphs are disorganized and difficult to understand. For example, the abstract is unstructured and confusing.

Author Response

Please find an attached file.

Yun Seop Yu

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The manuscript is ready to accept.

Best wishes.