This article about a prototype-based case study reports on results of a deep-learning-based human activity recognition approach in an edge computing scenario. The motivating use case seems to be primary care in healthcare facilities. However, motivating scenarios and use cases could be emphasized much more clearly throughout the complete paper, but especially in the introduction. The authors should consider stressing this aspect much more clearly in the Featured Application paragraph of their article. Sadly, this paragraph has still placeholder character.

However, the paper is understandable and easy to read. It fits into

• the machine learning and cloud computing subject area of the Computing and AI section or
• (partly) the applied industrial technologies section

although the healthcare domain is not covered explicitly by the journal.

Generally speaking, there is little to criticize (except a critical discussion on "threats to validity" of the study).

For the final polishing of the minor parts paper, I would like to list the following recommendations for the authors:

• Section 1: Introduction (please provide more motivating use cases here)
• Section 2: Related work (please provide more cloud and edge computing related survey and architecture studies to give the reader some hooks for in-depth follow-up readings: https://doi.org/10.3390/app8081368, https://doi.org/10.1016/j.jss.2017.01.001, https://doi.org/10.1109/FMEC.2018.8364046, https://doi.org/10.1109/FiCloud.2015.35).
• Section 3: DL-HAR Framework: Please explain the potential use cases in more details or list them decidedly presented in a summarizing table.
• Section 4: Experiment and Discussion: Please present the experimental setup (Section 4.3) graphically. Graphical experiment architectures let the reader get the core idea of the experiment faster. Consider my later points about threats to validity and how the overall study design has considered this.
• Figure 12: Is pretty useless because it is only composed of one bar. Think about a different way of presenting the data here.
• Section 5: Conclusion. It should be much more explained why Edge Computing is so vital for this paper. Furthermore, I would not stress nitty-gritty details that Python has been applied in the study. The used programming language should be not essential for the academic value of the paper (and for experts, it is more than expectable because of the machine learning domain).

My main point is a missing "threats to validity" discussion which give reasons for a minor revision of the paper.

It should become much more apparent for the reader that this study reports on a prototype implementation that has been fed with two datasets only, not with live data. So, the research has some synthetic character and only shows that the setting with Rasperry PIs can process these two datasets adequately. Therefore, I recommend adding a "threats to validity" section that reports on internal and external threats on the validity of this study, and how these threats have been addressed methodological by the study design. This "threats to validity" reporting is common academic practice and should also give some direction for the reader what can and what should not be derived from this study. For instance, the research can little say about processing live data outside the scope of the two datasets. The authors should also reflect how representative these two datasets are and why they have been chosen.