Semantic Technologies in Sensor-Based Personal Health Monitoring Systems: A Systematic Mapping Study

The authors present a mapping study on sensor-based health monitoring systems using semantic technology. A general background is given on the use of ontologies, KG and linked data for sensor data and health domain. The authors discuss other reviews that tackle related fields, but not the particular intersection that is targeted here. The PRISMA guidelines are used for guiding the mapping study. The study classifies and ranks 40 systems around 6 key challenges (and their sub-categories); describes common system architectures and data collection, development, and evaluation aspects. The authors observe multiple gaps, including: situational prediction and handling of uncertainty is inadequately addressed, nearly none of the systems utilize syntactic standards (e.g., FHIR), and there is little use of established guidelines to guide recommendations, among others.

The mapping study is quite impressive and comprehensive. I believe it has the potential to be a very useful review of the field in particular, and of the use of semantic technology in general. Further, it gives a good introduction on the topic in general. The paper is well written and the language is clear.

However, I have a few comments that I think should be resolved before publication:

The authors promise a thorough review of how semantic technology is used to overcome the challenges (this is objective 3 of the study). However, starting from the "context-awareness" part (Section 5.2), very little is mentioned about semantic technology. The authors mention that "Contextual information can be represented using semantic technologies, and most of the systems do so", and Table 7 lists the types of context that are covered by semantic technology. But, a (critical) reflection seems missing. For instance, what types of ontologies are popular for these particular purposes (e.g., context, situation); which types of knowledge graphs are often used (and using which platform, e.g., Jena, StarDog); what types of reasoning are implemented in CDSS (DL, first order); what types of rules / formalism are being used (SWRL, SHACL..)? I believe most of the content is in the paper, and may be resolved (to an extent) by re-shuffling some content from the background, semantic interoperability, and development languages section. This would allow the reader to get a better idea how semantic technology is used to overcome the challenges - currently this has to be puzzled together.

In a similar vein, the conclusion section could benefit from a reflection on (better) use of semantic technology in PHMS. Aside from the part on sensor ontologies, for instance, the part on explainability could be fleshed out; it is indeed an important feature of symbolic reasoning, but is it utilized in semantics-based PHMS? Are state-of-the-art reasoning technologies, such as SHACL, rdf-star, and Notation3 being utilized? Are ontologies being sufficiently re-used or are application-specific ontologies being used too often?

In the conclusion, I also invite the authors to revisit the final conclusion on generalizability of PHMS. Perhaps I'm misunderstanding, but it can be argued that disease-specific systems are more useful that a "one-size-fits-all" approach. Systems for COPD management need to focus on EMI to avoid flare-ups, whereas diabetes rather focuses on long-term behavior change (and apply psychological theory to that end). They have different management requirements and thus require different solutions. Regarding the lack of support for data standards (e.g., FHIR), perhaps this is due to the fact that systems (I assume) mostly rely on RDF? (It is true that FHIR has an RDF representation, but it is a bit awkward to use, requiring a whole lot of blank nodes).

The radar figures can be quite useful, but I feel they are currently giving a bit of a warped view. For instance, scoring "high" for DSS would involve quite a lot of DSS features, and go beyond the needs of a particular chronic illness. From Fig. 6, one would get the impression that most systems do not support DSS - they do, but simply not all possible DSS features the authors can think of.

The proposed reference architecture seems quite straightforward and even a bit reductive, compared to presented architectures. Why not separate important and distinct functionality, such as sensor, network and data storage, into separate layers? Why is it better to put them into a single layer? Why not go for a modular approach? There is insufficient motivation - the statement that it is "consistent with the layered architectures proposed in related reviews" is insufficient.

With regards to the evaluation setup: criterion 3 is not met by the mapping study, at least not as described by the authors. This criterion pertains to the quality of the individual studies (e.g., level of evidence, internal quality, sample size). (Note that some mapping studies even use the lack of robust evaluation (e.g., RCT) as _exclusion criteria_ for individual studies.) Instead, the discussion in Section 7 could be used to justify that the mapping study meets the criterion.

As a minor comment, the authors ruled out non open-access journals - I'm wondering how many papers were ruled out because of this? I don't think this is a common exclusion criterion, but I could be wrong.

Other comments:
- The authors provide a good description of 3 semantic technologies, but I'm unsure whether the basic intro's (in sections 2.2.1, 2.2.2, 2.2.3) are needed for the SWJ intended audience.

- Reasons are provided for using ontologies and LD in PHMS, but not for KG, it seems.

- Instead of ICD 10, refer to its latest (2019) version, ICD 11 (or, you could make a point that the latter is not being used in most systems)

- Provide a reference for: "Additionally, ambient sensors are increasingly being incorporated in health monitoring to monitor the state of the external environment [..]"

- The following should be better supported:
"systems that re-use existing ontologies have a higher degree of expressiveness for sensor and sensor data concepts than those that do not"

- This is a redundant statement: "Related to duration is frequency."

- What is meant by ambiguous sensor data?

- Please explain "photoplethysmography" when first used

- What is drawback of rules being based on existing knowledge?
"Secondly, the manual creation of rules is time-consuming, difficult to scale, and is often static and based on existing knowledge."

- Please elaborate on BSN acronym

- "Similarly, Ali et al. [93] tested compared" -> "Similarly, Ali et al. [93] compared"

The authors presented a comprehensive review of the frameworks, benefits, challenges, and future prospects of semantic technologies applied in the context of personal health monitoring systems. The authors have provided a thorough analysis of existing literature, making this paper a valuable resource for both researchers and practitioners interested in this field.
The paper is well-structured and enables readers to navigate through the wide research area of sensor-based personal health systems. The authors begin by defining the key concepts and terminologies associated with semantic technologies and their relevance to healthcare. They then delve into various applications, such as situation detection, disease prediction and decision support, highlighting the role of semantic technologies in enhancing the efficiency and effectiveness of these systems. The work presents a balanced coverage of various methodologies and techniques employed in the field. It explores different approaches, including semantic interoperability frameworks, ontology-based proposals for modeling healthcare data and contextual information.

While the survey paper offers a comprehensive overview of the topic, it is important to acknowledge that the depth of coverage may vary across different subtopics. Due to the breadth of the subject matter, aspects related to real-time processing and annotation of sensor data and the integration of wearable devices (sec. 3.2/3.3) could have benefited from more in-depth analysis.

Throughout the survey paper, the authors identify six basic key challenges in health monitoring systems. Some of them (such as context awareness and decision support) have been extensively described in previous works. Novel challanges should be addressed to enrich the discussion with forward-thinking perspectives. Two research areas that deserve special attention are explainable AI and human-centered design techniques, becoming increasingly sophisticated and crucial in current healthcare domains where trustworthiness and usability are essential.

Section 6 can be considered relatively simplistic compared to the other sections. The authors introduce a generic reference architecture, which could benefit from a more comprehensive and detailed description. To enhance the value and practicality of the reference architecture, the authors could consider dividing each layer into sub-layers and providing a more in-depth exploration of the techniques and tools required to address each key challenge. Examining the specific methodologies and technologies applicable to each sub-layer, the authors can offer a specific understanding of the implementation and operational aspects of the proposed architecture. Furthermore, by detailing these aspects, the authors can provide readers with practical guidance and facilitate the integration of semantic technologies and IoT tools/devices/protocols into personal health monitoring systems. This level of granularity would enable researchers and practitioners to gain insights into the technical and research issues of the three layers. This would strengthen the overall completeness and usefulness of the survey paper, providing readers with a clearer roadmap for implementing semantic technologies in novel personal health monitoring systems.

No resources or data file have been provided. It is highly recommended that the authors provide additional resources. One valuable suggestion to enhance accessibility and long-term repository discoverability is the creation of a GitHub page specifically dedicated to the survey paper. It can be used to share a variety of supplementary materials that complement the survey paper's content. This can include code links to datasets, ontologies, use cases, real-world scenarios and relevant resources. By sharing these materials, the authors enable interested readers to access and utilize the resources conveniently, fostering reproducibility and further experimentation. Moreover, a GitHub page offers the advantage of version control, ensuring that any future updates, revisions, or additional materials can be easily managed and accessed by the community.

The paper provides a broad overview in the context of semantic technologies and sensor-based monitoring, addressing key concepts and challenges in a manner that is accessible to those new to the topic. The study is comprehensive in its coverage, reviewing 40 systems and thoroughly discussing six main challenges in the field. However, it might benefit from a more in-depth technical analysis of each system for a more balanced view. This paper highlights the current uses and future potentials of these technologies in health monitoring.

Strengths:
+ The paper systematically reviews 40 systems, offering an extensive overview of current technologies and methodologies in the field.
+ Identifies and discusses six primary challenges in the domain, such as interoperability and context awareness, providing a clear understanding of the field's current state.
+ The papers are summarized into several tables making it easy to grasp the underlying core technologies easily.
+ The paper offers valuable insights and recommendations for future research, which can guide subsequent studies and developments in this domain.

Weaknesses:

- Some technical depth into the core technologies discussed in the review would have been helpful. For instance, the paper could benefit from a nuanced analysis of the technologies specified in Table 4, i.e., how aligned are the layered, multi-agent, service-oriented, modular, etc., architectures between all these different systems? Also, in Table 5, along with the type of data collected, an analysis of what type of devices were used to collect the data and how they interoperate with semantic systems could have been useful to summarize in the table or otherwise.

- The authors have identified six key challenges: interoperability, context awareness, situation detection, situation prediction, decision support, and uncertainty handling. How these were selected was not clearly articulated in the paper. For instance, other challenges with respect to sensors should have been addressed. These include challenges such as data privacy and security ( i.e., how can these semantic sensor systems ensure the confidentiality and integrity of sensitive health data collected by sensors?), scalability (i.e., how can you develop systems that can efficiently handle an increasing amount of data from a growing number of users and devices with semantic technologies?), user acceptance (i.e., how to ensure that these systems are user-friendly and acceptable to a diverse range of users, including those with varying levels of technological literacy), regulatory compliance (i.e., how could these semantic sensor systems could adhere to various healthcare regulations and standards, which can vary by region and are critical for the deployment of health monitoring systems?)