Approaches to Extracting Patterns of Service Utilization for Patients with Complex Conditions: Graph Community Detection vs. Natural Language Processing Clustering
, Hanieh Sadri 2, Ken Moselle 3, Ernie Chang 4,†, Yudi Santoso 2, Joseph Howie 2, Abraham Rudnick 5,*, Lloyd T. Elliott 6 and Alex Kuo 1Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsDear Authors,
Your manuscipt titled "Approaches to Extracting Patterns of Service Utilization for Patients with Complex Conditions: Graph Community Detection vs Natural Language Processing Clustering" is very intreresting. Please find my minor comments below.
1. Introduction section could be revised for the clarity and brevity. It's too long and confusing.
2. How were the resolution parameters chosen in the Louvain algorithm ?
3. Although it is stated in the manuscript as a limitation, what was the rationale behind choosing one algorithm over others ?
Best wishes.
Comments on the Quality of English Language
Minor editing of English language required
Author Response
Dear reviewer,
Thank you very much for reading our manuscript and for your feedback. Please, see below our responses to your questions.
Sincerely,
The authors
1. The introduction section could be revised for clarity and brevity. It's too long and confusing.
Answer: The introduction has been revised to ensure clarity.
2. How were the resolution parameters chosen in the Louvain algorithm?
Answer: We used the default value of the resolution parameter. We found that changing this parameter does not improve the modularity value of the output.
3. Although it is stated in the manuscript as a limitation, what was the rationale behind choosing one algorithm over others?
Answer: We did explore other algorithms as well. These include Fast-Greedy, Edge-Betweenness, and Leading-Eigen. However, we did not include them in the paper due to instability issues and low output quality. Subject matter experts found that Louvain, k-Mean, and Hierarchical clustering yield clusters that make the most sense from the clinical perspective. Regarding the additional algorithms that were not explored such as Word2Vec, tSNE, Gaussian mixture model, and structural clustering, as mentioned in the manuscript, they will be explored in future studies.
Reviewer 2 Report
Comments and Suggestions for AuthorsApplying community detection methods to group patients according to the services they receive is an excellent idea to improve the patient experience and the efficiency of medical services. The main contribution and challenging aspect of this paper is to establish a bipartite graph, or even construct a specific graph with services as nodes and common patients as edges, in which the authors utilize the natural language processing to assist the structuring process. This approach is straightforward to operate and implement within medical operation systems. I think it deserves to be published.
However, I have some points as below:
1. The data of the journey of patient is extracted by natural language processing, thus how to ensure the uniform quality of the description in the contents of patients? As we know, even the same type of patients might be described in distinct style. But I have not identify the contents concern this issue.
2. Although the size of modularity value verse number of iterations is illustrated, I still want to understand the reasonable size of the result and their explanation with the viewpoint of medicine treatment. I hope the authors might supply more descriptions on this content.
3. Finally, some recent references might be helpful, e.g.
Mehrdad Rostami, Mourad Oussalah, Kamal Berahmand and Vahid Farrahi, Community Detection Algorithms in Healthcare Applications: A Systematic Review, IEEE Access, 2023, vol.11, pp. 30247-30272.
Author Response
Dear reviewer,
Thank you very much for reading our manuscript and for your feedback. Please, see below our responses to your questions.
Sincerely,
The author
1. The data of the journey of patients is extracted by natural language processing, thus how to ensure the uniform quality of the description in the contents of patients? As we know, even the same type of patients might be described in distinct style. But I have not identify the contents concern this issue.
Answer: Different patients might be described differently from the vantage point, for example, of problems and diagnosis. However, every patient follows the same registration processes, and patients’ interaction with the health service system is described using the same location structure. The registration is done consistently using a rigorous application of data quality control procedures at the point of information captured. And our model is built out of registrations with services.
2. Although the size of modularity value verse number of iterations is illustrated, I still want to understand the reasonable size of the result and their explanation with the viewpoint of medicine treatment. I hope the authors might supply more descriptions on this content.
Answer: With the proposed iterative approach using Louvain community detection, the clusters of services get refined, going from heterogeneous communities of services to homogeneous communities of services. The iterative process continues until the number of communities of services no longer changes. To ensure that communities are statistically and clinically homogeneous, once the iterative process is completed, the optimal iteration is determined by SMEs using their clinical background, but also relying on the values of the weighted degrees. From a clinical perspective, homogeneous services refer to homogeneity with respect to features of the people and services associated with those features. It also refers to homogeneity with respect to the activities that are performed in relationship to a particular issue or problem.
3. Finally, some recent references might be helpful, e.g.
Mehrdad Rostami, Mourad Oussalah, Kamal Berahmand and Vahid Farrahi, Community Detection Algorithms in Healthcare Applications: A Systematic Review, IEEE Access, 2023, vol.11, pp. 30247-30272.
Answer: Thank you. This review paper is indeed useful for providing a broader view on graph clustering, and has included as a reference in our manuscript.
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsDear Authors,
I have no further comments.
Regards.
Reviewer 2 Report
Comments and Suggestions for AuthorsThis revised version responses all my concerns, as well as re-writing many paragraphs to improve the representation. I suggest to ACCEPT it for publication.
