Empowering Community Clinical Triage through Innovative Data-Driven Machine Learning

Round 1

Reviewer 1 Report

This paper deals with a very critical area of triage using a novel AI approach with excellent results. 

My only comment is that I believe that this sentence in line 451: "XGBoost's processing time for triaging 2,35,982 patients is remarkably swift, at 0.059 seconds, showcasing a level of efficiency  that far surpasses that achievable by a human-based triage model." is misleading, Yes, the computational speed is remarkable, however this cannot be compared to the actual real life process of triage in a hospital setting. A similar claim is made in the Abstract.

Author Response

Thank you for reviewing our article and providing comments.

Based on your comments, we applied the following changes:

“within 0.059 seconds” was removed from the Abstract.

“at 0.059 seconds, showcasing a level of efficiency that far surpasses that achievable by a human-based triage model” was replaced with “and takes less than 0.01 seconds in our case”.

Reviewer 2 Report

This is a paper that will be of interest to readers both in the Uk and other parts of the world. the triage targets may be different, but the methods will still apply. I liked the introduction to the healthcare system in the UK. I do wonder if your research could identify faults with this method of triage and distinguishing types of care.  Are there issues of the patients falling between the cracks and instances where delays could cause serious consequences.  You do briefly mention this problem.

I would like to see a little more discussion of the group of items mention on lines 45-46.  I have no idea what 111 means.  Is this a typo?  Tell me about Rapid access therapy team versus ambulance service. Who controls the ambulance service?  Is it tertiary care?  What is fall pickup?

I like a reference to the Emergency Security Index as well as the other scales.  Since you later mention the ESI, tell me a little more about it.  How many Levels are there?  What is Level 3?  I also would like to know more about the Heart Failure dataset.  Who creates it?  How often is it updated?  What kinds of data does it contain? Inpatient and outpatient? Specialty and primary care?  Who collects the data? Are regions in the UK important? Does the region impact the scoring? How sensitive are the predictions to the quality of data. In line 282, I assume 3 Oxygen saturation are counted as one? What is CVPU in Figure 2?

In Table 1, what is the importance of data type? Why is systolic BP data type int64 and Diastolic BP in floating64. Does datatype have any influence on the analytics? Are most of the symptoms not binary - have or does not have?  Are you doing degrees of symptoms (mild, moderate, severe)?

Figure 4 is informative.  Excellent way to display comparisons.

In line 432, it seems that you are using primarily statistics for your analytics.  It would be interesting to compare the use of genAI or LLMAI to the triage.  Maybe your conclusions can approach this topic.

You have a rich set of references.  Over all I like this paper.  It was easy to read and for the most part very informative.

Author Response

Thank you for reviewing our articles and your comments.

The uploaded file includes our responses (in blue print) and text modifications.

Author Response File: Author Response.pdf

Reviewer 3 Report

The study is about machine learning based prediction of triage for heart failure in UK.

The paper is well written. The major weak point of the study is that the considered data
 Major weak point is that thge considered data set did not have information on post-triage heart failure patients (line 235). The authors imputed the missing data on their own based on the patient's health record. This approach open the risk of wrong classification.

The following issue, however, must be tackled:

What is the novelty of the mansucript?  A similar study (DOI: https://doi.org/10.1016/j.array.2023.100281 ), not mentioned, in the list of references should be added and compared with.

minor issues

Author Response

Thank you for reviewing our articles and your comments.

Please find our responses (in Italics blue) and text additions below:

The study is about machine learning based prediction of triage for heart failure in UK.

The paper is well written. The major weak point of the study is that the considered data
 Major weak point is that thge considered data set did not have information on post-triage heart failure patients (line 235). The authors imputed the missing data on their own based on the patient's health record. This approach open the risk of wrong classification.

The following issue, however, must be tackled:

What is the novelty of the mansucript? The novelty includes 1) the application of triaging for ambulatory/community services (all previous papers discuss triaging in the emergency department only) and 2) the original feature generation approach to triaging cardiological patients in community services.

A similar study (DOI: https://doi.org/10.1016/j.array.2023.100281 ), not mentioned, in the list of references should be added and compared with.

The reference has been added and is accompanied by the following text: “Compared to the paper [47], the achieved accuracy is significantly higher, as we utilize a novel feature engineering approach incorporating feature selection and the NEWS2 scoring system”(lines 453-455).