A medical procedure-based patient grouping method for an emergency department
Graphical abstract
Introduction
Emergency department (ED) plays a critical role in most of the healthcare systems because it is the frontline of acute care and the main route of admission to hospital. The complexity of an ED system is featured by diverse patient conditions, fluctuating patient demands, and manpower variation. In Hong Kong, the EDs have been prone to violating the ED service pledges such as the maximum waiting time and length of stay due to large patient demand and persistent ED manpower shortage. Grouping patients with similar characteristics would be one of the feasible alternatives to reduce the complexity and uncertainty in the ED management, and hence ensure the delivery of quality and stable services to the ED patients. Some previous researchers have been dedicated to patient grouping but most of them are not successful in clustering patients in consideration of resource consumption [1]. Three commonly used patient grouping criteria are introduced below.
Casemix is a very common approach of patient grouping based on diagnosis-related grouping (DRG). As its name suggests, it only considers the patient clinical diagnosis categories [1]. The basic premise of Casemix is that patients with similar DRGs will consume a similar amount of resources. In many countries, local governments are using Casemix to fund public hospitals. However, people who against this approach may argue that the DRG method is built on the assumption of “average consumption” which could limit its predictive capabilities of identifying total resource consumption. For example, the elderly and frail patients would suffer under the DRG-based payment because they usually consume more ED resources, but only average resource level is assigned to them [2]. Moreover, the DRG method was originally developed to manage hospital patients. Its effectiveness under a more dynamic and pressing ED system has not been tested by any study.
Length of stay (LOS) grouping suggests that LOS is a good proxy measure of resource consumption when the direct measurement is difficult and costly [3], [4]. Many studies have adopted Gaussian mixture models to group hospital patients based on their LOS [4], [5], [6]. However, the appropriateness of using LOS as proxy to resource consumption is questionable when being applied to ED. In ED, most of the patients usually spend a large proportion of their LOS in the waiting for the first doctor consultation, during which little ED resources are consumed. Therefore, the time between the first consultation and disposition (discharge, admission, dead), or hereinafter known as length of treatment (LOT), would be a better measure of ED resource consumption than LOS.
Patient pathway grouping is based on patient physical movements within a healthcare facility. Isken and Rajagopalan [7] have proposed an approach of grouping patients according to the pathway they took within hospital area. Maruster et al. [8] have suggested an alternative approach of grouping patients with respect to the logistic perspective of treatment. Takakuwa and Shiozaki [9] have identified over 70 patterns of patient flows for 9 patient categories in a simulation project for an ED in Japan. One problem of this grouping method is that patients with different medical needs may share the same pathway. Therefore, the ED staff assigned to each pathway must possess a wide range of skills to meet various medical needs. Also, it is quite challenging to optimize the arrangement of material resources for different pathways.
Three grouping methods above cannot be explicitly linked up with the “actual” resource consumption in ED. Given the uncertain nature of the modern ED system, grouping patients requiring common medical procedures is worthy of investigation if one wants to modularize the ED management for better decision making especially under scarce resources. It is, thus, the objective of this paper to use a data-driven method to group patient with similar pattern of resource consumption. As the name indicates, the data-driven method makes solutions totally based on the data, and it does not involve any human knowledge which might bias the grouping result.
The organization of the paper is as follows: Section 2 introduces several grouping methods and describes the study workflow. Section 3 describes the implementation of the proposed method using the real data followed by result discussion. Section 4 discusses how the patient grouping can be helpful to the ED manpower planning and charging policy. Section 5 concludes this study together with future work.
Section snippets
The proposed methodology
Jain et al. [10] have provided a comprehensive survey of existing grouping techniques and some important applications. Nearly all the grouping algorithms aim to construct clusters with minimal intra-group diversity and maximal inter-group distinction. In this study, the patient grouping is a non-supervised learning process because no patient groups can be pre-defined from any specific ED knowledge. Therefore, three unsupervised grouping techniques are considered here, namely, hierarchical
Data collection, cleansing, and preprocessing
The data contains 2452 records of Category 3 and 4 patients in a random week of 2012. It was collected manually from the archived ED forms. The reason of selecting patient records of two out of the total five categories (Appendix) is that the two categories make up more than 90% of the total ED patients. Each data entry refers to one patient record including patient demographics, triage category, disposition, and medical procedures. The data is cleaned for unclear handwriting and inconsistent
Clinical interpretation
The grouping results show that SOM can successfully cluster patients with common medical procedures, and the inter-group difference in resource consumption is confirmed statistically significant using Wilcoxon rank sum test.
Although the grouping process does not involve patient conditions, each of the resulting groups is found distinctive in the patient's presenting complaints. Group 1 (10.9% Category 3 and 5.9% Category 4) consists of patients with cardiovascular complaints such as chest pain,
Conclusion
In this study, a medical procedure-based patient grouping method is proposed, and several grouping techniques are applied to the real ED data of Categories 3 and 4 patients because they account for more than 90% of the total visits. The best grouping method (i.e. SOM) is selected by the Davies-Bouldin's validity index which is one of the commonly used measures for data clustering. Using SOM, each of the resulting patient groups is featured by one “core” procedure with distinct mixture of other
Acknowledgements
The work described in this paper was partially supported by a grant from the City University of Hong Kong SRG project no. 7002773. The authors would like to thank several anonymous reviewers for their helpful comments and suggestions on this paper.
References (24)
- et al.
Logistic-based patient grouping for multi-disciplinary treatment
Artificial Intelligence in Medicine
(2002) - et al.
Integration of self-organizing feature map and k-means algorithm for market segmentation
Computers & Operations Research
(2002) - et al.
A comparison of SOM neural network and hierarchical clustering methods
European Journal of Operational Research
(1996) - et al.
Segmenting visitors of cultural events by motivation: a sequential non-linear clustering analysis of Italian Christmas Market visitors
Expert Systems with Applications
(2012) - et al.
Modeling daily patient arrivals at Emergency Department and quantifying the relative importance of contributing variables using artificial neural network
Decision Support Systems
(2013) - et al.
Case mix definition by diagnosis related groups
Medical Care
(1980) - et al.
Hospital case mix and average charge per case: an initial study
Health Services Research
(1979) - et al.
Classification trees: a possible method for iso-resource grouping in intensive care
Anaesthesia
(1998) - et al.
Length of stay-based clustering methods for patient grouping
Intelligent Patient Management Studies in Computational Intelligence
(2009) - et al.
Phase-type survival trees and mixed distribution survival tree for clustering patients’ hospital length of stay
Informatica
(2011)
Patient grouping optimization using a hybrid self-organizing map and Gaussian mixture model for length of stay-based clustering system
Data mining to support simulation modelling of patient flow in hospitals
Journal of Medical Systems
Cited by (17)
Queueing Problems in Emergency Departments: A Review of Practical Approaches and Research Methodologies
2022, Operations Research ForumA hybrid data mining approach for generalizing characteristics of emergency department visits causing overcrowding
2019, Journal of Library and Information StudiesMatching patients and healthcare service providers: a novel two-stage method based on knowledge rules and OWA-NSGA-II algorithm
2019, Journal of Combinatorial OptimizationEpidemic-logistics modeling: A new perspective on operations research
2019, Epidemic-logistics Modeling: A New Perspective on Operations Research