Clustering of a Health Dataset Using Diagnosis Co-Occurrences

Abstract

Assessing the health profiles of populations is a crucial task to create a coherent healthcare offer. Emergency Departments (EDs) are at the core of the healthcare system and could benefit from this evaluation via an improved understanding of the healthcare needs of their population. This paper proposes a novel hierarchical agglomerative clustering algorithm based on multimorbidity analysis. The proposed approach constructs the clustering dendrogram by introducing new quality indicators based on the relative risk of co-occurrences of patient diagnoses. This algorithm enables the detection of multimorbidity patterns by merging similar patient profiles according to their common diagnoses. The multimorbidity approach has been applied to the data of the largest ED of the Aube Department (Eastern France) to cluster its patient visits. Among the 120,718 visits identified during a 24-month period, 16 clusters were identified, accounting for 94.8% of the visits, with the five most prevalent clusters representing 63.0% of them. The new quality indicators show a coherent and good clustering solution with a cluster membership of 1.81 based on a cluster compactness of 1.40 and a cluster separation of 0.77. Compared to the literature, the proposed approach is appropriate for the discovery of multimorbidity patterns and could help to develop better clustering algorithms for more diverse healthcare datasets.

Clustering Diagnoses From 58 Million Patient Visits in Finland Between 2015 and 2018

“…Similar to the study by Wartelle et al [ 23 ], our study was also based on RR. However, there were 2 main differences.…”

“…Most existing studies on clustering are based on hierarchical agglomerative methods using heuristic criteria, either average or complete linkage [ 13 ]. Wartelle et al [ 23 ] extended hierarchical agglomerative clustering by directly optimizing clustering using RR. By default, this is a more solid approach than any linkage criterion (single, average, or complete).…”

“…However, there were 2 main differences. First, the agglomerative clustering algorithm in the study by Wartelle et al [ 23 ] needs to access the original data after each merge to recalculate the RR values, which is very time consuming with large data. We constructed the network only once, without any need to access the original data after that.…”

Multimorbidity clustering of the emergency department patient flow: Impact analysis of new unscheduled care clinics

“…To the best of our knowledge and based on the review of Padros-Torres et al focused on complex patient and chronic diseases [ 21 – 25 ], this is the first time that an analysis of multimorbidity patterns has been used to cluster healthcare visits using a temporal-based analysis within a specific population. With this novel method [ 34 ], we extended the use of these methods to a general ED population with a low multimorbidity context and with a comprehensive clustering system that was able to classify the vast majority of ED visits (92.93%). Other classifications based on the acuity or urgency of each patient such as the CIMU scale, CCMU scale or PS classification could have directly been used, especially since UCS targets low-acuity patients that should not require hospitalization or vital care that was shown with patients of CIMU levels 4 and 5.…”

“…A detailed study of the clustering method, concepts and results has been published elsewhere and is summarized below [ 34 ]. Using the co-occurrences of ICD10 diagnoses with block level representation, we applied a new design of Hierarchical Agglomerative Clustering (HAC) [ 35 ] with a new measure of similarity which targets relative risk: where p ij designates the probability of diagnoses i and j co-occurring in sets of patients, ‘visits less than 6 months focus on more meaningful co-occurrences links’, whereas, p i and p j are the marginal probabilities of occurrences used to weight the relation [ 36 ].…”