| Authors: |
Teresa Bastos Lopes, Mariana Dias, Marta Lopes, Luís Silva, Ana Januário, Andreia Mesquita, Carla Menino, Carla Vidinha, Fausto Honoré Silva, Mafalda Gonçalves, Maria José Guimarães, Pedro Casimiro, Rui Malha, Ana Londral and Federico Guede-Fernandez |
| Abstract: |
Overcrowding in emergency departments (ED) is a global public health problem with a negative impact on the quality of care provided, service efficiency, and resource utilisation. A small subgroup of patients, referred to as high users (HU), comprise approximately 3.5% to 29% of ED users yet account for more than 60% of all ED visits, and are characterised by multimorbidity, complex care needs, and a high psychosocial burden.
At Garcia da Orta Hospital, in Portugal, this reality led to the implementation of the Grupo de Resolução dos High Users (GRHU), a multidisciplinary team dedicated to the case management of HU of the ED. This programme integrates individualised multidisciplinary consultations (MC), with the aim of reducing ED utilisation and improving health outcomes. However,GRHU faces challenges in selecting patients for intervention, particularly in identifying those for whom the intervention is likely to have the greatest impact.
This study aims to identify patient profiles based on their response to the GRHU intervention and to develop predictive classification models to support patient selection. Clinical and sociodemographic data from the hospital database over three years (2022–2025) were used, covering 8250 patients, of whom 256 underwent the GRHU intervention.
Patient profiles capturing the level of impact of the GRHU intervention were identified using unsupervised analysis methods, including K-means, Ward’s hierarchical clustering, and Gaussian mixture models: K-means for identifying compact utilisation patterns, Ward’s method for variance-based and hierarchically interpretable groupings, and Gaussian mixture models for modelling heterogeneous and overlapping patient response profiles. The analysis used clinically defined impact metrics, including ED visits, specialist consultations, hospitalisations, and MCs, measured over periods of 3, 6, and 12 months around the intervention.
The clustering approach that best identified distinct response profiles was Ward’s hierarchical method with three clusters. The model used a 6-month window and the following variables: number of ED visits after the first MC, reduction in ED visits following the first MC, time between the first and last MC, and number of MC. This analysis was applied to the 145 patients with complete data within the 6-month window. This solution was selected based on internal validation metrics (Silhouette=0.55; Davies–Bouldin=0.94; Calinski–Harabasz=93.25) complemented by external criteria of clinical interpretability, reflecting clear and consistent patterns in changes in ED utilisation. Three distinct profiles were identified: a profile characterised by a marked reduction in ED visits shortly after the first MC, potentially indicative of a strong response to the intervention (n=106); a second profile associated with a longer intervention duration, followed by a reduction in ED visits (n=24); and a third profile displaying more heterogeneous trajectories, with limited or no reduction and in some cases an increase in ED visits(n=15).
As future work, based on these profiles, supervised classification models will be developed to predict each patient’s response profile based exclusively on pre-intervention variables. This way, the team can predict the level of impact that the intervention will have on the patient, supporting clinical decision-making and contributing to a more efficient allocation of resources and optimisation of the GRHU intervention. |