Unsupervised Clustering of 41,728 Emergency Department Visits: Insights into Patient Profiles and KTAS Reliability

Area of Science:

• Emergency Medicine
• Data Science
• Health Services Research

Background:

• Emergency department patient classification is crucial for timely care.
• Existing systems like the Korean Emergency Patient Classification Tool (KTAS) may not capture full patient heterogeneity.
• This can lead to under-classification and suboptimal resource allocation.

Purpose of the Study:

• To identify distinct patient subgroups in the emergency room using unsupervised learning.
• To explore patient heterogeneity beyond the current KTAS classification.
• To inform potential improvements in emergency triage and resource management.

Main Methods:

• Retrospective cross-sectional study of 41,728 emergency room patients.
• K-prototypes unsupervised cluster analysis using demographic, physiological, and clinical data.
• Validation of cluster numbers using Silhouette, Dunn, and Davies-Bouldin indicators; dimension reduction via UMAP.

Main Results:

• Two distinct patient clusters were identified.
• Cluster 0: Stable patients (mean age 58). Cluster 1: Younger, physiologically unstable patients (mean age 46, lower average arterial pressure).
• Significant differences in vital signs and pain scores between clusters; moderate association with KTAS classification (Cramer's V = 0.208).

Conclusions:

• Unsupervised learning can identify patient heterogeneity not captured by KTAS.
• This approach aids in early identification of high-risk patients and efficient resource allocation.
• Findings support developing a supplementary system for precision triage and patient-centered emergency care policies.