Machine Learning-Based Clusters of Vital Signs and Lactate Levels Predict Vasopressor Use in Sepsis

Jeong, Daun; Choi, Minyoung; Maeng, Seung Jin; Yoon, Hanbeom; Park, Jong Eun; Lee, Gun Tak; Hwang, Sung Yeon; Shin, Tae Gun; Chung, Sung Phil; Lim, Tae Ho; ,

doi:.0..

Original Article

doi: https://doi.org/10.15441/ceem.25.247 [Accepted]

Machine Learning-Based Clusters of Vital Signs and Lactate Levels Predict Vasopressor Use in Sepsis

Daun Jeong^1,2, Minyoung Choi³, Seung Jin Maeng³, Hanbeom Yoon³, Jong Eun Park³, Gun Tak Lee³, Sung Yeon Hwang³, Tae Gun Shin³, Sung Phil Chung⁴, Tae Ho Lim⁵; on behalf of the Korean Shock Society

¹Division of Critical Care Medicine, Department of Emergency Medicine, Chung-Ang University Gwangmyeong Hospital, Gwangmyeong-si, Gyeonggi-do, Republic of Korea
²Department of Emergency Medicine, Chung-Ang University College of Medicine, Seoul, Republic of Korea
³Department of Emergency Medicine, Samsung Medical Centre, Sungkyunkwan University School of Medicine, Seoul, Korea
⁴Department of Emergency Medicine, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Korea
⁵Department of Emergency Medicine, College of Medicine, Hanyang University, Seoul, Korea

Correspondence Tae Gun Shin Tel: +82-2-3410-2053, Fax: +82-2-3410-0012, Email: drshin88@gmail.com

Received: October 15, 2025. Revised: December 2, 2025. Accepted: December 29, 2025. Published online: January 14, 2026.
Daun Jeong, Minyoung Choi and Seung Jin Maeng contributed equally to this work.

ABSTRACT

Objective
Sepsis remains a major clinical challenge because of its complex, heterogeneous, and multidimensional clustering patterns. This study aimed to investigate the association between vasopressor administration and machine learning–derived clusters based on initial vital signs and lactate measurements obtained in emergency department (ED) and intensive care unit (ICU) settings.
Methods
A retrospective cohort analysis was performed using data from the Korean Shock Society Septic Shock (KOSS) Registry (septic shock in the ED) and the Marketplace for Medical Information in Intensive Care (MIMIC)-IV database (ICU patients with suspected infection). To derive clusters, k-means clustering was applied to six initial vital signs and serum lactate measurements. The primary outcome was vasopressor administration. Secondary outcomes included second vasopressor administration and 28-day mortality.
Results
A total of 17,500 patients were included in the analysis (KOSS cohort, n=7,130; MIMIC-IV cohort, n=10,370). K-means clustering identified three distinct clusters in each cohort. In the KOSS cohort, Cluster 3 was characterized by the lowest mean arterial pressure (MAP) (62 mmHg [IQR, 53–71]) and the highest diastolic shock index (DSI) (2.6 [2.3–3.0]). This cluster was associated with the highest rates of vasopressor administration (93.9%), second vasopressor administration (33.5%), and 28-day mortality (25.3%) (all p<0.001). Comparable physiological and clinical patterns were observed in the MIMIC-IV cohort, in which Cluster 3 likewise demonstrated the lowest MAP (68 mmHg [60–76]) and highest DSI (2.0 [1.8–2.3]). This group similarly exhibited the poorest outcomes, including vasopressor administration (41.0%), second vasopressor administration (16.7%), and 28-day mortality (29.0%).
Conclusion
Machine learning–derived clusters based on initial vital signs and serum lactate levels demonstrated different patterns of vasopressor use and mortality. The clinical utility of this approach for guiding timely or targeted vasopressor therapy requires prospective validation.

Keywords: Sepsis; Septic shock; Emergency department; Intensive care units; Cluster analysis