Phenotyping Sepsis at Emergency Department Presentation Using Early Clinical Data: A Multicenter Cohort Study

Sepsis is heterogeneous, and early emergency department (ED) management is constrained by limited information and dynamic clinical trajectories.

Using the nationwide Korean Sepsis Alliance (KSA) registry, we applied k-means clustering to nine routine ED variables (demographics, vital signs, Clinical Frailty Scale, mental status, lactate). The model was developed in KSA 3 (2019-2021) and tested for temporal reproducibility in the independent KSA 5 cohort (2022-2023). Phenotypes were compared on 7-day mortality, early sepsis-bundle compliance, and SOFA trajectories; stability was assessed by bootstrap resampling and alternative clustering. One-hour vasopressor non-compliance was decomposed into criterion-not-met versus delivered-with-delay.

Six phenotypes were identified in 9,430 patients and reproduced in 7,184 (mortality-rank ρ = 1.00; calibration drift, O/E 0.83-1.32). Seven-day mortality ranged 6.7%-31.8%. Extreme-severity phenotypes (A, F, E) were bootstrap-robust (Jaccard 0.65-0.91), whereas intermediate phenotypes (B, D) were less stable (0.21-0.47), consistent with a continuous severity spectrum. Non-compliance composition separated phenotypes that severity scores rank together: predominantly criterion-not-met (appropriate withholding) in low-acuity phenotype C versus delivered-with-delay in high-acuity phenotype E. Phenotypes B and F, matched on presenting physiology and mortality but differing in frailty and mental status, still differed in composition (criterion-not-met 49.6% vs 28.3%; p < 0.001). SOFA trajectories were highest in phenotype E.

ED-presentation phenotyping using routine variables identifies six cross-sectional signatures along the sepsis severity spectrum rather than definitive subtypes. Decomposing bundle non-compliance shows that it combines appropriate withholding and urgency-associated delay in phenotype-dependent proportions, a distinction one-dimensional severity scores cannot make, relevant to phenotype-stratified resuscitation warranting prospective evaluation.