Multicenter Validation of Clinical Sepsis Phenotypes

Importance: Four clinical phenotypes of sepsis based on data from electronic health records have been proposed. Although promising, the generalizability of these phenotypes remains uncertain, and multisite validation is needed. Objective: To validate, using the same methods and inclusion criteria, the 4 clinical phenotypes derived from Sepsis Endotyping in Emergency Care (SENECA) data. Design, Setting, and Participants: This multisite retrospective cohort study uses data on adult patients admitted to the emergency departments in Stockholm, Sweden (January 1, 2011, to September 1, 2023); Oxford, England (February 4, 2014, to June 20, 2021); and Oslo, Norway (January 4, 2019, to October 9, 2023) university hospitals. Included encounters are those with body fluid cultures taken, documented antibiotic administration, and Sequential Organ Failure Assessment scores of 2 or more, all within 6 hours of admission. Data analysis was conducted from November 2, 2023, to September 1, 2025. Main Outcomes and Measures: Consensus clustering with k-means was used to derive 4 clinical phenotypes at each site, comparing them with the SENECA-derived phenotypes, as well as with one another. Results: There were 30 865 patient encounters in Stockholm (mean SD age, 68 16 years; 18 165 men 59%), 15 575 in Oxford (mean SD age, 71 18 years; 9067 men 58%), and 1806 in Oslo (mean SD age, 71 17 years; 1068 men 59%). There was little consistency between the SENECA clinical phenotypes and each site's own phenotypes, with a Cohen κ of 0.32 for Stockholm, 0.37 for Oslo, and 0.40 for Oxford; the Adjusted Rand Indices were 0.21 for Stockholm, 0.27 for Oslo, and 0.26 for Oxford. There was also little consistency between the phenotypes derived in Stockholm, Oxford, and Oslo. Conclusions and Relevance: This study suggests that the 4 clinical phenotypes of the SENECA data are not generalizable across 3 independent cohorts. This calls for further exploration of possible underlying sepsis subgroups using alternative approaches that mitigate the inherent stochasticity in many unsupervised and semisupervised clustering methods.