Consensus-derived clinical sepsis profiles and mortality across six international cohorts
Ying-ying Shen, Ru-qin Yi, Wei Chen, Cheng-en Li, Ding Zhang, Sheng-hui Miao, Zhi-rong Zhang, Zhao-kun FanBackground:
Clinical phenotyping may help characterize sepsis heterogeneity, but the applicability of derived profiles across healthcare settings remains uncertain. We evaluated consensus-derived clinical profiles and their associations with mortality across six international cohorts.
Methods:
We analyzed 62,458 adults from six critical care databases in four countries. In the MIMIC-IV discovery cohort (n = 28,733), six clustering algorithms were applied to 24 routinely available clinical variables to derive consensus profiles. The profiles were assigned in five external cohorts using fixed discovery-cohort templates. We examined clinical characteristics, mortality associations, and exploratory treatment–outcome associations.
Results:
Three profiles were selected based on the Calinski–Harabasz peak and clinical interpretability. Profiles 1, 2, and 3 comprised 32.6%, 20.0%, and 47.4% of the discovery cohort, with in-hospital mortality of 30.3%, 13.8%, and 9.3%, respectively. Profile 1 showed greater multiorgan dysfunction, Profile 2 relatively preserved hemodynamics, and Profile 3 greater ventilatory support with less extrapulmonary dysfunction. Profile 1 had the highest mortality in all five external cohorts, with a decreasing mortality gradient across profiles in most cohorts. After adjustment for age and SOFA score, Profile 3 was associated with lower 1-year mortality than Profile 1 (hazard ratio 0.62, 95% CI 0.57–0.67). Treatment–outcome associations were exploratory.
Conclusions:
Consensus-derived profiles characterized clinically interpretable patterns of sepsis heterogeneity with mortality differences across the evaluated cohorts. These findings support further evaluation of clinical profiling for risk stratification and phenotype-informed research.