Inflammatory protein profiling shows limited ability to distinguish sepsis from sepsis mimics

Abstract

Sepsis mimics resemble sepsis clinically but may require different treatments. We evaluated whether inflammatory protein profiles could distinguish these conditions in a retrospective case-control study using 4 prospective Japanese cohorts. The study included 34 adults with sepsis and 46 with sepsis mimics. We measured 92 proteins using Olink, assessed between-group differences with false discovery rate (FDR) adjustment, and evaluated 3 prediction models using repeated nested cross-validation. No protein met the FDR threshold; the lowest FDR-adjusted P value was 0.345. The primary elastic-net model had an area under the receiver operating characteristic curve of 0.442 (95% confidence interval, 0.315-0.574), based on cross-validation predictions averaged for each participant. Random forest and gradient boosting also showed no clear evidence of discrimination above chance. Confidence intervals included chance performance in both the all-serum and quality-control-Pass sensitivity analyses. In this cohort, a single measurement of an inflammation-focused protein panel showed limited discrimination between sepsis and sepsis mimics. Integration with clinical information and longitudinal measurements warrants further investigation.