Abstract / Summary
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.