We prospectively examined 464 febrile patients (median age, 61 years) for predictors of in-hospital death, by use of univariate and multivariate logistic regression using clinical data (age, underlying disease, duration of fever, chills, and shock on admission) and plasma endotoxin, TNF-alpha, IL-6, IL-10, and procalcitonin levels. The mortality rate was 4.6-fold higher (95% confidence interval [CI], 1.8-12) in 31 patients with shock on admission, 7 of whom died; the strongest association with mortality was the endotoxin concentration (relative risk, 13.7; 95% CI, 1. 4-136), which predicted 5 of the deaths with a 5% false-positive rate. For 433 patients without shock on admission, mortality (26 deaths) was associated with age and underlying disease: clinical data predicted 30% of the deaths, whereas IL-6 and procalcitonin levels identified an extra 10% with a 5% false-positive rate. When febrile patients are screened on hospital admission to identify those with a high risk for mortality, clinical judgment on the basis of age, underlying disease, and recent history outweighs the predictive value of endotoxin, cytokine, and procalcitonin levels. Only in patients who present with shock will measurement of endotoxin levels help predict those who will likely die at the cost of few false-positive results.