Included in a 3-year population-based study were all patients (n = 64,281) admitted to a single tertiary care hospital (902 beds) using prospective hospital-wide surveillance for nosocomial infections. The objective of the study was to identify patients at risk for nosocomial bloodstream infection by using readily available hospital admission variables. After identifying potential risk factors for infection by univariate analyses, we derived multivariate models for predicting bloodstream infection by using logistical regression procedures. A total of 931 patients (1.45 per 100 admissions) developed a nosocomial bloodstream infection (2.2 episodes per 1000 patient-days) between 1 July 1987, and 30 June 1990. The crude mortality among infected patients was 34%, and the 319 deaths represented 22% of the total in-hospital mortality. Independent predictors of bloodstream infection were age, gender, primary diagnosis, and admission to a critical care unit. The sensitivity and specificity of the models for classifying patients as infected or noninfected were 81% and 81% for infants (1-11 months old) and 72% and 72% for adults, respectively. The negative predictive value of both models exceeded 99%. Applied to all patients on admission, the models we developed allowed us to survey only 28% of patients to identify more than 70% of those who will develop a nosocomial bloodstream infection.