Background: Fever and neutropenia are common clinical problems in pediatric oncology and frequently necessitate emergency hospitalization and immediate empiric broad spectrum antimicrobial therapy. Estimating the risk of bacteremia in fever and neutropenia is a challenge. The purpose of this study was to develop an algorithm predicting the risk of bacteremia and Gram-negative bacteremia in children and adolescents with fever and neutropenia, based on information accessible at presentation.
Methods: We collected information available within 2 h of presentation of children with fever and neutropenia and, on outcome, from all pediatric cancer patients presenting with fever and neutropenia from 1993 through 2001 in a retrospective single center cohort study. After univariate analyses a multivariate decision tree was constructed, and its performance was evaluated by cross-validation.
Results: Bacteremia was detected in 87 (24%) and Gram-negative bacteremia in 30 (8%) of 364 episodes of fever and neutropenia. At the predetermined sensitivity level, > or =95%, decision tree models reached cross-validated specificities of 37 and 43%, with negative predictive values of 96 and 99%, for bacteremia and Gram-negative bacteremia, respectively. Absence of a clinically or radiologically evident source of infection and previous episodes of fever and neutropenia were defined as two newly described factors associated with bacteremia.
Conclusions: Based on this retrospective analysis, it appears that bacteremia can be predicted with clinically useful specificity at a high level of sensitivity, using clinical information available at presentation in pediatric cancer patients with fever and neutropenia.