Rapid classification of positive blood cultures. Prospective validation of a multivariate algorithm

Abstract

Objective: To develop and validate a model predicting whether a positive blood culture represents a true positive or a contaminant in hospitalized patients, using only information available when the initial culture result becomes available.

Design: Prospective cohort study with derivation and validation sets.

Setting: Urban tertiary care hospital.

Patients: Clinical data were collected within 24 hours of the initial culture from a random sample of inpatients who had blood cultures performed, and data from the episodes in which growth was reported were included. There were 219 episodes in the derivation set and 129 episodes in the validation set.

Main outcome measure: True bacteremia. Reviewers blinded to potential clinical predictors and initial laboratory results classified 115 (53%) of the episodes in the derivation set and 57 (44%) of the episodes in the validation set as true positives.

Results: Independent multivariate predictors of bacteremia were organism type, days until the blood culture became positive, multiple positive cultures, and clinical risk score. These factors were used to develop a model stratifying patients into four risk groups. In the derivation set's lowest-risk group, 92% (65/71) of positives represented contaminants, and in the highest-risk group, 97% (86/89) of positives represented true positives. In the validation set, the misclassification rates were 14% (8/59) in the low-risk group, and 11% (5/44) in the high-risk group. These two groups together comprised 76% of all episodes.

Conclusion: This model can help clinicians quantify the likelihood that a given positive blood culture represents a true positive when the laboratory first calls, which may be helpful in subsequent decision making.