Rapid classification of positive blood cultures: validation and modification of a prediction model.

OBJECTIVE

1. To validate a previously developed prediction model to aid physicians in differentiating true positive blood cultures from contaminants when the laboratory first calls with a positive result, and 2) to determine whether it could be modified to make it more practical for clinical use without altering predictability.

DESIGN

A prospective cohort study of hospitalized patients (validation set) who had blood cultures done over a two-month period. Data collected included the seven independent predictors in the rapid classification of positive blood cultures model. The model was modified by eliminating one of the predictors (which required clinical data) but maintaining the laboratory components (morphologic and Gram stain characteristics, number of bottles positive, and time to positivity). The "blood culture episode" was the unit of evaluation. A blood culture episode was defined as a 48-hour period beginning with the drawing of blood for the culture and included any blood cultures obtained during that time period. Receiver operating characteristic (ROC) curve analysis was used to compare the predictabilities of these models.

SETTING

A 550-bed, university-affiliated county hospital that is a regional trauma center and has the only burn treatment unit in the region.

PATIENTS

All adult (> or = 16 years old) patients who had blood cultures done during the study period were eligible. Only patients with positive blood cultures were included in the study.

INTERVENTIONS

None.

MAIN RESULTS

Of 559 blood culture episodes identified, 139 (25%) included the growth of one or more organisms; 62 (45%) of the 139 episodes represented true bacteremia. By ROC curve analysis, there was no significant difference in the mean areas under the curve (AUCs) (+/- SE) of the model in the derivation set (the previously developed model) (0.93 +/- 0.02) compared with the validation set (0.89 +/- 0.03; p = 0.29). In the validation set there was no significant difference in the mean AUCs when the model was modified (0.89 +/- 0.03) by removing the clinical component vs the unmodified model (0.89 +/- 0.03; p = 0.98).

CONCLUSIONS

The rapid classification of blood cultures model was validated in a general hospital population. Predictability of the model was not altered significantly by eliminating one component that required clinical data. Because the modified model requires only laboratory information, this may allow reporting of the probability of true bacteremia at the time a positive blood culture is initially reported to physicians. This information may aid physicians in interpreting the positive blood culture.