Inconsistency of a model aimed at predicting bacteremia in hospitalized patients

Benjamin Mozes*, David Milatiner, Colin Block, Zvia Blumstein, Hillel Halkin

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review


Clinical prediction rules can help physicians determine the necessity for blood cultures in specific patients and/or in whom emp ric antibiotic treatment should be administered. Before adopting a prediction rule its validity must be evaluated in different settings. We revealed independent predictors of true bacteremia and developed a risk score based on them in one group of adult hospitalized patients (n = 474; derivation set). An attempt was made to validate this risk score in a second group of in-patients at the same hospital (n = 438; validation set). The derivation set included 540 blood culture episodes and the validation set 516. A blood culture episode was defined as one or more of all blood specimens withdrawn for culture from one patient over one 24 hour period. Independent multivariate predictors of true bacteremia' were: temperature of 39°C or higher, current immunosuppressive therapy, serum alkaline phosphatase > 100 IU and hospitalization in an intensive care unit. In the low risk group, defined by the absence of the said predictors, the rates of true bacteremia were 5.1 and 4.6% for the derivation and validation sets, respectively. As raised temperature is the main clinical feature guiding physicians to suspect bacteremia, we examined the probability of true bacteremia in patients with a temperature of less than 38°C and found it to be 5.6% in the two sets. The model identified high risk subset patient groups demonstrating true bacteremia in 38% of all episodes in the derivation set and the comparatively low rate of 12.1% (p < 0.01) for the validation set.' Application of a predictive model developed in our hospital was only partially successful in categorizing patients by risk of bacteremia. The risk in both sets, even among the patients at lowest risk, was judged too high to exclude collecting blood specimens for culture. These findings demonstrate the inconsistency of a model aimed at predicting bacteremia even within the same institution over different time periods.

Original languageEnglish
Pages (from-to)1035-1040
Number of pages6
JournalJournal of Clinical Epidemiology
Issue number9
StatePublished - Sep 1993
Externally publishedYes


  • Bacteremia
  • Clinical model validation
  • Clinical prediction rule


Dive into the research topics of 'Inconsistency of a model aimed at predicting bacteremia in hospitalized patients'. Together they form a unique fingerprint.

Cite this