 |
|
 |
|
 |
 |
|
 |
 |
 |
 |
 |
 |
 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
PEDIATRICS Vol. 106 No. 5 November 2000, pp. 977-982
From the * Division of Pediatric Emergency Medicine, Department
of Pediatrics, Eastern Virginia Medical School, Children's Hospital of
The King's Daughters, Norfolk, Virginia; and the Division of
Pediatric Emergency Medicine, Harvard Medical School, Children's
Hospital of Boston, Boston Massachusetts.
 |
ABSTRACT |
|---|
Top
Abstract
Methods
Results
Discussion
Conclusion
References
|
|---|
Purpose. To develop an improved model for the prediction of bacteremia in young febrile children.
Methods. A retrospective review was performed on patients 3 to 36 months of age seen in a children's hospital emergency department between December 1995 and September 1997 who had a complete blood count and blood culture ordered as part of their regular care. Exclusion criteria included current use of antibiotics or any immunodeficient state. Clinical and laboratory parameters reviewed included age, gender, race, weight, temperature, presence of focal bacterial infection, white blood cell count (WBC), polymorphonuclear cell count (PMN), band count, and absolute neutrophil count (ANC). Logistic regression analyses were used to identify factors associated with bacteremia, defined as growth of a pathogen in a blood culture. The model that was developed was then validated on a second dataset consisting of febrile patients 3 to 36 months of age collected from a second children's hospital (validation set).
Results. There were 633 patients in the derivation set (46 bacteremic) and 9465 patients in the validation set (149 bacteremic). The mean age of patients in the derivation and validation sets were 15.8 months (95% confidence interval [CI]: 15.2-16.5) and 16.6 months (95% CI: 16.5-16.8), respectively; the mean temperatures were 39.1°C (95% CI: 39.0-39.2) and 39.8°C (95% CI: 39.7-39.8); 56% were male in the derivation set and 55% male in the validation set. Predictors of bacteremia identified by logistic regression included ANC, WBC, PMN, temperature, and gender. Receiver operator characteristic (ROC) analysis showed similar performance of ANC and WBC as predictors of bacteremia. When placed into a multivariate logistic regression model, band count was not significantly associated with bacteremia. Information regarding focal infection was available for 572 patients in the derivation set. The percentage of patients diagnosed with bacteremia with a focal bacterial infection was not significantly different from the percentage who had bacteremia without a focal bacterial infection (16/200 vs 30/372). Based on this dataset, a logistic regression formula was developed that could be used to develop a unique risk value for each patient based on temperature, gender, and ANC. When the final model was applied to the validation set, the area under the ROC curve (AUC) constructed from these data indicated that the model retained good predictive value (AUC for the derivation vs validation data = .8348 vs 0.8221, respectively).
Conclusions. Use of the formulas derived here allows the clinician to estimate a child's risk for bacteremia based on temperature, ANC, and gender. This approach offers a useful alternative to predictions based on fever and WBC alone.bacteremia, detection, white blood cell.
Bacteremia occurs in 3% to 11% of febrile children 3 to
36 months of age.1-4 Because a percentage of children
with bacteremia will go on to develop serious bacterial infection,
clinicians have sought to identify clinical and laboratory variables
that aid in its prediction. Although information from the history and
physical examination are helpful, they do not provide a highly
sensitive screen for the detection of bacteremia in
children.1-3 Clinical predictors studied to date include
the age and appearance of the child and the height of their
fever.5-8 Laboratory predictors identified date include
the white blood cell count (WBC), C-reactive protein, erythrocyte
sedimentation rate, and other acute inflammatory
mediators.9-19
Debate currently exists as to the optimal management strategy for the
young child presenting with fever and no source. Practice guidelines
that were developed to aid clinicians in the management of the febrile
child have suggested the use of WBC as a discriminator between febrile
patients who could be observed without prophylactic antibiotic therapy
and those who merit treatment.20
The WBC differential is a routine component of the complete
blood count (CBC). While 2 recent investigations have commented on the
utility of the absolute neutrophil count (ANC),21,22
information derived from the WBC differential has not been incorporated into practice guideline recommendations. Furthermore, no studies have
attempted to develop a prediction tool that incorporates information from all predictors (such as age, temperature, WBC, and
ANC) to establish a unique risk value for each patient. Because the
ascertainment of whether a febrile pediatric patient has bacteremia remains an educated guess even in the best of hands, it makes sense to
use all information that might aid in this prediction.
A retrospective review was conducted of all children 3 to 36 months of age presenting to the Children's Hospital of The King's Daughters emergency department (ED) between December 1995 and December
1996 who had a CBC and a blood culture performed as part of their
evaluation for fever. The decision to order a CBC and culture on an
individual patient was made by the ED attending physician based on
clinical suspicion and standard patient-specific risk factors. Patients
were identified by generating a monthly query from the hospital's
computerized microbiology log of all blood cultures obtained from the
ED. During the second year of patient accrual (December 1996 through
August 1997), the only cases added to the database were those of
children who had both a blood culture and a CBC obtained and grew a
pathogen from their blood culture. Two control patients (eligible
patients presenting during the same period, who also had CBCs and blood
cultures drawn that were negative) per bacteremic patient were also
randomly selected. Patients excluded from the study included those
patients on antibiotic therapy within 48 hours of presentation, or
those with a known immunodeficiency state such as patients with sickle cell anemia, Down syndrome, immunoglobulin deficiency, or those currently on chronic steroid therapy. Records were reviewed on all
eligible patients. Data abstracted on each patient included: age in
months, presenting temperature in degrees centigrade, weight in
kilograms, the presence or absence of a focal bacterial infection (including otitis, pharyngitis, urinary tract infection, or bone and
joint infection), the type of differential count performed (manual vs
automated), and the CBC and blood culture results. Blood cultures
characterized as true pathogens included all cultures yielding
Streptococcus pneumoniae, Haemophilus influenzae,
Neisseria meningitidis, group A streptococcus, or
Salmonella species. Growth of Staphylococcus
epidermidis, Propionibacterium acnes, and diphtheroids from immunocompetent hosts with no history of cardiac disease, ventriculoperitoneal shunts, indwelling catheters, or other prosthetic devices were categorized as contaminants. The remaining organisms were
categorized on a case-by-case basis, factoring in the organism recovered, the clinical course of the patient, and the clinical diagnosis. ANC was computed as the total WBC × 103 cells/mm3 multiplied by
the sum of the percentage of bands and polymorphonuclear cell count
(PMN) divided by 100. In cases in which an automated differential was
computed, ANC was calculated as the WBC multiplied by the percentage of
granulocytes divided by 100.
Statistical Methodology
Analyses were conducted to assess cut-points of individual risk
factors and to develop an approach for multivariate assessment of risk
for bacteremia. To develop a multivariate model for risk, univariate
analyses were first conducted to identify potential risk factors for
bacteremia. The 2-sample Wilcoxon rank-sum test was used to compare the
median values of continuous variables between subjects with and without
bacteremia. The TABLE 1 TABLE 2 TABLE 3 TABLE 4
METHODS
Top
Abstract
Methods
Results
Discussion
Conclusion
References
2 test was used to assess the
association between gender and an outcome of bacteremia (Table
1). Logistic regression was then used to
develop a multivariate model for risk based on variables with
P < .05 in the univariate analyses. Final models include variables significantly associated with an outcome of bacteremia (Table 2). The presence of
interaction between variables was examined by inclusion of interaction
terms. The predictive power of models 1 and 2 and models based on
individual predictors was compared by constructing receiver operator
characteristic (ROC) curves and comparing the areas under these curves
(Fig 1). The final multivariate logistic
regression model based on ANC, temperature, and gender was used to
develop a formula for individual level risk that uses information from
all significant predictors of bacteremia (Table
3). This model was then applied to a
validation dataset derived from a large single-center study aimed at
determining the risk of bacteremia in children 3 to 36 months of age in
the post-H influenzae type b era.22 Data fields
present in both datasets included: age, WBC, PMN, band count,
temperature, gender, race, and blood culture results. The predictive
value of the model was assessed by applying it to the validation set
and computing the AUC for the ROC curve for the validation data. The
sensitivity and specificity of a number of risk values were also
tabulated and compared between both datasets. An assessment of
cut-points of clinical risk factors for bacteremia was then undertaken.
The best cut-point for each variable was defined as the value that
simultaneously maximizes sensitivity and specificity in a logistic
regression model that includes response as the outcome and a covariate
that takes value 1 if the variable exceeds the cutoff value and takes
value zero otherwise (Table 4). The crude
and adjusted odds ratios for bacteremia for those with high versus low
values were then computed and compared for each of the variables;
adjustment for confounding was performed in a logistic regression model
that included other significant predictors for bacteremia.
Results of Univariate Analyses in the Derivation
Sample*
Logistic Regression
Results*
View larger version (23K):
[in a new window]
Fig. 1.
ROC curves for each significant predictor of bacteremia and the final
model based on temperature, gender, and ANC are displayed. A poor test
will lead to a line with a slope of 45 degrees, as each increase in
sensitivity comes at a cost of an equivalent increase in the
false-positive rate. A good test will yield high sensitivity and low
false-positive rates and will correspond to a curve that lies left of,
and above, a worse test.
Experimental Sample (Validation Sample)
Best
Cut-Points
| |
RESULTS |
|---|
|
Top
Abstract
Methods
Results
Discussion
Conclusion
References
|
|---|
Six hundred thirty-three patients met study criteria. Fifty-six percent of patients in the derivation set were male and 63% were black. The mean age of the study population was 15.8 months (95% confidence interval [CI]: 15.2-16.5) and the mean presenting temperature was 39.1°C (95% CI: 39.0-39.2). Comparative data for the validation set showed: 55% of patients were male; mean age was 16.6 months (95% CI: 16.5-16.8); mean temperature was 39.8°C (95% CI: 39.7-39.8). The mean presenting temperature in the validation set was significantly higher than that in the derivation set (P < .001).
Forty-six cultures in the derivation set yielded a pathogen. S pneumoniae was the predominant organism identified (39 cases). The other organisms identified included N meningitidis (1), Streptococcus pyogenes (1), Salmonella species (1), Escherichia coli (1), Acinetobacter baumannii (1), and Staphylococcus aureus (2). There was no difference in pathogen recovery rate from patients who had a focal infection identified on the initial examination (16/200; 8%) versus those who had no focus identified (30/372; 8%). Contaminants identified included S epidermidis (15), Alpha streptococcus (6), diphtheroids (2), S aureus (1), Staphylococcus haemolyticus (1), A baumannii (1), Micrococcus (1), Bacillus species (1), and mixed contaminants (1).
Univariate analyses identified temperature, band count, PMN, ANC, WBC,
and gender as being significantly associated with an outcome of
bacteremia (Table 1). Female gender was also noted to be
significantly associated with the outcome of bacteremia (30/46 females
with bacteremia vs 247/586 females without bacteremia;
P = .002 by 2 testing).
Multiple logistic regression identified gender, WBC, ANC, and PMN count
as significant predictors of bacteremia. Because ANC values are
dependent on both WBC and PMN, 2 logistic regression models were
fit
one that included temperature, gender, WBC, band count, and PMN
and one that included temperature, gender, band count, and ANC. All
variables except band count and temperature remained significant
predictors of bacteremia in the final multivariate models (Table 2).
Because temperature was identified as a significant predictor in
earlier studies,1-3,21,22 it was retained as a covariate
in the final models. To compare the predictive power of the final
models, ROC curves were constructed and the areas under the curves were
compared. ROC curves were also constructed for each individual
predictor (Fig 1). The final logistic regression formula for a
subject's risk of bacteremia based on his or her values of
temperature, gender and ANC is given by:
|
12.454.
This model can be used to calculate an individual risk factor for each patient based on the above variables ("Appendix"). The sensitivity and specificity of a variety of treatment thresholds are displayed (Table 3). When the multivariate model based on temperature, gender, and ANC (model 1 of Table 2) was applied to the validation set, we found it to demonstrate good predictive power (area under the ROC curve: .8221). The area under the curve in both the derivation and validation set were similar (.8348 vs .8221, respectively). A variety of laboratory cut-points were found to show similar sensitivity and specificity values when applied to both datasets (Table 4).
The best cut-points for individual level clinical risk factors were then identified as those values that simultaneously maximize the sensitivity and specificity for our data. The sensitivity, specificity, and unadjusted and adjusted odds ratios for the best cutoff values for each risk factor are displayed (Table 4).
| |
DISCUSSION |
|---|
|
Top
Abstract
Methods
Results
Discussion
Conclusion
References
|
|---|
Our data suggest that the WBC differential provides useful
information to the clinician in assessing the chances that a given febrile child may have bacteremia. As current practice guidelines suggesting empiric treatment for all children 3 to 36 months of age
with temperatures of 
39°C and WBCs 15 000
cells/mm3 lead to the unnecessary treatment in
~85% of instances, it is important to search for screening tools
that may yield improved diagnostic accuracy for the detection of
bacteremia. Although the risk values derived from our equation are not
highly sensitive or specific, these values offer a useful alternative
to a screen based on the height of the fever and the total WBC.
Our data agree with 2 recent studies investigating similar study
questions. Kuppermann et al21 analyzed the utility of an
automated WBC differential for the detection of pneumococcal bacteremia
in children. His data, based on 164 cases of occult pneumococal
bacteremia occurring in 6579 patients, suggested that an ANC value of
10 000 cells/mm3 was a better discriminator of
bacteremia than a WBC of 15 000 cells/mm3.
From a large dataset derived from a computerized hospital database, Lee
and Harper22 assessed the risk of pneumococcal bacteremia
in febrile young children in the post-H influenzae type b
era. They concluded that adopting a cut-point of a WBC of 18 000
cells/mm3 might lead to less
unnecessary utilization of antibiotics without sacrificing
significant sensitivity for the detection of disease. Both of
these studies looked only at the discriminatory power of the WBC and
ANC for detecting pneumococcal bacteremia and did not assess the
utility of this test for detecting occult bacteremia in general.
Examination of the table of cut-point values (Table 4) suggests that subjects are at increased risk for having bacteremia if the ANC value exceeds 9.46 or if the WBC value exceeds 14.3. Exceeding the best cut-point for PMN (59) or for temperature (39.6) is not as highly predictive of bacteremia as is exceeding the cut-point for ANC or WBC. The adjusted OR for being positive for bacteremia for those in the high versus low PMN group is 1.89 for PMN and is 1.50 or 1.19 for temperature (Table 4). This suggests that having a high ANC or WBC value may be the strongest indication that a subject will develop bacteremia.
Although the ROC curve (Fig 1) corresponding to the model based on temperature, gender, and ANC as covariates is not markedly different from that based on temperature, gender, and WBC, it is important to note that even small improvements in sensitivity and specificity of the model have huge medical and economic implications. For example, when the practice guidelines developed by Baraff et al20 are applied to our dataset, they yield a sensitivity of 54% and a specificity of 86% for the detection of bacteremia. If one uses our model's cutpoint that has equivalent sensitivity to the practice guidelines, the corresponding specificity would be raised to 90% (an absolute increase in specificity of 4%). Alternatively, if we adopted a cut-point to have the same specificity (86%) as that yielded by the practice guidelines, our sensitivity would be 69%, an absolute increase in sensitivity of 15%. Translated over the thousands of febrile children evaluated yearly, these improvements in sensitivity and specificity represent a significant advance.
Little has been written regarding the incidence of bacteremia in children with focal infections. In a cohort of 1666 young children retrospectively identified who had otitis media and had a blood culture obtained, Schutzman et al23 noted a similar incidence of bacteremia (3%) to those seen in the same ED with fever and no source of infection. Although our study design does not allow for a true approximation of incidence, our data agree with that of Schutzman et al23 in that patients with focal bacterial infections who had blood cultures obtained had the same incidence of bacteremia as those with no source. Both studies may suffer from a selection bias, which may have artificially inflated the incidence of bacteremia in these groups.
One of the interesting findings noted during this study was the limited usefulness of the band count as a predictor of bacteremia, relative to WBC, PMN, and ANC. Our data agree with that of Kuppermann et al21 and Wack et al24 who also noted that band count may be more of a nonspecific indicator of stress. Band count has been shown to be elevated in patients with meningococcemia, yet the predictive utility of this result is limited in a disease of such low incidence.25
We were surprised to find that females in our derivation set were at increased risk for bacteremia. This finding is contrary to prevailing opinion, which has noted that males have a slightly increased risk for occult bacteremia. Because our distribution of pathogens mirrors that of most studies, we do not believe that our gender findings are related to the increased prevalence of urinary tract infections or other focal infection in our population. This finding remains unexplained.
Although the ROC curve does demonstrate that the model based on ANC, temperature, and gender derived from this study is a better predictor of bacteremia than that currently used by the practice guidelines, practical application of this information requires the entry of the individual patient's temperature, gender, and ANC into a scientific calculator that would then compute an individual risk for each patient based on our final regression formula (Table 3). This risk value incorporates more of the clinical information available than is currently used and can be used by clinicians to estimate the patient's probability of having bacteremia and choose a treatment strategy. The fact that the model applied equally well to the validation set supports its generalizability. Additional studies could also examine the usefulness of this model in other datasets.
It is important to note that none of the models derived have the exquisite sensitivity and specificity that clinicians demand. The management of the febrile pediatric patient will remain a clinical situation that calls for an educated guess of a patient's risk for bacteremia based on the available data at hand. Certainly if the experimental efficacy of the conjugate pneumococcal vaccine leads to a marked reduction in pneumococcal bacteremia, the management of the febrile child may change dramatically. Regardless of the treatment strategy that is chosen, close follow-up of these patients is still a sine qua non for effective management of this clinical problem.
Limitations of the study are related to the retrospective nature of the data collected, the entry of patients based on the clinician's determination of the need for CBC and blood culture rather than discreet entry criteria, and the small number of patients with the outcome of interest. It is important to note that as patient accrual in the second year was limited to bacteremic patients (with 2 controls added per bacteremic) our data cannot be used to estimate true incidence of disease. The fact that the mean temperature in the derivation set (39.2°C) was significantly lower than that in the validation set (39.8°C) reflects differences in entry criteria for the 2 study populations. Although some patients in the derivation set were entered with temperatures that were lower than the common testing threshold of 39.0°C, we might expect that the clinicians caring for these children believed a unique clinical indication for the test (such as history of higher fevers before presentation, history of rigors, or general toxicity on examination). Entry criteria into the derivation set were designed to assess the utility of the CBC as a prediction tool for occult bacteremia in a population of patients with suspected occult bacteremia derived from a pediatric ED comprised of fellowship-trained pediatric emergency physicians.
| |
CONCLUSION |
|---|
|
Top
Abstract
Methods
Results
Discussion
Conclusion
References
|
|---|
In summary, ANC seems to be equivalent to WBC as a predictor of bacteremia in young febrile children. We suggest that utilization of the formula derived in this study may yield more accurate estimates of the risk of bacteremia than are those derived from current practice guidelines. Band count alone does not seem to be as good a predictor of bacteremia as was previously suggested.
| |
APPENDIX. |
|---|
Application of the Formula for Calculating an Individual Patient's Risk
Patient A is a male with an ANC value of 23.0 and temperature of 40.3°C. Using model 1 (Table 2), we calculate his risk for bacteremia as:
|
|
Patient B is a male with an ANC value of 7.8 and temperature of 41.0. As for patient A, we calculate his risk for bacteremia as:
|
|
| |
ACKNOWLEDGMENTS |
|---|
We thank Arno Zaritsky, MD, and Ardythe Morrow, PhD, for their careful review and advice regarding the manuscript; Kimberley Kelly and Cindy Buckler for help with manuscript preparation; and Justin Klaff for help with data abstraction.
| |
FOOTNOTES |
|---|
Received for publication Sep 29, 1998; accepted Jan 7, 2000.
This work was presented in part at the Southern Society for Pediatric Research; February 6, 1997; New Orleans, LA; and at the Pediatric Academic Societies Annual Meeting; May 5, 1997; Washington, DC.
Reprint requests to (D.J.I.) Children's Hospital of the King's Daughters, 601 Children's Ln, Norfolk, VA 23507. E-mail: disaacma{at}chkd.com
| |
ABBREVIATIONS |
|---|
WBC, white blood cell count; CBC, complete blood count; ANC, absolute neutrophil count; ED, emergency department; PMN, polymorphonuclear cell count; ROC, receiver operator characteristic; CI, confidence interval.
| |
REFERENCES |
|---|
|
Top
Abstract
Methods
Results
Discussion
Conclusion
References
|
|---|
, interleukin-1
, and interleukin-6 levels in febrile, young children with and without occult bacteremia.
Pediatrics
1999;
104:1321-1326 This article has been cited by other articles:
 |
|
 |
|
  D H Wyllie, I C J W Bowler, and T E A Peto Relation between lymphopenia and bacteraemia in UK adults with medical emergencies J. Clin. Pathol., September 1, 2004; 57(9): 950 - 955. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. R. Casey, S. M. Marsocci, M. L. Murphy, A. B. Francis, and M. E. Pichichero White Blood Cell Count Can Aid Judicious Antibiotic Prescribing in Acute Upper Respiratory Infections in Children Clinical Pediatrics, March 1, 2003; 42(2): 113 - 119. [Abstract] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
 |
||
|
||
Home | My Pediatrics | Journal Information | Current Issue | Past Issues | Subscriptions & Services | Contact Us Click here for faster international access © 2000 American Academy of Pediatrics. All rights reserved. |
||
|
||
