CLINICAL PRACTICE

A Decision Rule for Predicting Bacterial

Meningitis in Children with Cerebrospinal
Fluid Pleocytosis When Gram Stain Is
Negative or Unavailable
Bema K. Bonsu, MBChB, Henry W. Ortega, MD, Mario J. Marcon, PhD, Marvin B. Harper MD

Abstract
Objectives: Among children with cerebrospinal fluid (CSF) pleocytosis, the task of separating aseptic
from bacterial meningitis is hampered when the CSF Gram stain result is unavailable, delayed, or nega-
tive. In this study, the authors derive and validate a clinical decision rule for use in this setting.
Methods: This was a review of peripheral blood and CSF test results from 78 children (<19 years) pre-
senting to Children’s Hospital Columbus from 1998 to 2002. For those with a CSF leukocyte count of
>7 ⁄ lL, a rule was created for separating bacterial from viral meningitis that was based on routine labo-
ratory tests, but excluded Gram stain. The rule was validated in 158 subjects seen at the same site
(Columbus, 2002–2004) and in 871 subjects selected from a separate site (Boston, 1993–1999).
Results: One point each (maximum, 6 points) was assigned for leukocytes >597 ⁄ lL, neutrophils >74%,
glucose <38 mg ⁄ dL, and protein >97 mg ⁄ dL in CSF and for leukocytes >17,000 ⁄ mL and bands to neu-
trophils >11% in peripheral blood. Areas under receiver-operator-characteristic curves (AROCs) for the
resultant score were 0.98 for the derivation set and 0.90 and 0.97, respectively, for validation sets from
Columbus and Boston. Sensitivity and specificity pairs for the Boston data set were 100 and 44%,
respectively, at a score of 0 and 97 and 81% at a score of 1. Likelihood ratios (LRs) increased from 0 at a
score of 0 to 40 at a score of ‡4.
Conclusions: Among children with CSF pleocytosis, a prediction score based on common tests of CSF
and peripheral blood and intended for children with unavailable, negative, or delayed CSF Gram stain
results has value for diagnosing bacterial meningitis.
ACADEMIC EMERGENCY MEDICINE 2008; 15:437–444 ª 2008 by the Society for Academic Emergency
Medicine
Keywords: meningitis, bacteria, enterovirus, children, prediction score, cerebrospinal fluid

T
he decline in acute bacterial meningitis brought
about by the introduction in the 1990s of a conju-
gate vaccine for Haemophilus influenzae Type b
and more recently by the introduction of a heptavalent
From the Department of Pediatrics, Division of Emergency
Medicine (BKB, HWO), and the Department of Laboratory
Medicine (MJM), Columbus Children’s Hospital, Columbus,
OH; the Division of Emergency Medicine, Children’s Hospitals
and Clinics of Minnesota (HWO), Minneapolis, MN; and the
Divisions of Emergency Medicine and Infectious Diseases,
Children’s Hospital Boston (MBH), Boston, MA.
Received October 10, 2007; revisions received January 2 and
January 22, 2008; accepted January 30, 2008.
Presented at the annual meeting of the Pediatric Academic Soci-
eties, San Francisco, CA, May 2004.
Address for correspondence and reprints: Bema K. Bonsu,
MBChB, e-mail: bonsub@pediatrics.ohio-state.edu.
vaccine for Streptococcus pneumoniae means that most
children who present with clinical signs of meningitis
and cerebrospinal fluid (CSF) pleocytosis will be diag-
nosed with aseptic meningitis.1,2 Because aseptic menin-
gitis is often self-limited, the challenge to clinicians is
how to strengthen decisions to discharge to home while
identifying the small number of children with acute bac-
terial meningitis.
A positive CSF Gram stain result eases this challenge
greatly because it strongly supports (apart from other
screening laboratory tests) a diagnosis of bacterial men-
ingitis and so mandates empiric antibiotic therapy and
hospitalization. Unfortunately, the ability to make such
decisions is severely hampered when this test is nega-
tive or unavailable, as occurs at sites (e.g., community
hospitals) that may only offer Gram staining at specific
intervals of the day. A negative Gram stain result, at a
false-negative rate of 40% and a true-negative rate of

ª 2008 by the Society for Academic Emergency Medicine ISSN 1069-6563

doi: 10.1111/j.1553-2712.2008.00099.x PII ISSN 1069-6563583 437
438
100%, translates to a clinically trivial negative likeli-
hood ratio (LR) of 0.4, barely altering the odds of bacte-
rial meningitis.3–6 In this scenario, but equally when a
report of the CSF Gram stain is unavailable or delayed,
alternative tools are needed not only to break the diag-
nostic impasse but also to inform decision-making
regarding initial antibiotic treatment, which, even when
lagging a mere few hours, is linked to adverse neuro-
logic outcomes.7–11
A number of extant tools are reported for aiding the
task of diagnosing meningitis.3,12–19 Of these, only two
are validated by independent investigators to be accu-
rate for diagnosing bacterial meningitis in children
other than those from which they were derived.20 These
include the bacterial meningitis score (BMS) by Nigro-
vic et al.3,13 and a fractional polynomial model by Bonsu
and Harper.12 The BMS, a weighted tally of five predic-
tors, is well established and simple to calculate at the
bedside; however, it relies heavily on the CSF Gram
stain result and so is limited when this test is delayed
or unavailable. The model by Bonsu and Harper, too,
while operational with or without the spinal fluid Gram
stain result, is impractical at sites that lack computer-
ized decision support capability or ready access to
preprogrammable probability calculators because it
requires substantial mathematical computation.12,20
A decision rule that combines the strengths of these
tools, namely, independence from the CSF Gram stain
and ease of use at the bedside, is likely to support deci-
sion-making in a wider array of clinical settings. In this
study, we provide a report of one such rule. This new
rule, which complements rather than supplants existing
rules, is intended to support decision-making in chil-
dren evaluated for signs of meningitis in the emergency
department (ED) when the CSF Gram stain result is
negative, delayed, or unavailable.
METHODS
Study Design
Our study is a retrospective cohort analysis of labora-
tory test results obtained from previously healthy chil-
dren with ages ranging from 4 weeks to 18 years who
presented to the ED with signs of acute meningitis.
Data were obtained from patient information archives
at Columbus Children’s Hospital and Children’s Hospi-
tal Boston after approval by institutional review boards
at both hospitals. Eligible subjects at Columbus Chil-
dren’s Hospital were children diagnosed with acute
viral meningitis or acute bacterial meningitis from June
1998 to June 2004. Eligible subjects at Children’s Hospi-
tal Boston were children who underwent lumbar punc-
ture from June 1992 to July 1999 for signs of acute
community-acquired meningitis.
Study Setting and Population
Both institutions are academic children’s hospitals with
EDs that serve diverse socioeconomic populations. At
Columbus Children’s Hospital, there are approximately
90,000 ED patient visits per year, of which 53% are by
whites, 36% are by African Americans, and 2% are by
Hispanics. Other groups, including those declining to
comment on their race, constitute 9% of children seen
Bonsu et al. • PREDICTING BACTERIAL MENINGITIS IN CHILDREN
at the hospital. At Children’s Hospital Boston there are
approximately 50,000 ED patient visits per year, of
which 35% are by whites, 25% are by African Ameri-
cans, and 20% are by Hispanics. Other groups consti-
tute 20% of children treated at the hospital.
Study Protocol
We reviewed, at both sites, laboratory test results of
children evaluated for acute infectious meningitis. Data
were entered into a database and included the white
blood cell (WBC) count in peripheral blood and CSF;
glucose, total protein, bacterial, and viral culture results
of CSF specimens; polymerase chain reaction (PCR)
results of CSF and throat specimens (performed rou-
tinely from late spring to early fall at Columbus Chil-
dren’s Hospital); and viral culture results of throat
specimens. At each site, data were linked by the subject’s
medical record number, a unique ED encounter number,
and the date and time of encounter. We excluded chil-
dren who did not have CSF pleocytosis (CSF leukocyte
count <8 cells ⁄ lL) as well as those with blood-contami-
nated CSF (>10,000 erythrocytes ⁄ lL) 12,18,20 and, for
some analyses, missing results of selected tests.
Predictor Selection. To create the clinical score, we
relied on data obtained from Columbus Children’s
Hospital for the interval June 1998 to February 2002
(derivation set). Operationally, healthy subjects were
defined by excluding from all data sets children with
International Classification of Diseases (ICD)-9–coded
diagnoses of neurosurgical, noninfectious neurologic,
and medical disease suggesting chronic poor health
clinically recognizable at the initial encounter.
Bacterial meningitis was coded to be present if stan-
dard culture of the CSF specimen isolated a pathogen
known to be a typical cause of this infection in an
otherwise healthy child. A diagnosis of bacterial menin-
gitis superseded a diagnosis of viral meningitis, irre-
spective of viral study results (no such case occurred).
The following bacteria grown from culture of CSF were
classified as contaminants and were removed from the
dataset: Viridans streptococci, Streptococcus bovis,
Staphylococcus epidermidis, Staphylococcus aureus,
Enterobacter spp., Corynebacterium spp., Bacillus spp.,
Flavobaterium spp., and Enterococcus spp. For score
creation, children for whom CSF did not grow a bacte-
rial pathogen or for whom an enterovirus could not be
identified were excluded. Viral meningitis was coded to
be present if culture or PCR of CSF or throat specimens
identified an enterovirus in a child with greater than
seven WBCs ⁄ lL in CSF. We reasoned that the score
had a good chance of generalizing to other settings if it
was able to distinguish between bacterial and viral
meningitis, diseases that are close in clinicopathologic
spectrum and, therefore most difficult to separate clini-
cally.
Predictors chosen for the score were tests commonly
used for diagnosis. These included the WBC concentra-
tion and ratio of bands to neutrophils (a surrogate for a
‘‘left-shifted’’ WBC differential) in peripheral blood; and
the WBC concentration, total protein, total glucose, and
the percentage neutrophils in CSF. The CSF Gram stain
result was not considered because: 1) a positive test
ACAD EMERG MED • May 2008, Vol. 15, No. 5 • www.aemj.org
result denotes the presence of bacteria, which by itself
indicates bacterial meningitis, and 2) a negative Gram
stain result barely alters the prior odds of bacterial
meningitis, so from a practical standpoint presents a
dilemma similar to that observed when the CSF Gram
stain report is delayed or absent, i.e., creates the need
to rely on other tests for diagnosis and treatment.
Data Analysis
We identified optimal cut-points for each selected pre-
dictor based on recursive partitioning, a technique that
uses nonparametric bivariate analysis to partition sam-
ples into subsets with the lowest probability of false-
negative and false-positive assignments with respect to
a given outcome.21,22 Sample partitioning continues
recursively with the option of penalizing misdiagnosis
depending on the seriousness of misclassification or to
enhance the performance of tests in the set used to
derive the rule.
There is modest risk with recursive partitioning of
overfitting the observed data, leading to poor perfor-
mance in new data sets. It is also established that the
usefulness of particular tests for predicting bacterial
meningitis varies between patients. A hierarchical pro-
cedure like recursive partitioning, therefore, that orders
tests sequentially based on their relative performance
in a particular group of patients may translate poorly to
new patients among whom other tests predominate.
Specifically, tests positioned early in the hierarchy of a
tree may be demoted when evaluated in new data sets.
Even slightly altering observed data sets can sometimes
lead to major changes in tree structure.21,22
For these reasons, and because CSF cell counts and
chemistries are generally assigned equal or nearly equal
weight for diagnosis, we adopted a rule that utilizes
these tests in parallel rather than in sequence. To do
this, we evaluated each test in a separate recursive par-
titioning tree, and, for robustness, chose test thresholds
at the first branch. Overfitting was lowered by avoiding
penalization for misclassification. Test thresholds so
selected were aggregated into a prediction rule that
gave equal weight to each test for determining whether
a particular child had bacterial meningitis. One or zero
points were assigned to each test (base classifier)
depending on whether values for that test fell within or
outside the selected intervals. A simple aggregation of
base classifiers for a maximum of 6 points was expected
to enhance the accuracy of the score in future data sets
without increasing its complexity inordinately. Recur-
sive partitioning was conducted with the RPART pro-
gram of the R statistical analysis environment (http://
www.r-project.org).
Validation. To validate our score, we followed recom-
mendations proposed in recent reviews.23–25 We first
estimated the accuracy of the score in the data set from
which it was derived (internal validation). This step was
followed by external validation of the score in a data
set assembled at the same site, but from a different time
interval (temporal transportability), and in a data set
assembled from a different site and time interval (geo-
graphic transportability). By using slightly different
inclusion criteria for parts of our analysis, we also
439
tested the methodologic transportability of our score.
This scheme tested the robustness (statistical and clini-
cal invariance) of our score to these factors.
Internal Validation of Score. Internal validation of the
model was achieved by bootstrapping.21,22 The boot-
strap procedure generates new data sets of the same
size as the observed data set by resampling from it with
replacement. Diagnostic performance is adjusted for
overoptimism (bias) by averaging from generated sam-
ples. The procedure was repeated 5,000 times. The area
under the receiver-operating-characteristic curve
(AROC), a common index of test performance, was esti-
mated for the prediction score.26,27 The ROC curve
plots sensitivity versus 1-specificity (LR) for ordered val-
ues of a test providing an estimate of global diagnostic
capability. The higher the AROC, the more favorable
the tradeoff is between sensitivity and specificity.28 An
AROC of 0.5 indicates that a diagnosis model has no
discriminating ability, whereas an AROC value of 1.0
indicates perfect discriminating ability. We adopted a
nominal value of 0.7 to be the threshold for stating a
model discriminated between bacterial and viral menin-
gitis.29
We also calculated sensitivity, specificity, and interval
LRs at a number of thresholds of the score.30 Because
of their discrete nature, we considered each threshold
of our prediction score to be an interval when calculat-
ing LRs. Based on published recommendations and
cognizant of serious complications of missed bacterial
meningitis, we adopted the following convention for
interpreting LRs: LRs of <0.5 lower the odds of infec-
tion (markedly, if <0.1); LRs of >2 increase the odds of
infection (markedly, if >10); and LRs from 0.5 to 2 do
not alter the odds of infection.31
External Validation. Our score was validated in two
sets of children. The first was a previously assembled
group of patients with a leukocyte count in CSF of
>9 ⁄ lL seen at Columbus Children’s Hospital from
March 2002 to June 2004 (the small difference in
thresholds for CSF leukocytes was considered insignifi-
cant). Because this sample was small, we did not
exclude subjects who had incomplete reporting of the
laboratory tests. Instead, the maximum score attained
based on available test results was substituted for the
full score.
The second validation set was children with a leuko-
cyte count in CSF of >7 ⁄ lL seen at Children’s Hospital
Boston from 1993 to 1999. Criteria for diagnosing acute
bacterial meningitis at this site were unchanged from
those specified for children at Columbus Children’s
Hospital. Because this sample was larger (see Results),
we were able to restrict our analysis to subjects who
had full reporting of laboratory test results.
It should be noted that controls chosen for external
validation of the rule had wider clinical distribution
than those chosen for creating the rule. Specifically, for
external validation, controls included every eligible sub-
ject with CSF pleocytosis not assigned a discharge
diagnosis of bacterial meningitis and irrespective of
viral culture and PCR findings of CSF. By choosing
this expanded definition, we were able to validate the
440
accuracy and transportability of the prediction score in
a more general sample of children.
Our primary goal was to develop an alternative tool
with at least equal, but preferably greater sensitivity
over the CSF Gram stain result for corroborating a
negative Gram stain report or for excluding bacterial
meningitis when the CSF Gram stain report is unavail-
able or delayed. With this goal in mind, we compared
the sensitivity of our score to that quoted in the medical
literature for CSF Gram stain: 60%–80% for untreated
and 40%–60% for antibiotic-pretreated patients in a
recent review.32 We considered our score sufficiently
useful if for all cases studied, 95% confidence intervals
(CIs) for sensitivity lay beyond the range reported for
the CSF Gram stain result. This definition is likely to
have intuitive interpretation, being more pragmatic
than an arbitrarily selected threshold.
RESULTS
Study Subjects
From June 1998 to February 2002, we identified 19 chil-
dren with laboratory-established bacterial meningitis
and 59 with laboratory-established enteroviral meningi-
tis at Columbus Children’s Hospital who met all study
criteria and had all ancillary test results recorded, for a
total of 78 subjects. The median age of children with
bacterial meningitis was 1.0 year (interquartile range
[IQR] = 0.4–2.2 years) and for children with enteroviral
meningitis, it was 6.6 years (IQR = 0.2–9.9 years;
p = 0.0001). Bacteria isolated from the CSF of children
evaluated at the Columbus Children’s Hospital included
S. pneumoniae (n = 12), Neisseria meningitidis (n = 6),
and Streptococcus agalactiae (n = 1). Information per-
mitting identification of the specific types of enterovi-
ruses isolated (culture ⁄ PCR) from individual patients
was unavailable.
Score
The following cut-points were identified by binary
recursive partitioning for discriminating between bacte-
rial and enteroviral meningitis: leukocyte concentration
in peripheral blood >17,000 cells ⁄ dL, ratio of bands to
neutrophils in peripheral blood >0.11 (11%), glucose
in CSF <38 mg ⁄ dL, protein in CSF >97 mg ⁄ dL
(5.38 mmol ⁄ L), WBCs in CSF >597 cells ⁄ lL, and
Table 1
Sensitivity and Specificity of Laboratory Test Thresholds Selected by Recursive Partitioning in Derivation Data Set*
Laboratory Test Incorporated Threshold Designated
into Score to be Positive
Peripheral blood leukocytes >17,000 cells ⁄ mL
Peripheral blood B:N ratio  >11%
CSF leukocyte concentration >597 cells ⁄ lL
CSF neutrophil concentration >74%
CSF total protein concentration >97 mg ⁄ dL
CSF glucose concentrationà <38 mg ⁄ dL
*Children seen at Columbus Children’s Hospital from 1998 to 2002.
 Ratio of bands to neutrophils (B:N) in peripheral blood.
àTo convert mg ⁄ dL to mmol ⁄ L divide by 18 (or multiply by 0.055).
CSF = cerebral spinal fluid.
Bonsu et al. • PREDICTING BACTERIAL MENINGITIS IN CHILDREN
neutrophils in CSF >74%. Sensitivity and specificity val-
ues for the derivation data set of cut-points selected for
each laboratory test by the partitioning procedure are
reported in Table 1. No individual laboratory test had
both good sensitivity and specificity (CSF glucose
<38 mg ⁄ dl did have a specificity ⁄ positive predictive
value of 100%). When combined, however, the accuracy
of the resultant 6-point prediction score was enhanced
substantially (Tables 2 and 3).
Internal validation (Columbus Children’s Hospital)
The resultant score discriminated accurately between
bacterial and enteroviral meningitis when validated by
the bootstrap procedure, AROC value 0.98 (95%
CI = 0.96 to 1.00), with no need to correct for bias. The
AROC for the derivation sample remained the same
(0.98, 95% CI = 0.96 to 1.00) when missing scores were
replaced with values based on a multiple imputation
procedure. The rule had a sensitivity of 100% and a
specificity of 34% at a score of 0 and a sensitivity of
100% and a specificity of 85% at a score of 1 (Table 2).
Two separate scores (0 and 1), therefore, were
observed to have 100% sensitivity for diagnosing bacte-
rial meningitis. LRs for separating bacterial from viral
meningitis were 0 (95% CI = 0 to 0.5) at a score of zero,
0 (95% CI = 0 to 0.3) at a score of one, 1 (95% CI = 0.3
to 3) at a score of two, 2 (95% CI = 0.1 to 16) at a score
of three, and 50 (95% CI = 7 to 350) at a score greater
or equal to four.
External Validation (Columbus Children’s Hospital)
Between March 2002 and June 2004, a total of 14
pathogens were isolated from children diagnosed with
bacterial meningitis at Columbus Children’s Hospital.
These included N. meningitidis (n = 6), S. pneumoniae
(n = 3), S. agalactiae (n = 4), and Escherichia coli
(n = 1). There were 144 controls (54 with confirmed
enteroviral meningitis) after removing children with
known neurologic or neurosurgical diseases. For 7 of
14 cases and 68 of 144 controls, there was insufficient
information to calculate the full score, so the maximum
score based on available test results was substituted.
The median age of children diagnosed with bacterial
meningitis was 1.54 years (IQR = 0.29–11.88 years).
Controls had a median age of 4.84 years (IQR = 0.22–
9.72 years).
Sensitivity, n (%), Specificity, n (%),
19 Bacterial Cases 59 Enteroviral Controls
12 (63) 55 (93)
16 (84) 29 (49)
12 (63) 55 (93)
14 (74) 49 (83)
16 (84) 55 (93)
14 (74) 59 (100)
ACAD EMERG MED • May 2008, Vol. 15, No. 5 • www.aemj.org 441

Table 2
Sensitivity and Specificity of the Decision Rule in Derivation and Validation Data Sets

Columbus Children’s Columbus Children’s Boston Children’s Cases and Controls from
Hospital Derivation Set Hospital Validation Set Hospital Validation Set All Three Data Sets
(1998–2002) (2002–2004) (1993–1999) (Columbus and Boston)
Sensitivity,* Specificity,* Sensitivity,  Specificity,  Sensitivity,à Specificity,à Sensitivity,§ Specificity,§
Prediction n (%), n (%), n (%), n (%), n (%), n (%), n (%) [CI], n (%) [CI],
Score 19 Cases 59 Controls 14 Cases 144 Controls 37 Cases 834 Cases 70 Cases 1,037 Controls
0 19 (100) 20 (34) 14 (100) 61 (42) 37 (100) 369 (44) 70 (100) 450 (43)
[95, 100] [40, 46]
1 19 (100) 50 (85) 14 (100) 114 (79) 36 (97) 675 (81) 69 (99) 839 (81)
[92, 100] [78, 83]
2 17 (90) 56 (95) 10 (71) 127 (88) 32 (87) 794 (95) 59 (84) 977 (94)
[74, 92] [93, 96]
3 16 (84) 58 (98) 5 (36) 133 (92) 25 (68) 820 (98) 46 (66) 1,011 (97)
[53, 77] [96, 98]
4 7 (37) 59 (100) 1 (7) 139 (97) 12 (32) 827 (99) 20 (29) 1,025 (99)
[18, 41] [98, 99]

CI = confidence interval.
*Sensitivity above and specificity below or equal to prediction score for cases with acute bacterial meningitis and controls with
acute enteroviral meningitis.
 Sensitivity above and specificity below or equal to stated prediction score for cases of acute bacterial meningitis and controls
based on the maximum score derived from available laboratory tests (see text for an explanation of how to calculate the score).
àSensitivity above and specificity below or equal to stated prediction score for children with complete reporting of laboratory
test results (see text).
§Sensitivity above and specificity below or equal to stated prediction score with 95% CIs for three data sets combined.

Table 3
Interval Likelihood Ratios (LR) of the Decision Rule and a Pragmatic Alternative* in External Data Set 

Original prediction rule Pragmatic prediction rule*

Prediction Cases, Controls, ILRà Cases, Controls, ILRà

Score n ⁄ 37 (%) n ⁄ 834 (%) (95% CI) n ⁄ 37 (%) n ⁄ 834 (%) (95% CI)
0 0 (0) 369 (44.2) 0 (0, 0.2) 0 (0) 333 (39.9) 0 (0, 0.2)
1 1 (2.7) 306 (36.7) 0.07 (0.01, 0.5) 1 (2.7) 331 (39.6) 0.07 (0.01, 0.5)
2 4 (10.8) 119 (14.3) 0.8 (0.3, 1.9) 3 (8.1) 124 (14.9) 0.5 (0.2, 1.6)
3 7 (18.9) 26 (3.1) 6 (3, 14) 9 (24.3) 30 (3.6) 7 (4, 13)
‡4 25 (67.6) 14 (1.6) 40§ (23, 71) 24 (64.8) 16 (1.9) 34§ (20, 58)

To convert mg ⁄ dL to mmol ⁄ L divide by 18 (or multiply by 0.055).

CI = confidence interval.
*The pragmatic (easier to memorize) rule substitutes the following cut-points for thresholds in the original rule (compare
Table 1): leukocytes >600 cells ⁄ lL, neutrophils >75%, total protein >100 mg ⁄ dL, glucose <40 mg ⁄ dL in CSF, and bands-to-
neutrophils ratio >10% in peripheral blood. The threshold for the peripheral blood leukocyte count stays unchanged at
>17,000 cells ⁄ mL. Each positive test result is assigned a single point and aggregated into a final score.
 Boston Children’s Hospital: largest data set; provides the most precise estimates of interval LRs.
àInterval LR at stated prediction score with 95% CIs.
§Interval LR at or above a prediction score of 4; interval LR at or above a prediction score of 3 was 17 (95% CI = 12 to 24) for ori-
ginal rule and 16 (95% CI = 12 to 22) for pragmatic rule.

Despite the fact that a full report of tests to permit
complete scoring was unavailable in approximately
50% of cases and controls, the AROC in these children
was 0.90 (95% CI = 0.88–0.92). No case of bacterial men-
ingitis was missed at a score of 0 or 1 (Table 2). Sensitiv-
ity and specificity values were 100 and 42% at a score of
>0 and 100% and 79% at a score of >1. The LR increased
from 0 at scores of 0 and 1 to 9 at a score of >2.
External Validation (Children’s Hospital Boston)
Thirty-seven pathogens were isolated from children
diagnosed with bacterial meningitis at the Children’s
Hospital Boston. Pathogens included S. pneumoniae
(21), N. meningitidis (7), S. agalactiae (5), E. coli (2), and
H. influenzae Type b (2). There were 834 controls after
removing children with known neurosurgical ⁄ neuro-
logic diseases. The median age of children diagnosed
with bacterial meningitis was 0.97 years (IQR = 0.39–
2.73 years). Controls had a median age of 0.19 years
(IQR = 0.12–1.02 years)
The AROC among children evaluated at Children’s
Hospital Boston was 0.97 (95% CI = 0.94 to 0.99). No
case of bacterial meningitis was missed at a score of 0
(Table 2). This score had a sensitivity of 100% and a
specificity of 44%. LRs were more precise and
increased from 0 and 0.07 at scores of 0 and 1, respec-
442
Figure 1. Receiver-operator-characteristic (ROC) curve analysis:
proportion of cases and controls at or above sequential
cut-points of the original prediction rule, Boston Children’s Hos-
pital. Area under the curve = 0.97. Numbers (0 to 6) that are
adjacent to the points on the curve are sequential thresholds of
the prediction rule.
tively, to 40 at a score of ‡4. Unlike the aforementioned
sets, one case of bacterial meningitis was missed at a
score equal to 1. This was an 8-year-old boy diagnosed
with pneumococcal meningitis who had the following
laboratory test results: CSF leukocytes 279 cells ⁄ lL,
CSF percent neutrophils 87%, CSF protein concentra-
tion 31.8 mg ⁄ dL, CSF glucose concentration 63 mg ⁄ dL
(3.5 mmol ⁄ L), peripheral WBC count 9360 cells ⁄ mL, and
a bands-to-neutrophils ratio in peripheral blood 0.01
(1%).
Similar test performance characteristics (Table 3)
were observed for a pragmatic rule that for ease of
memorization replaced thresholds in the original score
with the following values: bands-to-neutrophils ratio in
peripheral blood >10%; leukocytes >17,000 cells ⁄ mL
(left unchanged); and neutrophils >75%, leukocytes
>600 cells ⁄ lL, total protein >100 mg ⁄ dL, and glucose
<40 mg ⁄ dL (2.22 mmol ⁄ L) in CSF. The AROC for this
rule (Boston data set) was 0.96 (95% CI = 0.94 to 0.98;
Figure 1).
Finally, at a prediction score of 0, the sensitivity of
our score when combining 70 cases of acute bacterial
meningitis from the three data sets was 100% (95%
CI = 95% to 100%; Table 2) and beyond the range pub-
lished for the CSF Gram stain result (40%–80%). Speci-
ficity at the same score when combining 1037 controls
from the three data sets was 43% (95% CI = 40% to
46%). At a score of 1, sensitivity and specificity values,
respectively, when combining all cases and controls in
the three data sets were 99% (95% CI = 92% to 100%)
and 81% (95% CI = 78% to 83%).
DISCUSSION
In this study, we present a new rule for diagnosing bac-
terial meningitis that does not require complex mathe-
matical computation and is independent of the CSF
Gram stain result, but appears sufficiently accurate to
permit the timely diagnosis of this infection in children
with CSF pleocytosis. We believe that this rule can be
Bonsu et al. • PREDICTING BACTERIAL MENINGITIS IN CHILDREN
used to support clinical decisions relating to initial
treatment and hospitalization of children suspected to
have meningitis in clinical settings for which other vali-
dated rules published in the medical literature are
limited or inapplicable.
It is worth reemphasizing that our prediction score
extends rather than replaces existing rules. Utilized
within a Bayesian framework of diagnosis, our score
adjusts the prior odds of acute bacterial meningitis
based on information gathered during the clinical
examination to derive the final odds of this infection.30
It is likely to be most beneficial when clinical findings
are inconclusive and Gram stain results of CSF are neg-
ative or unavailable. In this setting, a low score is likely
to corroborate a truly negative CSF Gram stain result
and to exclude bacterial infection, allowing clinicians to
pursue less aggressive management strategies. Alterna-
tively, low odds of bacterial meningitis may be adjusted
higher if the score shows a high likelihood of bacterial
meningitis, avoiding delays to diagnosis and antibiotic
treatment.
Our prediction score is not a substitute for careful
clinical examination and judgment. Indeed, proper use
presupposes good clinical assessment skills and acu-
men. Additionally, within a Bayesian framework, diag-
nosis that remains inconclusive after application of our
score may be further strengthened by combining LRs
reported here with those estimated for other laboratory
tests. Tests utilized at other sites that are reported to be
helpful for diagnosing bacterial meningitis include
serum procalcitonin and C-reactive protein.33 More
research, however, is needed to reliably estimate the
incremental value of newer (over traditional) laboratory
tests for clinical diagnosis and for formalizing these
and other diagnosis tools for decision-making.
LIMITATIONS
Our study inherits all the limitations generally associ-
ated with a retrospective cohort design. It avoids, how-
ever, biases and ambiguities inherent to rules derived
from a review of subjective findings by relying on
quantitative test results alone. To create our rule, we
relied exclusively on cases of bacterial and enteroviral
meningitis confirmed by culture or nucleic acid amplifi-
cation techniques. We chose this strict definition to
avoid disease misclassification resulting from the
unpredictable yield of bacteria from CSF culture in chil-
dren receiving antibiotic treatment.15,34,35 Even so, and
despite the imposition of reasonable safeguards, over-
fitting due to the recursive partitioning procedure is
possible. We also expect data sets to include a fraction
of children pretreated with antibiotics because informa-
tion about prior treatment is difficult to verify, retro-
spectively. This last limitation is compounded by the
operational difficulty that even when the interval from
prior antibiotic therapy to the ED encounter is recorded
reliably, it still is difficult to select a realistic ⁄ sensible
time frame beyond which to assert indisputably that
such therapy has (vs. has not) ceased to confound
spinal fluid test results.
These restrictions (as well as others that were poten-
tially missed) had the potential to lower the
ACAD EMERG MED • May 2008, Vol. 15, No. 5 • www.aemj.org
performance of the rule in new subjects. Reassuringly,
our resultant rule showed no substantial bias when
evaluated by the bootstrap procedure, and when vali-
dated in children from other sites and time intervals,
demonstrated good temporal, geographic, and method-
ologic transportability. We believe, therefore, that it
captures robust predictors of acute bacterial meningitis
(those that are invariant to these factors), generalizing
sufficiently well for clinical diagnosis and management
in diverse settings.23–25 Validation by other investiga-
tors is needed to solidify our findings and to further
clarify the usefulness and practical application of this
decision rule.23–25
CONCLUSIONS
The new rule reported in this study and intended,
primarily, for children with unavailable, delayed, or
negative CSF Gram stain and for sites that lack auto-
mated decision support capability appears to have
adjunctive value for identifying acute bacterial menin-
gitis at the initial clinical encounter. At a score of 0,
it substantially lowers the prior odds of bacterial
meningitis in children with CSF pleocytosis. Pending
verification, this rule appears able to support judi-
cious hospitalization and treatment in a wide array of
children evaluated for meningitis in an outpatient
setting.
References
1. Schuchat A, Robinson K, Wenger JD, et al. Bacte-
rial meningitis in the United States in 1995. Active
Surveillance Team. N Engl J Med. 1997; 337:970–6.
2. Haddy RI, Perry K, Chacko CE, et al. Comparison
of incidence of invasive Streptococcus pneumoniae
disease among children before and after introduc-
tion of conjugated pneumococcal vaccine. Pediatr
Infect Dis J. 2005; 24:320–3.
3. Nigrovic LE, Kuppermann N, Malley R. Develop-
ment and validation of a multivariable predictive
model to distinguish bacterial from aseptic meningi-
tis in children in the post-Haemophilus influenzae
era. Pediatrics. 2002; 110:712–9.
4. Geiseler PJ, Nelson KE, Levin S, Reddi KT, Moses
VK. Community-acquired purulent meningitis: a
review of 1,316 cases during the antibiotic era,
1954-1976. Rev Infect Dis. 1980; 2:725–45.
5. Wald ER, Kaplan SL, Mason EO Jr, et al. Dexametha-
sone therapy for children with bacterial meningitis.
Meningitis Study Group. Pediatrics. 1995; 95:21–8.
6. Marton KI, Gean AD. The spinal tap: a new look at
an old test. Ann Intern Med. 1986; 104:840–8.
7. Aronin SI, Peduzzi P, Quagliarello VJ. Community-
acquired bacterial meningitis: risk stratification for
adverse clinical outcome and effect of antibiotic tim-
ing. Ann Intern Med. 1998; 129:862–9.
8. Auburtin M, Wolff M, Charpentier J, et al. Detri-
mental role of delayed antibiotic administration and
penicillin-nonsusceptible strains in adult intensive
care unit patients with pneumococcal meningitis:
the PNEUMOREA prospective multicenter study.
Crit Care Med. 2006; 34:2758–65.
443
9. Lepur D, Barsic B. Community-acquired bacterial
meningitis in adults: antibiotic timing in disease
course and outcome. Infection. 2007; 35:225–31.
10. Proulx N, Frechette D, Toye B, Chan J, Kravcik S.
Delays in the administration of antibiotics are asso-
ciated with mortality from adult acute bacterial
meningitis. QJM 2005; 98:291–8.
11. Miner JR, Heegaard W, Mapes A, Biros M. Presen-
tation, time to antibiotics, and mortality of patients
with bacterial meningitis at an urban county medi-
cal center. J Emerg Med. 2001; 21:387–92.
12. Bonsu BK, Harper MB. Differentiating acute bacte-
rial meningitis from acute viral meningitis among
children with cerebrospinal fluid pleocytosis: a mul-
tivariable regression model. Pediatr Infect Dis J.
2004; 23:511–7.
13. Nigrovic LE, Kuppermann N, Macias CG, et al. Clin-
ical prediction rule for identifying children with
cerebrospinal fluid pleocytosis at very low risk of
bacterial meningitis. JAMA. 2007; 297:52–60.
14. Hoen B, Viel JF, Paquot C, Gerard A, Canton P.
Multivariate approach to differential diagnosis of
acute meningitis. Eur J Clin Microbiol Infect Dis.
1995; 14:267–74.
15. Leblebicioglu H, Esen S, Bedir A, Gunaydin M, San-
ic A. The validity of Spanos’ and Hoen’s models for
differential diagnosis of meningitis. Eur J Clin
Microbiol Infect Dis. 1996; 15:252–4.
16. Spanos A, Harrell FE Jr, Durack DT. Differential
diagnosis of acute meningitis. An analysis of the
predictive value of initial observations. JAMA. 1989;
262:2700–7.
17. Baty V, Viel JF, Schuhmacher H, Jaeger F, Canton
P, Hoen B. Prospective validation of a diagnosis
model as an aid to therapeutic decision-making in
acute meningitis. Eur J Clin Microbiol Infect Dis.
2000; 19:422–6.
18. Freedman SB, Marrocco A, Pirie J, Dick PT. Predic-
tors of bacterial meningitis in the era after Haemo-
philus influenzae. Arch Pediatr Adolesc Med. 2001;
155:1301–6.
19. Oostenbrink R, Oostenbrink JB, Moons KG, et al.
Application of a diagnostic decision rule in children
with meningeal signs: a cost-minimization study. Int
J Technol Assess Health Care. 2003; 19:698–704.
20. Dubos F, Lamotte B, Bibi-Triki F, et al. Clinical deci-
sion rules to distinguish between bacterial and
aseptic meningitis. Arch Dis Child. 2006; 91:647–50.
21. Hastie T, Tibshirani R, Friedman JH. The Elements
of Statistical Learning: Data Mining, Inference, and
Prediction: With 200 Full-Color Illustrations.
Springer Series in Statistics: Corrected Print. New
York, NY: Springer, 2002.
22. Harrell FE. Regression Modeling Strategies: With
Applications to Linear Models, Logistic Regression,
and Survival Analysis. Springer Series in Statistics.
New York, NY: Springer, 2001.
23. Justice AC, Covinsky KE, Berlin JA. Assessing the
generalizability of prognostic information. Ann
Intern Med. 1999; 130:515–24.
24. Altman DG, Royston P. What do we mean by validat-
ing a prognostic model? Stat Med. 2000; 19:453–73.
444
25. Stiell IG, Bennett C. Implementation of clinical deci-
sion rules in the emergency department. Acad
Emerg Med. 2007; 14:955–9.
26. Hosmer DW, Lemeshow S. Applied Logistic
Regression. Wiley Series in Probability and Statis-
tics. Texts and References Section, 2nd ed. New
York, NY: Wiley, 2000.
27. Zweig MH, Campbell G. Receiver-operating charac-
teristic (ROC) plots: a fundamental evaluation tool
in clinical medicine. Clin Chem. 1993; 39:561–77.
28. Hanley JA, McNeil BJ. The meaning and use of the
area under a receiver operating characteristic
(ROC) curve. Radiology. 1982; 143:29–36.
29. Swets JA. Measuring the accuracy of diagnostic
systems. Science. 1988; 240:1285–93.
30. Gallagher EJ. Clinical utility of likelihood ratios.
Ann Emerg Med. 1998; 31:391–7.
31. Jaeschke R, Guyatt GH, Sackett DL. Users’ guides
to the medical literature. III. How to use an article
Bonsu et al. • PREDICTING BACTERIAL MENINGITIS IN CHILDREN
about a diagnostic test. B. What are the results and
will they help me in caring for my patients? The
Evidence-Based Medicine Working Group. JAMA.
1994; 271:703–7.
32. Seehusen DA, Reeves MM, Fomin DA. Cerebrospinal
fluid analysis. Am Fam Physician. 2003; 68:1103–8.
33. Gendrel D, Raymond J, Assicot M, et al. Measure-
ment of procalcitonin levels in children with bacte-
rial or viral meningitis. Clin Infect Dis. 1997;
24:1240–2.
34. Shohet I, Shahar E, Meyerovich J, Barzilay Z. Diag-
nosis of bacterial meningitis in previously treated
children. South Med J. 1985; 78:299–301.
35. Mandell GL, Douglas RG, Bennett JE, Dolin R.
Mandell, Douglas, and Bennett’s Principles and
Practice of Infectious Diseases, 5th ed. Philadelphia,
PA: Churchill Livingstone, 2000.