Bacterial meningitis remains an important cause

Transcription

1 Development and Validation of a Multivariable Predictive Model to Distinguish Bacterial From Aseptic Meningitis in Children in the Post-Haemophilus influenzae Era Lise E. Nigrovic, MD*; Nathan Kuppermann, MD, MPH ; and Richard Malley, MD* ABSTRACT. Context. Children with meningitis are routinely admitted to the hospital and administered broad-spectrum antibiotics pending culture results because distinguishing bacterial meningitis from aseptic meningitis is often difficult. Objective. To develop and validate a simple multivariable model to distinguish bacterial meningitis from aseptic meningitis in children using objective parameters available at the time of patient presentation. Design. Retrospective cohort study of all children with meningitis admitted to 1 urban children s hospital from July 1992 through June 2000, randomly divided into derivation (66%) and validation sets (34%). Patients. Six hundred ninety-six previously healthy children aged 29 days to 19 years, of whom 125 (18%) had bacterial meningitis and 571 (82%) had aseptic meningitis. Intervention. Multivariable logistic regression and recursive partitioning analyses identified the following predictors of bacterial meningitis from the derivation set: Gram stain of cerebrospinal fluid (CSF) showing bacteria, CSF protein >80 mg/dl, peripheral absolute neutrophil count > cells/mm 3, seizure before or at time of presentation, and CSF absolute neutrophil count >1000 cells/mm 3. A Bacterial Meningitis Score (BMS) was developed on the derivation set by attributing 2 points for a positive Gram stain and 1 point for each of the other variables. Main Outcome Measure. The accuracy of the BMS when applied to the validation set. Results. A BMS of 0 accurately identified patients with aseptic meningitis without misclassifying any child with bacterial meningitis in the validation set. The negative predictive value of a score of 0 for bacterial meningitis was 100% (95% confidence interval: 97% 100%). A BMS >2 predicted bacterial meningitis with a sensitivity of 87% (95% confidence interval: 72% 96%). Conclusions. The BMS accurately identifies children at low (BMS 0) or high (BMS >2) risk of bacterial meningitis. Outpatient management may be considered for children in the low-risk group. Pediatrics 2002;110: ; prediction model, bacterial meningitis, aseptic meningitis. From the *Department of Medicine and Divisions of Infectious Diseases and Emergency Medicine, Children s Hospital and Harvard Medical School, Boston, Massachusetts, and the Department of Internal Medicine, Division of Emergency Medicine and the Department of Pediatrics, University of California, Davis, School of Medicine, Davis, California. Received for publication Feb 19, 2002; accepted May 20, Reprint requests to (R.M.) Divisions of Infectious Diseases and Emergency Medicine, Children s Hospital, 300 Longwood Ave, Boston MA richard.malley@tch.harvard.edu PEDIATRICS (ISSN ). Copyright 2002 by the American Academy of Pediatrics. ABBREVIATIONS. CSF, cerebrospinal fluid; WBC, white blood cell count; Hib, Haemophilus influenzae type b; ANC, absolute neutrophil count; RBC, red blood cells; ROC, receiver operating characteristic; BMS, Bacterial Meningitis Score; NPV, negative predictive value; CI, confidence interval; PPV, positive predictive value. Bacterial meningitis remains an important cause of morbidity and mortality in children despite the advent of highly effective bacterial conjugate vaccines. In the United States each year, there are close to 6000 new cases of bacterial meningitis, of which approximately half occur in children younger than 18 years of age. 1 Streptococcus pneumoniae and Neisseria meningitidis are the major bacterial pathogens in children beyond the immediate neonatal period 1 and have associated overall mortality rates of 6% to 12% 2 4 and 3% to 5%, 1,5 respectively. Despite antimicrobial therapy, up to 25% of survivors of bacterial meningitis have significant sequelae, such as neurologic deficits or hearing loss. 6 Because discrimination between bacterial meningitis and aseptic meningitis at the time of presentation is often difficult, children with cerebrospinal fluid (CSF) pleocytosis are routinely admitted to the hospital to receive broad-spectrum antibiotics pending bacterial culture results. Previously, investigators have evaluated means to distinguish bacterial from aseptic meningitis before the results of blood and/or CSF cultures are available. When evaluated on patients not previously treated with antibiotics, the CSF Gram stain has been reported to have a sensitivity between 60% and 92%. 7,8 Other rapidly available parameters, such as the peripheral white blood cell count (WBC) and the CSF WBC with differential, have wide zones of overlap in patients with bacterial and aseptic meningitis Measurements of CSF leukocyte aggregation, CSF lactate, and serum procalcitonin lack either sensitivity or specificity and are not routinely available The serum concentration of C-reactive protein has been evaluated as well, but has limited value as it may be low early in bacterial disease or elevated in patients with viral infections. 23,24 Endogenous inflammatory mediators such as CSF tumor necrosis factor-, interleukin-1, or interleukin-6 have good predictive power but are not routinely available in the clinical setting. 25,26 Two multivariable models have previously been developed to distinguish bacterial versus aseptic meningitis. The first study which was derived from 712 PEDIATRICS Vol. 110 No. 4 October 2002

2 an all-ages population and before the widespread introduction of Haemophilus influenzae type b (Hib) conjugate vaccine, and which used CSF absolute neutrophil count (ANC), CSF-blood glucose ratio, age, and time of year demonstrated excellent sensitivity and specificity for the prediction of bacterial versus aseptic meningitis. 27 A second, more recent study was based on a pediatric population in the era of universal Hib vaccination and used the following dichotomous predictors: CSF WBC 30 cells/mm 3, CSF-blood glucose ratio 40%, CSF glucose 40 mg/dl, CSF protein 45 mg/dl, positive Gram stain, peripheral band count 500 cells/mm 3, and age 6 months. 13 This model also had excellent predictive ability; however, the authors did not attempt to validate their findings, which makes it difficult to assess the robustness and generalizability of their model. Because most children admitted to the hospital for meningitis have viral rather than bacterial infections, we wished to develop and validate a simple model to predict bacterial versus aseptic meningitis in a pediatric population in the era of Hib vaccination. We also sought to identify a low-risk group of patients who may be candidates for outpatient management. METHODS Patient Population We reviewed the charts of all patients aged 29 days to 19 years who were admitted to Children s Hospital, Boston, between July 1, 1992, and June 30, 2000, with a final diagnosis of meningitis. A computerized search tool was used to identify patients with the following International Classification of Diseases diagnosis codes: bacterial meningitis ( ), viral meningitis ( ), tuberculous meningitis (013.0), and unspecified meningitis ( ). Definitions Patients were defined as having bacterial meningitis if either of the following 2 criteria were met: 1) the CSF culture was positive for a bacterial pathogen; or 2) the presence of CSF pleocytosis ( 7 WBC per mm 3 ) 28 and either a positive blood culture or CSF latex agglutination test positive for S pneumoniae, N meningitidis, H influenzae, orstreptococcus agalactiae (group B Streptococcus). Patients in whom CSF culture revealed an organism usually associated with contamination (such as Staphylococcus epidermidis or Propionobacterium acnes in previously healthy patients) were excluded. Patients were categorized as having aseptic meningitis if they had CSF pleocytosis with negative bacterial cultures as well as negative CSF latex agglutination tests, if performed. Patients meeting this definition but who had been given systemic antibiotics within 72 hours before lumbar puncture and blood culture (n 108) were not included in the analysis because the accuracy of these pretreated cultures could not be assured and thus, we could not be certain that these patients truly had aseptic meningitis. Exclusion Criteria To eliminate patients who would have required hospitalization regardless of the risk of bacterial meningitis, the following criteria were developed. We excluded patients who recently underwent neurosurgical procedures (n 34), presented with clinical sepsis or purpura fulminans (n 22), or were immunosuppressed (because of chemotherapy for oncologic diagnoses [n 5], who underwent organ transplantation [n 3], had human immunodeficiency infection [n 4], or had primary immunodeficiencies [n 2]). We also excluded patients (n 10) who had other focal bacterial infections requiring inpatient treatment that were recognized at the time of admission, including children younger than 3 months with urinary tract infections (n 7), as well as children with brain abscesses (n 2) or periorbital cellulitis (n 1). Data Collection We collected the following variables by a review of the computerized and/or written medical records and entered the information onto structured data sheets: date of birth, date of admission and discharge, gender, previous and present medical history conditions, duration and peak of fever at admission, antibiotic pretreatment, occurrence and timing of seizures, peripheral WBC and differential, CSF WBC and differential, CSF red blood cell count (RBC), CSF glucose, CSF protein, Gram stain, blood and CSF cultures, and latex agglutination studies if performed. The presence of seizures was based on classification of the treating physician. The enteroviral season was defined as the time period between June 1 and October The laboratory values obtained closest in time to the lumbar puncture were used. When 2 CSF cell counts were performed, the tube with fewest RBCs was used regardless of the recorded tube number. Statistical Methods Univariate and multivariate analyses were conducted using the Statistical Program for the Social Sciences. 30 Confidence intervals for proportions were calculated using Stata statistical software. 31 Answer Tree statistical software was used to perform recursive partitioning of the data. 32 The database was randomly divided into a derivation set comprising two thirds of the patients and a validation set comprising the remaining patients. The Student t test was used to compare means between the derivation and validation sets and to identify variables in the derivation set that were significantly associated (P.05) with bacterial meningitis. To limit the number of variables considered, only those that were objective, biologically plausible, readily available at the time of hospital presentation, and correlated with bacterial meningitis in previous studies were chosen. Because of the colinearity of some of these variables, we selected peripheral ANC over WBC and CSF ANC over CSF WBC based on inspection of the receiver operating characteristic (ROC) curves of these variables against bacterial meningitis. To obtain a simple and practical model, continuous predictors were dichotomized using rounded whole number values that allowed effective discrimination between bacterial meningitis and aseptic meningitis. We selected these cutoffs by identifying predictor variable values associated with a point of the ROC curve with the best sensitivity-specificity trade off. These optimal cutoff points typically occur where the ROC curve turns the corner, at which point each incremental gain in sensitivity results in a substantial loss of specificity. We then rounded off these values for numerical and clinical simplicity. We then performed 2 multivariable statistical analyses on the derivation set to identify significant predictors of bacterial meningitis. Only those patients with complete information on all candidate variables were considered in the multivariable analyses. First, we entered the candidate variables as described above into a forward stepwise multivariable logistic regression analysis. Variables independently associated (P.05) with bacterial meningitis in this analysis were then ranked according to the magnitude of the -coefficient. Second, we performed binary recursive partitioning using the candidate variables described above (ie, the same dichotomous predictors used in the logistic regression analysis) using a classification tree algorithm 32 to identify the most important predictors of bacterial meningitis. As another form of multivariate analysis, recursive partitioning may be superior to logistic regression when the goal is to correctly classify patients with the outcome of interest with the highest sensitivity rather than obtain overall model accuracy. Recursive partitioning also explores interactions between predictor variables natively in the algorithm. Finally, the trees developed by the recursive partitioning software are easily interpreted by investigators and clinicians alike and are well suited for use in clinical decision-making. 33,34 In the decision tree development of the derivation set, we considered the same candidate variables as were considered in the logistic regression analysis, and we then hand pruned the resultant tree for simplicity and plausibility. For ease of use in the clinical setting, we then created a Bacterial Meningitis Score (BMS), using the predictor variables identified by ARTICLES 713

3 both of these multivariable statistical methods. The relative weighting of each component variable of the BMS score was based on its value in the logistic regression analysis. The performance of the BMS was then evaluated on patients in the validation set. We also conducted 2 subset analyses of the data. To estimate the performance of the model in the era of routine pneumococcal conjugate vaccination, we performed a subgroup analysis excluding cases of S pneumoniae meningitis. Patients with missing values were excluded in the multivariable analyses. We also analyzed the subset of patients between 29 and 60 days of age to determine how well the prediction model identified children with bacterial meningitis among this younger, high-risk group. Approval for review of the medical records was granted by the institutional review board of Children s Hospital, Boston. RESULTS Patient Characteristics We identified 696 children, including 125 (18%) with bacterial meningitis and 571 (82%) with aseptic meningitis. Cases of bacterial meningitis were caused by the following pathogens: S pneumoniae (79, 63%), N meningitidis (21, 17%), group B Streptococcus (13, 10%), H influenzae (6, 5%, all nontypeable), Listeria monocytogenes (3, 2%), enteric Gram-negative rods (2, 2%), and other streptococci (1, 1%). No cases of tuberculous meningitis were identified. The bacterial pathogen was identified in both CSF and blood culture in 60 patients (48%), CSF culture alone in 35 (28%), blood culture alone in 28 (22%), and by latex agglutination alone in 2 (2%). Of the pretreated patients with bacterial meningitis, the pathogen was identified in both CSF and blood culture in 18 patients (35%), CSF culture alone in 12 (23%), blood culture alone in 22 (42%) and by CSF latex agglutination alone in 0. Of patients with bacterial meningitis, 42% had antibiotics administered within 72 hours before lumbar puncture (42% oral and 58% parenteral antibiotics). Latex agglutination tests were performed in 61 patients. Four patients died, all with bacterial meningitis (3% of total patients with bacterial meningitis). Comparison of the Derivation and Validations Sets Patients were randomly divided into 2 sets: the derivation set (n 456, 66% of patients) and the validation set (n 240, 34% of patients). Patients in the derivation set were, on average, older at presentation and more likely to be male (Table 1). There were no other statistically significant differences between the derivation and validation sets. Univariate Analysis of the Derivation Set Patients with bacterial meningitis were significantly more likely to be admitted in the nonenteroviral season and to have had seizures at or before presentation than patients with aseptic meningitis. Other variables commonly associated with bacterial infection were also significantly different in patients with bacterial compared with aseptic meningitis, including the peripheral ANC, CSF ANC, CSF glucose, CSF protein, and positive Gram stain (Table 2). Optimum cut offs for continuous variables were then selected by analyzing the ROC curve for each variable, as previously described (Fig 1). The following dichotomous variables were thus selected for consideration in the multivariate analyses: admission during nonenteroviral season, seizure at or before presentation, peripheral ANC cells/mm 3, CSF ANC 1000 cells/mm 3, CSF glucose 35 mg/ dl, CSF protein 80 mg/dl, and positive Gram stain. Multivariate Analyses Logistic Regression The variables listed above were entered into a forward stepwise logistic regression analysis, to identify independent predictors of bacterial meningitis (Table 3). The presence of a positive Gram stain was the most significant predictor of bacterial meningitis ( 5.6). The other variables independently associated with bacterial meningitis in this analysis were CSF protein 80 mg/dl, peripheral ANC cells/mm 3, seizure at or before presentation, and CSF ANC 1000 cell/mm 3. Only 4% of patients were excluded from this analysis because of missing values. Recursive Partitioning Recursive partitioning was used to identify important predictors in the derivation set (Fig 2). The same TABLE 1. Characteristics of Patients in the Derivation Set and Validation Set* Variable Derivation Set ( Standard Derivation) N 456 Validation Set ( Standard Deviation) N 240 Median age in mo (interquartile range) 5.0 ( ) 3.0 ( ) Male gender, n (%) 281 (62%) 126 (53%) Peak fever ( C) Duration of fever (d) Antibiotic pretreatment, n (%) 36 (8%) 16 (7%) Enteroviral season, n (%) 291 (64%) 154 (64%) Seizure at or before presentation, n (%) 33 (7%) 11 (5%) Mortality, n (%) 4 (1%) 0 (0%) Peripheral WBC ( 10 3 cells/mm 3 ) CSF WBC (cells/mm 3 ) CSF glucose (mg/dl) CSF protein (mg/dl) Positive Gram stain, n (%) 64 (14%) 31 (13%) Bacterial meningitis, n (%) 86 (19%) 39 (16%) * Means and their standard deviations are given for all continuous variables (with the exception of age, which is presented as a median because of substantial skewness). P PREDICTIVE MODEL TO DISTINGUISH BACTERIAL FROM ASEPTIC MENINGITIS

4 TABLE 2. Univariate Comparisons of Candidate Predictor Variables in the Derivation Set Between Patients With Bacterial and Aseptic Meningitis* Variable Bacterial Meningitis N 86 Aseptic Meningitis N 370 Difference Between Means or Proportions (95% CI) P Value Enteroviral season, n (%) 22 (26%) 269 (73%) 47% (37 57).001 Seizure at or before 19 (22%) 14 (4%) 18% (9 27).001 presentation, n (%) Peripheral ANC ( ( ).001 cells/mm 3 ) CSF ANC (cells/mm 3 ) ( ).001 CSF glucose (mg/dl) (12 26).001 CSF protein (mg/dl) (20 226).001 Positive Gram stain, n (%) 62 (72%) 2 (0.5%) 72% (62 81).001 * Means are given for continuous variables and proportions for categorical variables. Fig 1. ROC curves for each of the continuous univariate candidate predictors: peripheral ANC, CSF ANC, CSF glucose, and CSF protein. Arrow heads indicate the identified cut offs for dichotomization of the variables (see text). The 45 line through the origin represents the ROC curve of a test whose decision ability is no better than chance. TABLE 3. Multivariate Logistic Regression Analysis* Predictor Coefficient P Value 95% CI Gram stain CSF protein 80 mg/dl Peripheral ANC cells/mm Seizure at or before presentation CSF ANC 1000 cells/mm * Table gives the coefficient, P value, and odds ratio for each independent predictor (P.05). variables identified as independent predictors in the logistic regression analysis were identified as important predictors in the recursive partitioning analysis. The most important variable, identified at the top of the resultant decision tree, was the Gram stain result. For the patients with negative CSF Gram stains, the other predictors in the decision tree accurately discriminated bacterial from aseptic meningitis. At each step of the decision tree, the risk of bacterial meningitis can be estimated for any particular combination of predictor variables. Prediction Model Based on the results of both the logistic regression and the recursive partitioning analyses, we developed the BMS. Because the presence of a positive Gram stain was the strongest independent predictor of bacterial meningitis using both analytical methods, with the relative magnitude of the -coefficient roughly twice that of the other variables, 2 points were attributed to a positive Gram stain and 1 point for each of the remaining variables. The range of the resulting BMS was thus 0 to 6 points (Table 4). ARTICLES 715

5 Fig 2. Results of the recursive partitioning analysis on the derivation set. Each box gives the number (and percentage) of patients with bacterial and aseptic meningitis given the particular combination of predictor variables. Patients missing predictor variables have been excluded (n 19). The performance of the BMS was then tested on the validation set (Fig 3). A BMS of 0 accurately identified patients with aseptic meningitis without misclassifying any child with bacterial meningitis as having aseptic meningitis. The negative predictive value (NPV) of a score of 0 for bacterial meningitis was 100% (144/144, 95% confidence interval [CI]: 97% 100%), and the specificity was 73% (144/196, 95% CI: 91% 100%). A BMS 2 predicted bacterial meningitis with a sensitivity of 87% (33/38, 95% CI: 72% 96%), and a positive predictive value of 87% (33/38, 95% CI: 72% 96%). When applied to the entire data set, the BMS misclassified only 3.3% of the patients. Of all patients withabms 0 in the derivation and validation data sets (predicted to have a low likelihood of bacterial meningitis), 2 of 404 patients (0.5%; 95% CI: 0.06% 1.8%) actually had bacterial meningitis. The 2 patients thus misclassified were both in the derivation set. One was a 7-month-old child with pneumococcal meningitis who had a CSF WBC of 11 cells/mm 3 and peripheral ANC of cells/mm 3. This patient had been pretreated with parenteral antibiotics before the lumbar puncture. The second patient was a 6-month-old child with meningococcal meningitis TABLE 4. Bacterial Meningitis Score* Predictor Present Points Absent Positive Gram stain 2 0 CSF protein 80 mg/dl 1 0 Peripheral ANC cells/mm Seizure at or before presentation 1 0 CSF ANC 1000 cells/mm * Possible scores range from 0 to 6. Fig 3. Performance of the BMS for patients with aseptic and bacterial meningitis in the validation set. Error bars represent the upper end of the 95% CI. Above each bar is the number and percentage of patients with aseptic or bacterial meningitis with a given score. Patients missing predictor variables have been excluded (n 6). with CSF WBC of 10 cells/mm 3 and peripheral ANC of cells/mm 3. Subset Analyses We performed 2 subset analyses. In the first subset analysis, we evaluated the performance of the model after exclusion of patients infected with S pneumoniae. In our study, 63% of the children with bacterial meningitis cases were infected with S pneumoniae. Because the conjugate pneumococcal vaccine reduces the incidence of invasive pneumococcal infections attributable to vaccine-type strains, 35 the prediction rule was reexamined after exclusion of 716 PREDICTIVE MODEL TO DISTINGUISH BACTERIAL FROM ASEPTIC MENINGITIS

6 children with pneumococcal meningitis. When applied to the validation set, a BMS of 0 predicted aseptic meningitis with 73% sensitivity (144/196, 95% CI: 67% 80%), 100% specificity (11/11, 95% CI: 72% 100%), and positive predictive value of 100% (144/144, 95% CI: 97% 100%). A BMS 2 predicted bacterial meningitis with 100% sensitivity (11/11, 95% CI: 72% 100%), 97% specificity (191/196, 95% CI: 94% 99%), and 69% positive predictive value (11/16, 95% CI: 41% 89%). We performed a second subanalysis to evaluate the performance of the prediction rule on patients younger than 2 months old, because these children are typically difficult to evaluate clinically and likely have different risk of bacterial meningitis than older patients. Of the 77 patients between 29 to 60 days of age in the validation set, the prediction model accurately identified the 3 patients with bacterial meningitis. DISCUSSION Because accurately distinguishing bacterial meningitis from aseptic meningitis at the time of presentation is difficult, children with CSF pleocytosis are routinely admitted to the hospital for treatment with broad-spectrum antibiotics while awaiting results of bacterial cultures. We have developed a scoring system based on multivariate logistic regression and recursive partitioning analyses of data obtained from hospitalized children with bacterial and aseptic meningitis. This prediction rule is based solely on data that are objective and readily available at the time of presentation. As shown by the performance of the BMS on the validation set, this score accurately identifies patients both at high and very low risk of bacterial meningitis. Several findings of the current study are consistent with prior investigations. We found Gram stain, seizure at or before presentation, peripheral ANC, CSF ANC, and CSF protein to be significant univariate predictors of bacterial meningitis, as shown in previous analyses. 7 11,13,14 The considerable overlap of these variables between the patients with aseptic and bacterial meningitis, however, limits their discriminative ability when applied as univariate predictors. Our study differs from previous studies in being the first validated multivariate model derived from a pediatric population in the postconjugate Hib vaccine era. As compared with the multivariate model developed by Spanos and colleagues, 27 our study population comprises only children from an era in which meningitis attributable to Hib has become exceedingly rare because of widespread vaccination (and indeed, no cases of Hib meningitis were found in our population). In the model by Spanos, age was a highly predictive variable, whereas the age of the patient was not a significant predictor in our study. This difference may reflect the fact that their study included patients of all ages ( 1 month), whereas our study included only pediatric patients (29 days 18 years of age). Finally, although aseptic meningitis is more likely to occur in the enteroviral season, we did not find that time of the year provided significant independent discriminative power with respect to the prediction of bacterial versus aseptic meningitis. The predictive model by Freedman and colleagues 13 was developed on a pediatric population in the era of widespread Hib vaccination, but the absence of validation makes it difficult to ascertain the generalizability of their findings. The use of the BMS score could have important implications for clinicians caring for children with meningitis. First, it is simple to use, as it is based on variables that are readily available at the time of initial patient evaluation. Second, it allows for rapid and accurate discrimination of patients at either high or very low risk of bacterial meningitis. The determination that a patient has a high risk of bacterial meningitis would allow for prompt recognition of the severity of illness and immediate antibiotic administration and hospitalization, with transfer to a tertiary pediatric facility, if necessary. Conversely, patients with a BMS score of 0 can be considered to be at very low risk of bacterial meningitis. Hospitalization may not necessarily be warranted for some of these patients. Four hundred four (60%) admitted patients in our study had a BMS of 0 and thus could have been considered to be at very low risk of bacterial meningitis (with a misclassification rate of 0.5%). A subset of these patients including those who were very well appearing, well hydrated, and without any signs of neurologic or hemodynamic compromise could conceivably have been followed as outpatients, perhaps after the administration of a long-acting parenteral antibiotic. 36,37 Potential limitations of our study should be considered. Ideally, a prediction rule should be derived, and then be validated prospectively on a separate population. 38 However, because the incidence of bacterial meningitis is relatively low, our study was based on a retrospective analysis of data from hospitalized children over an 8-year period. We confirmed the validity and robustness of this model by evaluating the performance of the BMS on a separate, randomly selected, subset of patients. Although evaluation of the general appearance of the patients and prospective validation of the model would be ideal, the fact that all the variables used in the BMS are objective historical and laboratory parameters gathered at the time of initial presentation makes it unlikely that the results would greatly change had the study been conducted prospectively using the same predictor variables. We also had high rate of complete data collection; 96% of the patients in the database had no missing data for the analyzed variables. Because patients with aseptic meningitis who were pretreated were excluded from the analysis, this model should not be applied to patients who have received systemic antibiotics within 72 hours of lumbar puncture. The incidence of bacterial meningitis in our study was 18%, considerably higher than reported population estimates. This likely reflects referral bias, as patients with bacterial meningitis were transferred from other hospitals for admission more often than those with aseptic meningitis (52% vs 6%). Because this bias affects the prevalence of bacterial meningitis in our patient population, the PPVs and NPVs of the ARTICLES 717

7 BMS must be interpreted with this in mind. In contrast, however, the sensitivity and specificity of our scoring system are not affected by this potential bias. Finally, several of the lumbar puncture results may have been confounded by the presence of a high CSF RBC ( cells/mm 3 ). From the outset of the study, we chose not to attempt to apply any correction factor in these cases, as there is no routinely accepted method to adjust for CSF WBC or protein in the presence of elevated CSF RBC. 39,40 We believe that this decision did not adversely affect our results, however, because the number of children in which the CSF RBC was elevated was relatively low (n 23, 3% of all patients), but similar to other centers, 13 and all of these patients had aseptic meningitis. This would tend to bias against our model because of the corresponding increase in CSF WBC and protein in these children with aseptic meningitis who had elevated CSF RBC. The epidemiology of bacterial meningitis in the United States has been substantially transformed over the past decade with the advent of universal Hib immunization. The incidence of bacterial meningitis in children is likely to be reduced further over the next few years as a consequence of pneumococcal conjugate immunization. Nevertheless, because of the multiple pathogens responsible for pediatric meningitis, 1 the possibility of pneumococcal serotype replacement and the realities of underimmunization 44 and vaccine failures, 45,46 a prediction rule such as the BMS that identifies patients at very low risk of bacterial meningitis and who may be considered for outpatient therapy may be beneficial to patients and result in important cost savings. ACKNOWLEDGMENTS We wish to acknowledge the support provided by a resident research grant from the American Academy of Pediatrics (to Dr Nigrovic) and a grant by the Meningitis Research Foundation (to Dr Malley). REFERENCES 1. Schuchat A, Robinson K, Wenger JD, et al. Bacterial meningitis in the United States in N Engl J Med. 1997;337: Syriopoulou V, Daikos GL, Soulis K, et al. Epidemiology of invasive childhood pneumococcal infections in Greece. Acta Paediatr Suppl. 2000; 89: Arditi M, Mason EO, Bradley JS, et al. Three-year multicenter surveillance of pneumococcal meningitis in children: clinical characteristics and outcome related to penicillin susceptibility and dexamethasone use. Pediatrics. 1998;102: Stanek RJ, Mufson MA. A 20-year epidemiological study of pneumococcal meningitis. Clin Infect Dis. 1999;28: Malley R, Inkelis SH, Coelho P, Huskins WC, Kuppermann N. Cerebrospinal fluid pleocytosis and prognosis in invasive meningococcal disease in children. Pediatr Infect Dis J. 1998;17: Wald ER, Kaplan SL, Mason EO, et al. Dexamethasone therapy for children with bacterial meningitis. Pediatrics. 1995;95: Marton KI, Gean AD. The spinal tap: a new look at an old test. Ann Intern Med. 1986;104: Hristeva L, Bowler I, Booy R, King A, Wilkinson AR. Value of cerebrospinal fluid examination in the diagnosis of meningitis of the newborn. Arch Dis Child. 1993;69: Sato Y, Ohta Y, Honda Y, Kaji M, Oizumi K. Diagnostic value of atypical lymphocytes in cerebrospinal fluid from adults with enteroviral meningitis. J Neurol. 1998;245: Baker RC, Lenane AM. The predictive value of cerebrospinal fluid differential cytology in meningitis. Pediatr Infect Dis J. 1989;8: Negrini B, Kelleher KJ, Wald ER. Cerebrospinal fluid findings in aseptic versus bacterial meningitis. Pediatrics. 2000;105: Berkley JA, Mwangi I, Ngetsa CJ, et al. Diagnosis of acute bacterial meningitis at a district hospital in sub-saharan Africa. Lancet. 2001;357: Freedman SB, Marrocco A, Pirie J, Dick PT. Predictors of bacterial meningitis in the era after Haemophilus influenzae. Arch Pediatr Adolesc Med. 2001;155: Dunbar SA, Eason RA, Musher DM, Clarridge JE. Microscopic examination and broth culture of cerebrospinal fluid in diagnosis of meningitis. J Clin Microbiol. 1998;36: Garty BZ, Berliner S, Liberman E, Danon YL. Cerebrospinal fluid leukocyte aggregation in meningitis. Pediatr Infect Dis J. 1997;16: Michelow I, Nicol M, Tiemessen C, Chezzi C, Pettifor J. Value of cerebrospinal fluid leukocyte aggregation in distinguishing the cause of meningitis in children. Pediatr Infect Dis J. 2000;19: Bailey EM, Domenico P, Cunha BA. Bacterial or viral meningitis? Measuring CSF lactate in CSF can help you know quickly. Postgrad Med. 1990;88: Controni G, Rodriquez WJ, Hicks JM, et al. Cerebrospinal fluid lactic acid levels in meningitis. J Pediatr. 1977;91: Leib SL, Bosacci R, Gratzl O, Zimmerli W. Predictive value of cerebrospinal fluid (CSF) lactate level versus CSF/blood glucose ratio for the diagnosis of bacterial meningitis following neurosurgery. Clin Infect Dis. 1999;29: Schwarz S, Bertram M, Schwab S, Andrassy K, Hacke W. Serum procalcitonin levels in bacterial and abacterial meningitis. Crit Care Med. 2000;28: Gendrel D, Raymond J, Assicot M, et al. Measurement of procalcitonin levels in children with bacterial or viral meningitis. Clin Infect Dis. 1997;24: Viallon A, Zeni F, Lambert C, et al. High sensitivity and specificity of serum procalcitonin levels in adults with bacterial meningitis. Clin Infect Dis. 1999;28: Hansson LO, Axelsson G, Linne T, Aurelius E, Lindquist L. Serum C-reactive protein in the differential diagnosis of acute meningitis. Scand J Infect Dis. 1993;25: Lembo RM, Marchant CD. Acute phase reactants and risk of bacterial meningitis among febrile infants and children. Ann Emerg Med. 1991; 20: Dulkerian S, Kilpatrick L, Costarin AT, et al. Cytokine elevation in infants with bacterial and aseptic meningitis. J Pediatr. 1995;126: López-Cortés LF, Marquez-Arbizu R, Jimenez-Jimenez LM, et al. Cerebrospinal fluid tumor necrosis factor-alpha, interleukin-1beta, interleukin-6 and interleukin-8 as diagnostic markers of cerebrospinal fluid infection in neurosurgical patients. Crit Care Med. 2000;28: Spanos A, Harrell FE Jr, Durack DT. Differential diagnosis of acute meningitis: an analysis of the predictive value of initial observations. JAMA. 1989;262: Barone MA. Laboratory values. In: McMillan JA, DeAngelis CD, Feigin RD, Warshaw JD, eds. Oski s Pediatrics: Principles and Practice. 3rd ed. New York, NY: Lippincott Williams & Wilkins; 1999: Strikas RA, Anderson LJ, Parker RA. Temporal and geographic patterns of isolates of nonpolio enteroviruses in the United States, J Infect Dis. 1986;153: SPSS for Windows. [computer program]. Version Release Chicago, IL: SPSS, Inc; Stata Statistical Software [computer program]. Version Release 6.0. College Station, TX: StataCorp; Answer Tree [computer program]. Version Release 2.0. Chicago, IL: SPSS Inc; Brieman L, Friedman JH, Olshen RA, Stone CJ. Classification and Regression Trees. Washington, DC: Chapman & Hall; Stiell IG, Wells GA. Methologic standards for the development of clinical decision rules in emergency medicine. Ann Emerg Med. 1999;33: Black S, Shinefield H, Fireman B, et al. Efficacy, safety and immunogenicity of heptavalent pneumococcal conjugate vaccine in children. Pediatr Infect Dis J. 2000;19: Bradley JS, Ching DK, Phillips SE. Outpatient therapy of serious pediatric infections with ceftriaxone. Pediatr Infect Dis J. 1988;7: Congeni BL, Bradley JS, Hammerschlag MR. Safety and efficacy of once daily ceftriaxone for the treatment of bacterial meningitis. Pediatr Infect Dis J. 1986;5: Laupacis A, Sekar N, Stiell I. Clinical prediction rules: a review and suggested modifications of methodological standards. JAMA. 1997;277: PREDICTIVE MODEL TO DISTINGUISH BACTERIAL FROM ASEPTIC MENINGITIS

8 39. Bonadio WA, Smith DS, Goddard S, Burroughs J, Khaja C. Distinguishing cerebrospinal fluid abnormalities in children with bacterial meningitis and traumatic lumbar punctures. J Infect Dis. 1990;162: Bonadio WA. The cerebrospinal fluid: physiologic aspects and alterations associated with bacterial meningitis. Pediatr Infect Dis J. 1992;11: Eskola J, Kilpi T, Palmu A, et al. Efficacy of a pneumococcal conjugate vaccine against acute otitis media. N Engl J Med. 2001;344: Lipsitch M. Bacterial vaccines and serotype replacement: lessons from Haemophilus influenzae and prospects for Streptococcus pneumoniae. Emerg Infect Dis. 1999;5: Lipsitch M. Interpreting results from trials of pneumococcal conjugate vaccines: a statistical test for detecting vaccine-induced increases in carriage of nonvaccine serotypes. Am J Epidemiol. 2001;154: National, state and urban area vaccination coverage levels among children aged months United States, July 1994 June MMWR Morb Mortal Wkly Rep. 1996;45: Breukels MA, Spanjaard L, Sanders LA, Rijkers GT. Immunological characterization of conjugated Haemophilus influenzae type b vaccine failure in infants. Clin Infect Dis. 2001;32: Heath PT, Booy R, Azzopardi H, et al. Antibody concentration and clinical protection after Hib conjugate vaccination in the United Kingdom. JAMA. 2000;284: DOGS HELP PROTECT PEOPLE PRONE TO EPILEPTIC SEIZURES Florida Is Poised to Recognize Their Role... In Florida... the State Legislature unanimously passed a bill last week that extends to people who have seizure disorders the right to be accompanied in public places by a trained service dog... Florida may be the first state to pass such a measure... A study of patients and their seizure-alert dogs by the University of Florida s Epilepsy Clinic in 1998 determined that some dogs do detect seizures, but the scientists did not have the money to investigate whether early detection was a spontaneous reaction or trainable behavior... One nonprofit agency that trains the dogs, Canine Assistants, Inc. of Alpharetta, Georgia, says it has placed 17 Golden and Labrador Retrievers with epileptic owners in 7 years, expects to place 10 this year alone and has a waiting list for 60 dogs. Canedy D. New York Times. March 29, 2002 Noted by JFL, MD ARTICLES 719

9 Development and Validation of a Multivariable Predictive Model to Distinguish Bacterial From Aseptic Meningitis in Children in the Post- Haemophilus influenzae Era Lise E. Nigrovic, Nathan Kuppermann and Richard Malley Pediatrics 2002;110;712 DOI: /peds Updated Information & Services including high resolution figures, can be found at: /content/110/4/712.full.html References Citations Subspecialty Collections Permissions & Licensing Reprints This article cites 41 articles, 10 of which can be accessed free at: /content/110/4/712.full.html#ref-list-1 This article has been cited by 25 HighWire-hosted articles: /content/110/4/712.full.html#related-urls This article, along with others on similar topics, appears in the following collection(s): Infectious Disease /cgi/collection/infectious_diseases_sub Information about reproducing this article in parts (figures, tables) or in its entirety can be found online at: /site/misc/permissions.xhtml Information about ordering reprints can be found online: /site/misc/reprints.xhtml PEDIATRICS is the official journal of the American Academy of Pediatrics. A monthly publication, it has been published continuously since PEDIATRICS is owned, published, and trademarked by the American Academy of Pediatrics, 141 Northwest Point Boulevard, Elk Grove Village, Illinois, Copyright 2002 by the American Academy of Pediatrics. All rights reserved. Print ISSN: Online ISSN:

10 Development and Validation of a Multivariable Predictive Model to Distinguish Bacterial From Aseptic Meningitis in Children in the Post- Haemophilus influenzae Era Lise E. Nigrovic, Nathan Kuppermann and Richard Malley Pediatrics 2002;110;712 DOI: /peds The online version of this article, along with updated information and services, is located on the World Wide Web at: /content/110/4/712.full.html PEDIATRICS is the official journal of the American Academy of Pediatrics. A monthly publication, it has been published continuously since PEDIATRICS is owned, published, and trademarked by the American Academy of Pediatrics, 141 Northwest Point Boulevard, Elk Grove Village, Illinois, Copyright 2002 by the American Academy of Pediatrics. All rights reserved. Print ISSN: Online ISSN: