Identifying Low-risk Patients for Bacterial Meningitis in Adult Patients with Acute Meningitis

Transcription

1 ORIGINAL ARTICLE Identifying Low-risk Patients for Bacterial Meningitis in Adult Patients with Acute Meningitis Yasuharu Tokuda 1, Masahiro Koizumi 2, Gerald H. Stein 3,4 and Richard B. Birrer 5 Abstract Objective To derive and validate a clinical prediction model with high sensitivity for differentiating aseptic meningitis (AM) patients from bacterial meningitis (BM) patients. Methods We developed the model using the derivation cohort in a community rural hospital in Okinawa and assessed its performance using the validation cohort in a metropolitan urban hospital in Tokyo. There were 66 (39.5%) and 5 (17.9%) adult patients with BM among the derivation (n=167) and the validation cohort (n=28), respectively. Recursive partitioning analysis was used to determine the important classification variables and to develop a sensitive model to safely exclude BM. Results The model produced high- and low-risk groups based on the following: 1) Gram stain, 2) CSF neutrophil percent!15%, 3) CSF neutrophil count!150 cells/mm 3, and, 4) mental status change. Among the derivation cohort, there were 65 patients with BM in the high-risk group (n=76), while only one patient with BM was noted (sensitivity, 99%) in the low-risk group (n=91). Among the validation cohort, there were 5 patients with BM in the high-risk group (n=7), while no patient was classified with BM (sensitivity, 100%) in the low-risk group (n=21). Conclusion This simple and sensitive model might be useful to safely identify low-risk patients for BM who would not require antibiotic treatment. Key words: central nervous system bacterial infections, aseptic meningitis, cerebrospinal fluid, clinical prediction rule, recursive partitioning analysis (Inter Med 48: , 2009) () Introduction In cases of acute meningitis, immediate diagnosis and treatment is essential for patients with bacterial meningitis (BM). A delay in starting appropriate antibiotics may worsen the prognosis in patients with BM (1-3). On the other hand, treatment for aseptic meningitis (AM) is generally supportive and does not require antibiotics, since it is typically caused by viral infection (4, 5). Because the initial symptoms and signs of BM are frequently similar to those of AM, laboratory examination is always required. Above all, direct examination of cerebrospinal fluid (CSF) using Gram stain is mandatory for rapid confirmation of bacterial pathogen. However, in many patients with BM bacteria in CSF is not revealed in Gram stain. The conventional CSF examinations, such as CSF total leukocyte count, CSF-blood glucose ratio or others, are not accurate enough to differentiate BM from AM (6, 7). Moreover, the newly developed laboratory tests for the diagnosis of BM, such as CSF procalcitonin and bacterial polymerase chain reaction (PCR) test, are still not suitable for widespread clinical application due to their non-immediate results and high costs (8-12). Therefore, the current guidelines recommend starting empiric antibiotics for patients with acute meningitis whenever bacterial etiology is at least possible (13, 14). In addition, early use of corticosteroids may be beneficial for improving Center for Clinical Epidemiology, St. Luke s Life Science Institute, Tokyo, Department of Medicine, Okinawa Chubu Hospital, Okinawa, Department of Medicine, University of Florida School of Medicine, Gainesville, FL, USA, Department of Medicine, University of Hawaii John A. Burns School of Medicine, Honolulu, HI, USA and Department of Medicine, Cornell University School of Medicine, New York, NY, USA Received for publication November 1, 2008; Accepted for publication December 26, 2008 Correspondence to Dr. Yasuharu Tokuda, tokuyasu@orange.ocn.ne.jp 537

2 outcome in patients with BM (15), although a recent study did not confirm any benefit in adult BM patients in sub- Saharan Africa (16). Thus, physicians must determine whether or not to use steroids for patients with acute meningitis at initial presentation. Several studies have developed clinical prediction models for differentiating BM from AM based on multivariable logistic regression modeling (17-22). However, since physicians need to perform algebraic calculation at the bedside in order to apply these models, they are not eager to use these prediction rules (23). In fact, several authors have reported the infrequent clinical use of clinical prediction rules due to the need for complicated calculation and the low sensitivity (24, 25). In contrast, simple and sensitive prediction models, for instance the rule for ankle injuries (the Ottawa Ankle Rule), are popular and being used in widespread clinical practice (26). Thus, the simpler the prediction model, the more likely physicians are to remember and use it (23). Thus, the aim of this study was to develop a simple and sensitive prediction model for BM in adults with acute meningitis. We used a recursive partitioning analysis, since this statistical technique is well suited to develop a sensitive and simple classification model (27, 28). We developed our model using a derivation cohort in a community rural hospital in Okinawa, Japan and validated the model using a validation cohort in a metropolitan urban hospital in Tokyo, Japan. Study participants Methods We investigated consecutive adult patients (aged 16 years or older) with community-acquired acute meningitis presenting to the emergency departments (ED) at two teaching hospitals, Okinawa Chubu Hospital and St. Luke s International Hospital, to derivate and validate the clinical prediction model, respectively. The hospital for derivation cohort (between 1990 and 2000) is a 550-bed hospital that serves the rural community with population of approximately 400,000 in Okinawa, Japan, and provides primary and specialty care. The ED in this hospital receives approximately 50,000 annual outpatient visits. The hospital for validation cohort (between 2003 and 2007) is a 520-bed hospital that serves the metropolitan community with population of about 300,000 in Tokyo, Japan, and similarly it provides primary and specialty care. The ED in this hospital receives about 40,000 annual outpatient visits. The approvals from the hospital ethics committees were obtained. For these two hospitals, we enrolled patients with community-acquired acute infectious meningitis. The three inclusion criteria were: 1) acute onset with intervals!10 days from symptom to presentation to ED; 2) symptoms consistent with acute meningitis (fever, chills, headache, nausea, or vomiting); and, 3) CSF total leucocytes of!7 cells/mm 3 or presence of bacteria on CSF Gram stain. We excluded patients with: 1) onset with intervals >10 days from symptom to presentation to ED; 2) hospital-acquired meningitis; 3) post-neurosurgery or post-traumatic meningitis; 4) encephalitis; 5) fungal meningitis; 6) tuberculous meningitis; 7) meningitis carcinomatosa; 8) diagnosis of multiple sclerosis, Guillain-Barre syndrome, or central nervous system vasculitis; or, 9) patients with no CSF examination. As the purpose of this study was to identify low-risk patients for a diagnosis of BM among adults with acute meningitis, our methodology was based on previous similar studies (17, 19, 22) and thus we excluded those with encephalitis, fungal meningitis, or tuberculous meningitis. Therefore, we did not intend to differentiate these important pathologic conditions for meningitis or encephalitis in this study, It is important to note that careful attention should be paid to the possibility of these conditions among patients initially suspected as having AM or BM. Data collection Data collection involved clinical history, physical examination, and laboratory data. Clinical history included demographics. Data for physical examination included mental status change and nuchal rigidity. Presence of mental status change was indicated by the Glasgow coma scale!14. We used data for nuchal rigidity as present or absent which was assessed by ED physicians at the time of the initial evaluation. Laboratory data included peripheral white blood cells (WBC), blood cultures (two sets), and serum glucose concentration and CFS examinations, including bacterial identification on Gram stain, total leukocyte count, neutrophil count, neutrophil percent (%), protein, glucose, and CSF/serum glucose concentration ratio. Data for physical examination and laboratory examination at the initial workup at the emergency departments or outpatient units were collected for our analysis. All patients with acute meningitis underwent CSF Gram stain test. BM was defined: 1) if a bacterial pathogen was identified in CSF or blood; 2) if a significant concentration of bacterial antigen in blood, urine or CSF was obtained; 3) if Leptospira spp were identified in blood, urine or CSF; 4) if BM was considered as the likely cause based on infectious disease specialist consultation. All other patients were given a diagnosis of AM. Viral cultures were not performed in these two hospitals during the study period. Statistical analysis We compared clinical and laboratory data between patients with BM and AM in the derivation and validation cohorts. Fisher s exact test or Student t-test was used to compare categorical or continuous variables, where appropriate. P-values were based on a 2-sided test and considered statistically significant at <0.05. General statistical analyses were performed with SPSS version 15.0J (SPSS-Japan, Tokyo, Japan). The prediction model was developed based on the deriva- 538

3 Table1. ClinicalCharacteristicsofDerivationandValidationCohorts Characteristic Derivation cohort (n=167) Validation cohort (n=28) Bacterial Aseptic P-value Bacterial Aseptic p-value (n=66) (n=101) (n=5) (n=23) Mean age, yr (SD) 56.9 (17.5) 31.8 (10.4) < (22.3) 40.4 (15.3) Male gender, n (%) 36 (54.5) 63 (62.4) (60) 10 (43.5) Mental status change, n (%) 45 (68.2) 0 (0) < (80) 1 (4.3) Nuchal rigidity, n (%) 54 (81.8) 31 (30.7) < (80) 8 (34.8) Mean Peripheral WBC count, /mm3 (SD) (7016) 8903 (4062) < (12854) 8443 (3474) SD=standard deviation. WBC=white blood cell. P-values for comaprison between patients with bacterial meningitis and those without aseptic meningitis by Fisher's exact test for nominal variables and by t-test for continuous variables. tion cohort and evaluated using the validation cohort. Recursive partitioning analysis was used to develop the model. This analysis sequentially divides the range of each variable to obtain an optimal binary split into two subgroups (daughter nodes). Gini criteria was used as the node splitting rule to reduce impurity of proportions of patients with BM and those with AM within each daughter node (29). The resulting sequence of trees was pruned by 10-fold cross-validation procedure. We grouped the several terminal nodes into two risk levels (high- and low-risk groups) based on the relative prevalence of BM within each terminal node. Since our goal was to generate a prediction model that would exclude BM at reasonable sensitivity, we set the misclassification cost for labeling bacterial meningitis as AM at 15 times higher than for labeling aseptic meningitis as bacterial meningitis. Varying misclassification costs in this manner tends to obtain a final model with high sensitivity. We assessed the classification characteristics of the final model by estimating prevalence of BM, positive and negative predictive value, as well as sensitivity, specificity, and positive and negative likelihood ratios among high risk and low risk groups, respectively. Recursive partitioning analysis was performed with CART version 5.0 (Salford, San Diego, California). Results We analyzed the derivation cohort of 167 patients with acute meningitis (101 with AM and 66 with BM) to develop a clinical prediction model. Next, we assessed performance of the model using the validation cohort of 28 patients with acute meningitis (23 with AM and 5 with BM) as an external validation study. The frequently recognized causative bacteria (n>1) among BM in the derivation cohort included Streptococcus pneumoniae (n=18), Escherichia coli (n=6), Neisseria meningitidis (n=5), Listeria monocytogenes (n=3), Klebsiella pneumoniae (n=2), and Staphylococcus aureus (n=2). Listeria monocytogenes meningitis was identified in three patients with connective tissue disorders receiving long-term steroid therapy. The causative bacteria among BM in the validation cohort were Streptococcus pneumoniae (n=4) and Corynebacterium (n=1). The patient with Corynebacterium meningitis (23- year-old woman) had received cytotoxic chemotherapy for acute lymphoblastic leukemia. Table 1 shows the clinical characteristics between BM and AM both in the derivation and validation cohorts. The characteristics significantly associated with BM among the derivation cohort included older age, mental status change, and nuchal rigidity, higher blood leukocyte count. Among the validation cohort, mental status change was significantly associated with BM. Among the derivation cohort, sensitivity of mental status change for BM was 68% and its specificity was 100%. Among the validation cohort, its sensitivity was 80% and its specificity was 96%. Table 2 shows CSF examinations between BM vs. AM both in the derivation and validation cohorts. Total leukocyte count, neutrophil count, neutrophil percent, protein, glucose, and CSF/blood glucose ratio were significantly different between patients with BM and those with AM both in the derivation and validation cohorts. Direct Gram stain was positive in 27 patients (40.9% of those with BM) of the derivation cohort and in two patients (40% of those with BM) of the validation cohort. In five patients with meningococcal meningitis, the range of CSF neutrophil count was between 30 and 8,160 cells/mm 3 (mean 2,849). There was no statistical significance of CSF neutrophil count between patients with meningococcal meningitis and with other BM (P=0.98). Recursive partitioning analysis selected the variables for the best performance of classifying BM and AM. The most important variables were: CSF Gram stain, CSF neutrophil percent, and CSF neutrophil count. Figure 1 shows the optimal tree generated by this analysis. The tree had four splits which produced five terminal nodes. The terminal nodes were grouped into high- and low-risk groups based on the relative prevalence of BM. The high-risk group included: those with positive bacteria on CSF Gram stain (Terminal Node 1); those with negative Gram stain, CSF neutrophil percent >15%, and CSF neutrophil >150/mm 3 (Terminal Node 2); those with negative Gram stain, CSF neutrophil proportion >15%, CSF neutrophil!150/mm 3, and mental status change (Terminal Node 3). The low-risk group included: those with negative Gram stain and CSF neutrophil percent!15% (Terminal Node 4) and those with negative 539

4 Table2. CSFExaminationsofDerivationandValidationCohorts Characteristic Derivation cohort (n=167) Validation cohort (n=28) Bacterial Aseptic P-value Bacterial Aseptic p-value (n=66) (n=101) (n=5) (n=23) Positive Gram stain, n (%) 27 (40.9) 0 (0) < (40) 0 (0) Mean total leukocyte count, /mm3 (SD) 6957 (14692) 184 (277) (16059) 438 (376) <0.001 Mean neutrophil count, /mm3 (SD) 4509 (13279) 78 (183) (15243) 52 (78) <0.001 Mean neutrophil percent, % (SD) 84.8 (17.2) 39.5 (33.7) < (41.1) 15.4 (25.4) Mean protein concentration, mg/dl (SD) 345 (465) 89 (55) < (393) 75 (43) <0.001 Mean glucose concentration, mg/dl (SD) 45.6 (39.0) 58.9 (10.8) (18.2) 55.5 (12.3) <0.001 Glucose ratio (CSF/serum), (SD) (0.218) (0.108) < (0.043) (0.111) <0.001 SD=standard deviation. CSF=cerebrospinal fluid. P-values for comaprison between patients with bacterial meningitis and those without aseptic meningitis by Fisher's exact test for nominal variables and by t-test for continuous variables. Bacteria on CSF Gram stain Positive Negative High-Risk (TN1) CSF Neutrophil Percent (%) >15% <=15% CSF Neutrophil Count (cells/mm3) Low-risk (TN4) >150/mm3 High-Risk (TN2) <=150 /mm3 Mental Status Change Positive Negative High-Risk (TN3) Low-risk (TN5) Figure1. Theclinicalpredictionmodelforbacterialvs.asepticmeningitisinadultpatientswith acutemeningitis.positivementalstatuschangeisdefined GlasgowComaScaleof14. TD:terminalnode,CSF:cerebrospinalfluid Gram stain, CSF neutrophil proportion >15%, CSF neutrophil!150 /mm 3, and no mental status change (Terminal Node 5). Table 3 shows performance of the model for BM vs. AM in the derivation cohort. Ninety-one patients (54.5%) were classified into the low-risk group and 76 (45.5%) into the high-risk group. The misclassification rate in the low-risk group was 1/91 (1%). The patient, who was classified in terminal node 5 of the low-risk group but was identified as having BM, was a 44-year-old woman with leptospirosis but no mental status change. Her CSF examinations showed negative Gram stain, neutrophil percent of 19%, and CSF neutrophil count of 71/mm 3. In the high-risk group, there were 65 patients with BM (85.5%). Thus, the positive predictive value of 86% and negative predictive value of 99% were obtained along with sensitivity of 99% and specificity of 89% in the derivation cohort. After application of the constructed model to the validation cohort, performance of the model for BM vs. AM between the high-risk and the low-risk groups were observed as follows. Among the validation cohort, 21 (75%) were classified into the low-risk group and seven (25%) into the high-risk group. The misclassification rate in the low-risk group was 0/91 or 0%. In the high-risk group, there were five patients with BM (71%). Thus, in the validation cohort, we obtained a positive predictive value of 71% and a negative predictive value of 100%, along with sensitivity of 100% and specificity of 91%. Overall, the model was simple in its structure and showed high sensitivity. The moderate specificity and positive predictive values were likely to 540

5 Table3. PerformanceoftheModelforBacterialvs.AsepticMen ingitis Characteristic Value Derivation cohort (n=167) Prevalence of bacterial meningitis, (%) High risk group n=76 Terminal Node 1 26/26 (100) Terminal Node 2 35/46 (76) Terminal Node 3 4/4 (100) Low risk group n=91 Terminal Node 4 0/41 (0) Terminal Node 5 1/50 (2) Positive predictive value of high risk group, % 65/76 (86) Negative predictive value of low risk group, % 90/91 (99) Sensitivity, % 99 Specificity, % 89 Positive likelihood ratio 9.04 Negative likelihood ratio 0.02 Validation cohort (n=28) Prevalence of bacterial meningitis, (%) High risk group n=7 Terminal Node 1 2/2 (100) Terminal Node 2 3/4 (75) Terminal Node 3 0/1 (0) Low risk group n=21 Terminal Node 4 0/17 (0) Terminal Node 5 0/4 (0) Positive predictive value of high risk group, % 5/7 (71) Negative predictive value of low risk group, % 0/21 (100) Sensitivity, % 100 Specificity, % 91 Positive likelihood ratio Negative likelihood ratio 0 be the trade-offs for high sensitivity and high negative predictive value. Discussion This study derived and validated the clinical prediction model based on a recursive partitioning analysis to maximize the model sensitivity for safely identifying patients with low-risk for BM. The important classification variables were CSF Gram stain, CSF neutrophil percent, CFS absolute neutrophil counts, and mental status change. In the classification performance, our simple prediction algorithm was highly sensitive (99% for the derivation cohort and 100% for the validation cohort). The moderate degree of specificity (89% for the derivation cohort and 91% for the validation cohort) was the trade-off for obtaining high sensitivity with the model. Use of this model for predicting BM might improve diagnostic sensitivity for suspicion of BM and could guide rational decision making for initiating empiric antibiotics and/ or corticosteroids. Nevertheless, on the grounds of the importance of physicians clinical judgment, all available clinical information should be also considered for these critical decisions. In addition, impact analysis using the actual application of this model is needed to confirm its good performance. The accurate diagnosis of BM is often difficult. Physicians frequently face diagnostic uncertainty at the time of initial evaluation for patients with acute meningitis in ED. In most cases with acute meningitis, physicians need to initiate empiric treatment with a broad spectrum antibiotic until the CSF cultures prove negative, since missed antibiotics therapy for BM would lead to serious consequences including mortality. Therefore, a clinical prediction model for BM must be highly sensitive for not labeling those with BM into the low-risk group. In addition, prediction models should be simple, since most physicians are likely to follow only simple, but still sensitive, models (23, 31). Thus, the high sensitivity and its simple structure of our model seems appropriate for widespread practical use (30). Although many variables were significantly different between patients with BM and those with AM, recursive partitioning analysis selected the following important variables for splitting into binary nodes with low- and high-risk groups; CSF Gram stain, CSF neutrophil percent, CSF absolute counts, and mental status change. First, all patients with positive Gram stain were consequently identified as BM. One advantage of this test is that we can notice the morphological identity of bacteria for guiding antibiotic choice; for instance, when we would see lancet-formed Gram positive 541

6 diplococci in the CSF Gram stain, this finding would indicate Pneumococcal, or Strep. pneumoniae, meningitis. Second, the increase in neutrophil percentage, as well as in absolute number of neutrophils, in CSF examinations is also important determinant. Thus, microscopic examination for the differentiation of CSF leukocytes should be available, regardless of day or night and weekday or weekend, for more accurate diagnosis of patients with acute meningitis. Lastly, but not least, mental status change in patients with acute meningitis suggests bacterial etiology. If there are patients with suspected BM and mental confusion or agitation, lumbar puncture to obtain CSF fluid should be safely performed. The most common bacteria isolated from our patients with BM were Streptococcus pneumoniae, followed by Escherichia coli and Neisseria meningitides. Patients with meningococcal meningitis are known to have a lower number of neutrophils in the CSF early in the disease process even with signs and symptoms of meningitis. However, in five patients with meningococcal meningitis of our study, the range of their CNS leukocyte count was similar to that of patients with other BM. In fact, the CNS neutrophil percent in all these patients was >15%. Consequently, these patients were correctly classified into the high risk group. There are several limitations in this study. First, although we conducted an external validation study evaluating our model in a different hospital, this model requires prospective validation in independent settings before widespread practical application. For its generalizability to other settings, validation studies are needed in other countries, with different background populations and different patient spectrum. The varying prevalence effect of HIV infection, as a risk factor for opportunistic infection such as Listeriosis, may show possible performance variation. Moreover, it is preferable to conduct an impact analysis for evaluating its safety profile and performance in actual settings (33). Second, the precise etiology in most patients with AM was unclear. In previous studies, viral meningitis is the most frequent cause for aseptic meningitis, although Mycoplasma spp, drug reactions, or other conditions may result in the similar syndrome (32). However, it was frequently difficult to differentiate the individual etiology of AM and the physicians in our institutions generally did not test for specific viruses for most patients with AM. Third, we did not include the new laboratory technology such as CSF procalcitonin or bacterial polymerase chain reaction tests. If these tests become available for clinical practice, we might need to redevelop a prediction model by incorporating these new tests. However, the usefulness of these tests remains to be determined. Fourth, we could not obtain data for prior use of antimicrobials before visiting our hospitals. There might have been a few patients with BM who were classified as AM due to no growth of bacterial pathogen in CSF and blood. However, as we reviewed the comments about the etiology by infectious disease specialist physicians in the medical records of all patients, we thought that a clinically reasonable analysis was conducted for the classification between AM and BM. Fifth, our methodology was based on the previous studies which developed clinical prediction models for acute meningitis of Spanos et al (17), Hoen et al (19), and Brivet et al (22), by using only data from patients with AM or BM; we did not collect data for meningitis patients with these etiologies such as tuberculosis or fungus. Tuberculous or fungal meningitis can sometimes be indistinguishable from AM and BM when they show acute onset with classical symptoms and without other neurological complications. Further studies are needed to predict these etiologies in patients with acute meningitis. Sixth, we excluded patients with encephalitis from our cohorts. Patients with mental status change and mild CSF pleocytosis might have resulted from encephalitis. Identifying these patients is important, since specific antiviral agents, such as acyclovir, should be started when this possibility is considered. Thus, physicians need to pay particular attention to the possibility of encephalitis in those with AM who develop mental status change. Finally, our two study cohorts had a different prevalence of BM. There might be a difference in terms of geographical settings throughout Japan. In particular, patients with BM related to Strongyloides stercoralis infection is more likely to be seen in subtropical regions such as Okinawa. This endemicity of parasitic infection may be a reason for the different prevalence between two hospitals. Positive predictive value would become lower in the settings where its prevalence would be lower. However, in the validation cohort, the model performance also showed a high sensitivity. Thus, high sensitivity of the model in the derivation cohort was confirmed, which is the most important characteristic for this model. In summary, our clinical prediction model could be used to help identify low-risk patients for BM at the initial evaluation for adult patients with acute meningitis. This simple, sensitive, and tree-based classification model performs with high sensitivity. In particular, patients with negative Gram stain and CSF neutrophil percent!15% or those with CSF neutrophil count!150/mm 3 with normal mental status are less likely to have BM. This simple tree-based classification algorithm may need prospective validation and impact studies before application to clinical practice. Acknowledgement We thank all attending physicians, residents, and nurses of the emergency departments in our hospitals for their care of adult patients with acute meningitis. 542

7 References 1. Mace SE. Acute bacterial meningitis. Emerg Med Clin North Am 26: , Sinner SW, Tunkel AR. Antimicrobial agents in the treatment of bacterial meningitis. Infect Dis Clin North Am 18: , Flores-Cordero JM, Amaya-Villar R, Rincon-Ferrari MD, et al. Acute community-acquired bacterial meningitis in adults admitted to the intensive care unit: clinical manifestations, management and prognostic factors. Intensive Care Med 29: , Hausfater P, Fillet AM, Rozenberg F, et al. Prevalence of viral infection markers by polymerase chain reaction amplification and interferon-alpha measurements among patients undergoing lumbar puncture in an emergency department. J Med Virol 73: , Nowak DA, Boehmer R, Fuchs HH. A retrospective clinical, laboratory and outcome analysis in 43 cases of acute aseptic meningitis. Eur J Neurol 10: , Tunkel AR, Scheld WM. Acute bacterial meningitis. Lancet 346: , Seehusen DA, Reeves MM, Fomin DA. Cerebrospinal fluid analysis. Am Fam Physician 68: , Jereb M, Muzlovic I, Hojker S, Strle F. Predictive value of serum and cerebrospinal fluid procalcitonin levels for the diagnosis of bacterial meningitis. Infection 29: , Chadwick DR, Lever AM. The impact of new diagnostic methodologies in the management of meningitis in adults at a teaching hospital. QJM 95: , Inada K, Takahashi K, Ichinohe S, et al. A silkworm larvae plasma test for detecting peptidoglycan in cerebrospinal fluid is useful for the diagnosis of bacterial meningitis. Microbiol Immunol 47: , Saravolatz LD, Manzor O, VanderVelde N, Pawlak J, Belian B. Broad-range bacterial polymerase chain reaction for early detection of bacterial meningitis. Clin Infect Dis 36: 40-45, Schuurman T, de Boer RF, Kooistra-Smid AM, van Zwet AA. Prospective study of use of PCR amplification and sequencing of 16S ribosomal DNA from cerebrospinal fluid for diagnosis of bacterial meningitis in a clinical setting. J Clin Microbiol 42: , Fitch MT, Abrahamian FM, Moran GJ, Talan DA. Emergency department management of meningitis and encephalitis. Infect Dis Clin North Am 22: 33-52, Talan DA, Zibulewsky J. Relationship of clinical presentation to time to antibiotics for the emergency department management of suspected bacterial meningitis. Ann Emerg Med 22: , van de Beek D, de Gans J, McIntyre P, Prasad K. Corticosteroids for acute bacterial meningitis. Cochrane Database Syst Rev 1: , Scarborough M, Gordon SB, Whitty CJ, et al. Corticosteroids for bacterial meningitis in adults in sub-saharan Africa. N Engl J Med 357: , Spanos A, Harrell FE Jr, Durack DT. Differential diagnosis of acute meningitis. An analysis of the predictive value of initial observations. JAMA 262: , McKinney WP, Heudebert GR, Harper SA, Young MJ, McIntire DD. Validation of a clinical prediction rule for the differential diagnosis of acute meningitis. J Gen Intern Med 9: 8-12, Hoen B, Viel JF, Paquot C, Gerard A, Canton P. Multivariate approach to differential diagnosis of acute meningitis. Eur J Clin Microbiol Infect Dis 14: , Leblebicioglu H, Esen S, Bedir A, Gunaydin M, Sanic A. The validity of Spanos and Hoen s models for differential diagnosis of meningitis. Eur J Clin Microbiol Infect Dis 15: , 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 19: , Brivet FG, Ducuing S, Jacobs F, et al. Accuracy of clinical presentation for differentiating bacterial from viral meningitis in adults: a multivariate approach. Intensive Care Med 31: , Katz MH. Multivariable analysis: a practical guide for clinicians. Cambridge University Press, New York, 1999: Pearson SD, Goldman L, Garcia TB, Cook EF, Lee TH. Physician response to a prediction rule for the triage of emergency department patients with chest pain. J Gen Intern Med 9: , Wasson JH, Sox HC. Clinical prediction rules. Have they come of age? JAMA 275: , Stiell I, Wells G, Laupacis A, et al. Multicentre trial to introduce the Ottawa ankle rules for use of radiography in acute ankle injuries. Multicentre Ankle Rule Study Group. BMJ 311: , Breiman L. Classification and regression trees. Wadsworth International Group: The Wadsworth statistics/probability series. Chapman & Hall, London, Zhang H, Singer B. Recursive partitioning in the health sciences. Statistics for biology and health. Springer, New York, Strobl C, Boulesteix AL, Augustin T. Unbiased split selection for classification trees based on the Gini Index. Computational Statistics and Data Analysis 52: , Wasson JH, Sox HC, Neff RK, Goldman L. Clinical prediction rules. Applications and methodological standards. N Engl J Med 313: , Laupacis A, Sekar N, Stiell IG. Clinical prediction rules. A review and suggested modifications of methodological standards. JAMA 277: , Reilly BM, Evans AT. Translating clinical research into clinical practice: impact of using prediction rules to make decisions. Ann Intern Med 144: , Ropper AH, Adams RD, Victor M, Brown RH, Victor M. Adams and Victor s Principles of Neurology. 8th ed. McGraw-Hill, New York, The Japanese Society of Internal Medicine 543