Article 267Vol. 14. No. 4

Eur. J. Clin. Microbiol. Infect. Dis., 1995, 14: 267-274

Multivariate Approach to Differential Diagnosis ofAcute Meningitis

B. Hoen1*, J.F. Vie12, C. paquot1, A. Gérard1, P. Canton:

A previously reported statistical model based on a combination of four parameters(total polymorphonuclear cell count in cerebrospinal fluid (CSF), CSF/blood glucoseratio, age and month of onset) appeared effective in differentiating acute viralmeningitis (AVM) from acute bacterial meningitis (ABM). The objectives of this studywere to validate this model on a large independent sample of patients with acutemeningitis and to build and validate a new model based on this sample. Of 500 con-secutive cases of community-acquired meningitis reviewed retrospectively, 115 wereABM, 283 were AVM and 102 were of uncertain etiology. For each of the ABM andAVM cases, the probability of ABM versus AVM (pABM) was calculated for both mod-els. Sensitivity, specificity and predictive values as weil as areas under the receiveroperating characteristic (ROC) curves were calculated for both models. The originalmodel proved an accu rate and reliable diagnostic test. Its area under the ROC curvewas 0.981. For pABM = 0.1, its negative and positive predictive values were 0.99 and0.68, respectively. The new model retained four slightly different independent vari-ables: CSF protein level, total CSF polymorphonuclear cell count, blood glucoselevel and leukocyte cou nt. Its area under the ROC curve was 0.991 and, forpABM = 0.1 , its negative and positive predictive values were 0.99 and 0.85, respec-tively. ln conclusion, both models provide a valuable aid in differentiating AVM fromABM. They should be further evaluated in a prospective appraisal of thei r contributionto therapeutic decision making.

Accurate and rapid diagnosis is essential forpatients with acute bacterial meningitis (ABM).Although examination of the œrebrospinal fluid(CSF) often provides immediate confirmation ofABM, it sometimes fails to differentiate betweenABM and acute viral meningitis (A VM). A posi-tive Grani-stained CSF smear is virtually 100 %diagnostic of ABM, but in about 25 % of cases ofABM, the initial Gram-stained CSF smear isnegative (1). For years, in cases of difficult differ-ential diagnosis between ABM and A VM, physi-cians have preferred to treat viral meningitis withantibiotics rather than Dot to treat a bacterial one.However, the cost of antibiotic therapy and its at-tendant hospitalization as weIl as its potential side

effects.. raised concern about giving unnecessaryantibiotics in A VM. Various blood and CSF find-ings, such as CSF lactate and blood or CSF C-reactive protein, have been proposed as markeTSfor differential diagnosis between ABM andA VM (2-5). Unfortunately, no single value forany CSF or blood parameter has been found to bediscriminant, since for each potential parameterthe distribution for ABM may overlap the entirerange of values found in AVM (5-7). Thisprompted Spanos et al. (6) to undertake a retro-spective study of a large number of bath ABMand A VM cases with the aim of assessing whetherthe final diagnosis could be predicted accuratelyusing a statistical model based on a combinationof parameters. They identified four independentparameters for predicting the likelihood of ABM:total polymorphonuclear leukocyte (PMN) countin CSF, CSF/blood glucose ratio, age and thenumber of months from August 1 at the time ofthe onset of meningitis. ln any case of meningitis,

1 Département de Maladies Infectieuses et Tropicales,Centre Hospitalier Universitaire de Nancy, Hôpitaux deBrabois, 54511 Vandoeuvre Cedex, France.

2 Département de Santé Publique, Faculté de Médecine, 2place Saint-Jacques, 25030 Besançon, France.

Eur. J. Clin. Microbiol. Infect. Dis.268

the probability of ABM versus A VM (pABM)could either be calculated according to the logis-tic model equation or estimated from a nomo-gram derived from this model, given the value ofthe four diagnostic parameters (6). The completemathematical expression of Spanos' model was asfollows:

l.pABM = "L,where(1 + e- )

L = 0.52 x number of months from August 1-12.76 x CSF-blood glucose ratio (if ratio

exœeds 0.6 use 0.6)+ 0.341 x (PMNs in CSF x 106/1)°.333+ 2.29 x age + 2.79 (if age ~ 1 year),-2.71 x age +7.79 (if 1 year < age ~2years),-0.159 xage+ 2.69 (if 2 yeaTS < age~22 yeaTS) or+ 0.100 x age - 3.01 (if age > 22 years).

If the Gram stain was positive, a probability of0.99 was assumed.

The aims of our study were to validate the modelof Spanos et al. (6) on a new sample of patientswith meningitis, as suggested by these authorswho urged that their model should be validated inother hospitals; and to perform an independentmultivariate logistic regression analysis with anobjective of determining the most accurate com-bination of parameters for predicting the likeli-hood ofbacterial meningitis.

Materials and Methods

for the treatment of meningitis. Cases that could be cate-gorized neither as ABM nor as AVM according to thesedefinitions were considered of uncertain etiology andwere neither considered for Spanos' model validationnor for building our model.From the initial database that contained 500 cases ofmeningitis, 102 cases (20.4 %) were categorized asmeningitis of uncertain etiology. CSF Gram stain wasnegative in ail these cases. ln-bouse antibiotics had beengiven prior to lumbar puncture to 49 % of these 102patients as compared to 27 % of ABM cases and 21 %of AVM cases (p < 0.0001). Of the 398 remaining cases,115 were ABM and 283 were A VM. The two groups werecompared for their main characteristics using chi-square(X2) and Mann-Whitney rank-suffi tests for qualitativeand quantitative variables, respectively.For each case of meningitis, we calculated the probabilityof ABM versus AVM (pABM) according to Spanos'model. For the validation of this model, we calculatedthe sensitivity, specificity, positive (PPV) and negative(NPV) predictive values of the model for different cut-offpoints of pABM. Results of these calculations were dis-played on a receiver operating characteristics curve(ROC curve) (8) whose area under the curve (AUC)was then calculated according to the rank-suffi testmethod (9). We then performed a stepwise regressionanalysis (entering and removing limits equal to 0.10 and0.05, respectively) on the whole sample of cases since nodata were missing for any of the 398 cases (10). Theanalysis was based on the following variables: age, monthof onset, blood leukocyte count, blood glucose level, CSFglucose level, CSF/blood glucose ratio, CSF protein level,total and differential CSF leukocyte count and total CSFpolymorphonuclear cell count. As did Spanos et al. (6),we entered the variable month as the number of monthsfrom the estimated peak incidence of AVM (August 1)and the glucose ratio as follows: (CSF glucose +0.1 mmol/l)/(blood glucose + 0.1 mmol/l). Total CSFleukocyte count and total CSF polymorphonuclear countwere entered into the model in a slightly modified form([integer[CSF-cell count/100]+0.5]x100) in order todecrease the number of covariate patterns for these twovariables. As a result, ail values within the same hundredwere assigned the me an value of the hundred. For ex-ample, CSF leukocyte count values of 230 x 106/1 and270 x 106/1 were both assigned the mean value 250 x 1Q6/l.We did not enter the variable race because ail but threepatients were Caucasians.As we did for the validation of Spanos' model, we cal-culated sensitivity, specificity, PPV and NPV of our 10-gistic model for varions cutpoints of pABM and built aROC curve whose AUC was calculated and comparedto AUC of Spanos' model (11). Finally we assessed theaccuracy ofboth Spanos' and our model in two subgroupsof ABM, those with CSF positive on direct Gram stainand those with CSF negative on direct Gram stain.Ali the computations were performed using softwarefrom BMDP Statistical Software, USA.

We reviewed the charts of ail patients with a final diag-nosis of acute meningitis hospitalized in the Departmentof Infectious Diseases of the U niversi ty of Nancy MedicalCenter, Nancy, France, between August 1983 and Decem-ber 1991. Cases of meningitis were sought out by theçoded diagnoses assigned to each patient file using acomputer-assisted research. Like Spanos et al. (6), weselected cases of community-acquired acute meningitisthat occurred in patients older than one mon th. Casessecondary to neurosurgical procedures, cases of neoplas-tic or tuberculous meningitis and those with missing basicCSF (cell count, protein and glucose) data and bloodparameter data needed for multivariate analysis wereexcluded a priori. Conversely, we retained cases for whichoutpatient antibiotic treatment was given prioNo admis-sion, both to fit Spanos' selection criteria and becausethis is a common feature at presentation. We also usedthe same diagnostic criteria as Spanos.Cases were categorized as ABM if an appropriate etiologicbacterium was cultured from CSF or blood or if bacterialantigen was demonstrated in patients' CSF, blood or urineby latex agglutination. Cases were categorized as AVMif a virus was isolated from culture (on human diploidfibroblasts) of blood, stool or CSF or if the dischargediagnosis was viral meningitis and no etiology other thanviral infection was found and no antibiotics were given

Results

Patients' characteristics and blood and CSF find-ings for bath ABM and A VM cases are presentedand compared in Table 1. For an studied parame-

269Vol. 14,1995

Table 1: Patients' characteristics and blood and CSF findings in cases of bacterial and viral menin-gitis. Quantitative variables are expressed as mean with standard deviation and range in brackets.

P value'Bacterial meningitis(n = 115)

Viral meningitis(n = 283)

33.7:1: 23.2 (0.1-83)59/56

19.9:1:10.1 (4.5-52.7)16.3:1: 9.1 (3.9-50.9)82.7:1: 13.9 (10.0-97.0)

9.4 :1: 4.8 (2.4-33.0)4990 :1: 5000 (2.0-30000)4750:1: 5026 (0-29700)83.6:1: 24.4 (0-100)

3.6 :1:3.1 (0.2-20.0)2.1 :1:2.2 (0-10.2)0.2:1: 0.3 (.001-2.3)

18.0:1: 13.6 (1-66)168/1158.9:1: 3.4 (2.9-25.4)6.2 :1:3.0 (1.2-23.0)

68.3:1: 13.0 (27.0-91.0)5.3:1:1.2 (2.7-9.6)311 :1: 400 (6-3500)66:1: 134 (0-1260)

26.9 :1: 29.3 (0-95)0.5:1: 0.3 (0.07-2.4)3.1 :1:0.6 (1.1-4.7)0.6 :1:0.2 (0.2-1.3)

<.00010.14< .0001<.0001<.0001<.0001< .0001<.0001<.0001<.0001< .0001<.0001

Age (years)Gender (MlF)Leukocyte count (1 09/1)PMNcount(109/1)Percent PMNsBlood glucose (mmol/l)CSF leukocyte count (1 06/1)CSF PMN count (1 06/1)Percent PMNs in CSFCSF protein (g/l)CSF glucose (mmol/l)CSF/blood glucose ratio

. p value of Mann-Whitney test for the comparison of quantitative variables and of Pearson's chi-square test for

gender.

ters but gender, we showed a strongly significantdifferenœ between ABM and A VM cases. How-ever, aIl of these parameters had a wide distribu-tion range often with a large overlap betweenABM and A VM. The distribution of month ofonset was also quite different for ABM (peak in-cidence in January, February and March) than forAVM (peak incidenœ in June, July and August).Among the 115 ABM cases, 95 (83 %) had a posi-tive CSF culture, 12 had positive CSF Gram stainand CSF antigen latex agglutination tests, and 8had only a positive CSF antigen latex agglutina-tion test. CSF Gram stain was negative in 53 of the115 ABM cases (46 %). The 62 positive Gramstains included 30 gram-positive cocci, 24 gram-negative cocci, 7 gram-negative rods and 1 gram-positive rod. The microorganisms responsible forthe 95 culture-proven cases were Neisseriameningitidis (n = 33), Streptococcus pneumoniae(n = 26), Listeria monocytogenes (n = Il),Haemophilus influenzae (n = 10) and others{n = 15). The 20 microorganisms identified by apositive antigen latex agglutination result wereStreptococcus pneumoniae (n = 12), Neisseriameningitidis (n = 5) and Haemophilus influenzae(n = 3). Twelve of these microorganisms were alsodetected by CSF Gram stain. The breakdown ofcases into categories of ABM is shawn in Table 2.

PPV of 82 % and 68 %, respectively. ln otherwords, for a patient whose Spanos' model derivedpABM is lower than 0.1, one might Iule out thediagnosis of ABM with only a 1 % chance of er-roneous diagnosis. A cut-off point of 0.2 offersbetter specificity and PPV for a very milddecrease of NPV which remains as high as 98 %.

Our regression procedure resulted in the identifi-cation of four different independent variables.These were CSF protein level, total CSF poly-morphonuclear count, blood glucose level andleukocyte count. The complete mathematical ex-pression of our model was as follows:

1pABM = L ' where(1 + e- )

L = 32.13 x 10-4 x CSF PMN count (106/1)+ 2.365 x CSF protein (g/l)+ 0.6143 x blood glucose (mmol/l)+ 0.2086 x white blood cell count (109/1) -11.

The value of the Hosmer-Lemeshow goodness-of-fit statistic (10) computed with BMDP LRprogram is 4.59, and its corresponding p-valuecomputed with the chi-square distribution is 0.71.This indicates that the model fits quite weIl (10).Sensitivity, specificity and predictive values of ourmodel for the same cut-off points as defined forSpanos' model are displayed in Table 3. Using acut-off point of 0.1 pABM is associated with sen-sitivity and NPV values identical to those foundwith Spanos' model for the same cut-off point.Moreover, using this cut-off point, PPV is as highas 85 %. A cutpoint of 0.2 is associated with ahigher PPV, 90 %, only counterbalanced by a 1 %reduction of NP\/: Areas under ROC curves were0.981 and 0.991 in Spanos' model and our model,respectively, not significantly different from eachother.

Sensitivity, specificity and predictive values ofSpanos' model for different cui-off points ofpABM are displayed in Table 3. ln the predictionof bacterial meningitis, a high sensitivity and ahigh NPV are the most important criteria. ln anaddition al search for good specificity and positivepredictive value, the cui-off point 0.1 for pABM isassociated with sensitivity and NPV values of 97 %and 99 %, respectively, and with a specificity and

Eur. J. Clin. Microbiol. Infect. Dis.270

We then focused our attention on the subgroup ofABM with negative Gram stain, which is prob-ably the most difficult diagnostic situation. Weapplied both models to this subgroup and toGram stain positive ABM cases. The results ofthis analysis are displayed in Table 4. The sameproportion of patients received ill-bouse antibiot-ics prior to lumbar puncture in both subgroups.Laboratory data tended to be less specific forABM cases with negative Gram stain than incases with positive Gram stain. CalculatedpABMs were significantly greater in Gram stainpositive cases. ln this subgroup, aIl the values ofpABM were equal to 0.99 using Spanos' model, inwhich aIl cases with Gram positive stain were as-signed this value. Using our model which did notpresuppose the value of pABM in Gram stainpositive cases, we found a mean pABM value of0.97. ln cases with negative Gram stain, the meanpABM value was 0.83 for both models, signifi-cantly lower than in Gram stain positive cases(p = 0.03). However, when applying the models tothe only cases of Gram stain negative ABM, thenegative predictive value calculated for the cut-off point pABM = 0.1 was 0.99, equal to that cal-culated when applying the models to the wholegroup of ABM.

Discussion

At first the present sindy confirmed that differen-tiating A VM from ABM is often a difficult task,since in our initial series of 500 cases, 20 % couldneither be categorized as ABM nor as A VM ac-cording to the case definition used. ln most ofthese cases the final diagnosis had remained un-certain because the patients were given antibioticsfor the treatment of Gram stain negative and cul-ture-negative meningitis. We feel that ibis per-œntage of 20 % is a good reflection of the propor-tion of patients with acute meningitis in whom thedifferential diagnosis is problematic (1). To ap-plaise whether the exclusion of these 102 casesmight have biased our results, we comparedthemto ABM and A VM definite cases. The sex ratiowas Dot different within the three groups. Forsncb quantitative variables as age, CSF and bloodlaboratory parameters, cases with uncertain eti-ology had mean values between those round inABM and A VM cases and a wide range overlap-ring those of ABM and A VM cases. This was alsothe case for model derived pABMs (Table 5),either Spanos' or ours, whose distributions were

Vol. 14,1995 271

Table 3: Sensitivity, specificity, positive and negative predictive values of Spanos' model and our model for different cut-offpoints of the model-derived probability of bacterial versus viral meningitis (pABM).

Present modelSpanos' model

Positive

predictivevalue(%)

pABM Specificity(%)

Positivepredictivevalue(%)

Negativepredictivevalue(%)

Sensitivity(%)

Specificity(%)

Negativepredictivevalue (%)

Sensitivity(%)

99999998989898979696949391

98979794939392908787868481

879093969798999999

100100100100

768085909295969797

100100100100

99999998979797969595959493

0.050.0750.10.20.30.40.50.60.70.80.90.950.99

97979796959494929089858377

6878829294959797989999

100100

5564688386899293959698

100100

Table 4: Proportion of patients treated with in-house antibiotics priorto lumbar puncture (pre-treated),laboratory findings (median/range) and model-derived probability of bacterial versus viral meningitis(pABM) in the 62 cases of gram-positive CSF and the 53 cases of gram-negative CSF bacterial

meningitis.

P value"Gram stain positive(n=62)

Gram stain negative(n=53)

0.7b0.08

0.12

0.04

0.04

0.06

0.03

0.03°

0.03

16 (26%)20.1/5.1-52.7

8.6/4.0-33.04.9/0.12-29.54.5/0.006-29.23.2/0.6-15.71.3/0-9.30.99 (0.99-0.99)0.97 (0.06-1.00)

15(28%)17.0/4.5-477.9/2.4-24.92.0/0.002-30.01.7/0-29.72.2/0.2-201.8/0-10.20.83 (0.007-1.0)0.83 (0.01-1.0)

Pre-treated, n(%)Leukocyte count (x 109/1)Blood glucose level (mmol/l)CSF leukocyte count ~x 109/1)CSF PMN count (x 10 /1)CSF protein (g/l)CSF glucose (mmol/l)Spanos'pABM, mean (range)Present pABM, mean (range)

"Mann-Whitney rank-sum test unless specilied otherwise.b Pearson's chi-square test.

Ct test lor the comparison 01 sample mean to expected mean.

both clearly bimodal, with one mode for pABMvalues lower than 0.1 and a second mode forpABM values greater than 0.9. We can thereforereasonably assume that cases with uncertain eti-ology included viral and bacterial meningitis in aproportion not very different from that of the 398cases with definite diagnosis. As a result, if a biashad been introduced by our analysis process, itshould have been only minor.

this was also the case for leukocyte count andPMN count. Not surprisingly, CSF protein leveland leukocyte count were retained in the multi-variate model.

ln a further step we validated Spanos' model on anew, large and independent sample of patientswith acute meningitis. We confirmed that pABM,the parametergenerat~d by Spanos' model, couldbe used as an accurate diagnostic test since thearea under ROC curve for this parameter wasvery close to 1, even higher than that found bySpanos et al. in their test sampl~; For the cul-offpoint pABM = 0.1, sensitivity and NPV of themodel were quite high. Specificity was only 82 %,meaning that for this cul-off point, up to 18 % ofA VM cases would be categorized as ABM ac-cording to Spanos' model. ln our model, for the

We then showed once again that no single para-meter could fuie out the diagnosis of bacterialmeningitis since the distribution of aIl but one ofthe CSF parameters tested overlapped the entirerange of values found in A VM. The exceptionconcerned CSF protein level for which somevalues found in A VM were lower than the lowestvalue found in ABM. Among blood parameters

Eur. J. Clin. Microbiol. Infect. Dis.272

Table 5: Mean values and range of probability of bacterial versus viral meningitis (pABM) according toSpanos' mode! and ours, forproven cases ofbacterial and viral meningitis as weil as for cases of uncer-tain etiology.

Viral meningitis(n = 283)

pABM Bacterial meningitis(n = 115)

Uncertain(n = 102)

Spanos'modelmean

range

0.920.007-1

0.070.0003-0.93

0.390.0005-1

Present modelmeanrange

0.040.0006-0.80

0.350.001-1

0.910.01-1

patient's age and month of onset. ln fact, althoughthe likelihood of developing a bacterial versusviral meningitis does vary with the period of theyear, it seems odd that, once defined by its clinicaland laboratory characteristics, the diagnosis ofmeningitis could switch from viral to bacterial de-pending on the month of onset. An advantage ofnot using the date of onset in the model is that theclinical prediction Iule could be used anywhere inthe world, not just in the northern hemisphere.

The following case report illustrates this facto A55-year-old woman was admitted on 12 November1992 with acute meningitis. Initiallumbar punc-ture showed 310 leukocytes/mm3 with 95 % lym-phocytes, protein and glucose levels of 0.8 g/1 and2.04 mmol/l, respectively. CSF Gram stain wasnegative. Blood leukocyte count was 4.6 x 109/1and blood glucose level was 5.61 mmol/l. She wastreated with antibiotics. Two days after admissionshe developed herpes zoster. Later on CSF viralcultures grew varicella-zoster virus. The calcu-lated pABM values according to Spanos' modeland our model were 0.62 and 0.01, respectively.Should the same patient have presented an iden-tical zoster meningitis in August, ber Spanos'pABM would have been 0.32. Owing to the re-sults of our model (pABM = 0.01), one mighthave decided not to treat the patient since nega-tive predictive value of the model for this cut-offpoint is greater than 99.5 %. Even in the' Augustversion,' pABM calculated according to Spanos'model was too high for ruling out the diagnosis ofABM. This is probably due to the patient's age, atwhich a bacterial meningitis is more likely tooccur than a viral one.

CSF polymorphonuclear count was the only vari-able common to both Spanos' model and ours. lnour model this variable predicted ABM betterthan either the percentage of polymorphonuclearrelIs in the CSF or the total CSF leukocyte count.This may not be surprising since it is the product

same cut-off point pABM = 0.1, we found roughlythe same sensitivity and NPV values, whereasspecificity was better since less than 7 % of A VMwere categorized ABM. However, it should bementioned that calculating sensitivity, specificity,predictive values and area under the ROC curveof our model on the same data set used to build itmay artificially favor our model. These perform-ance indexes might weli decrease if our modelwere tested with independent data, as it was forour validation of Spanos' model.

With both models, 3 of 115 patients with ABMhad a pABM value lower than 0.1. The responsiblepathogens were Streptococcus pneumoniae,Haemophilus influenzae and Staphylococcusepidermidis. ln ali three patients, pABMs werecalculated based on the results of the first CSF ex-amination, as planned in the design of the study.ln these patients, the first lumbar tap was doneearly in the course of the disease and was flot sug-gestive ofbacterial meningitis at that time. A sec-ond lumbar tap, performed within 24 bOUTS of thefirst one, was then typical ofbacterial meningitis.

We found that both predictive models accuratelyâifferentiated ABM from AVM in the subgroupof ABM with negative Gram stain in which differ-ential diagnosis is the most difficult. The diagno-sis of ABM can be ruled out with good confidence(1 % chance of error) in negative Gram stainmeningitis when model derived pABM is equal toO.lorlower.

The tact that the distribution of causative micro-organisms in ABM was different in our series ascompared to Spanos' (more N. meningitidis andless H. influenzae) does flOt seem to have alteredthe validity of the clinical prediction Iule.

Ali four variables retained in our model werelaboratory parameters obtained from patient'sblood or CSF immediately after bis or ber admis-sion to hospital. They were independent of the

273Vol. 14,1995

3.5 mmol/1. Blood leukocyte count was 6.3 x 109/1,and serum glucose was 7 mmol/1. CSF Gram stainwas negative. On one band the patient had noclinical evidence of ABM, but on the other bandCSF was cloudy and contained more than500 x 106 PMN/1.. Calculated pABM according toSpanos' model and ours was 0.065 and 0.069 re-spectively, bath faT lower than 0.10, the value ofpABM associated with a NPV of 99 %. We hy-pothesized that the patient had A VM with pre-dominant PMN in CSF due to early lumbar taroAccordingly, we decided to delay the administra-tion of antibiotics. The patient became afebrilewithout treatment and 48 hauTs after admission,lumbar tap showed a clear CSF containing500 x 106 lymphocytes/1. The diagnosis of A VMwas accepted, although CSF viral culture re-mained negative, and the patient was dischargedon the fourth day.

ln this case report, the model provided an aid indeciding not to treat the patient. The pABM cut-off point value that allows ruling out the diagnosisof bacterial meningitis may be chosen individu-ally by each clinician, according to the NPV as-sociated with a given pABM. ln our ongoing pro-spective study, we considered that a pABM valueequal to 0.10 might be an accurate cut-off point,bath discriminant and harmless since it is as-sociated with a NPV equal to 99 %. For a differ-ynt management strategy, other clinicians mightchoose other cutpoint values using the data givenin Table 3.

Finally, we share the opinion of Spanos et al. (6)that a probability is only a probability, not thefinal answer. We likewise recommend that pABMshould be regarded as one piece of diagnosticinformation among others and should never besubstituted entirely for a careful diagnostic evalu-ation of each individu al case.

Acknowledgement

We are deeply indebted to C. Dobrovolny for reviewingthis manuscript.

of the two. Nye et al. (12) has previously shownthat CSF differentialleukocyte COllot was a betterguide to diagnosis than the total COllot. However,they did DOt appraise the diagnostic value of thetotal polymorphonuclear cell COllOt.

v

CSF protein level was the first step~entered para-meter of the model. After this jirst regressionstep, CSF glucose level and CSF/blood glucoseratio dramatically lost capability of improving themodel while this was DOt the case foI:~blood glu-cose level. One explanation is that CSF glucose isstrongly inversely correlated to CSF protein, asevidenced in our series (r = -0.416, P < 0.001). lnaddition, blood glucose level which is known to beelevated in ABM (12, 13) appeared in this studyas an independent predictor of bacterial menin-gitis, irrespective of CSF glucose level.

The results of our study support the accuracy ofSpanos' model. Although we found anothermodel based on slightly different variables, wecao consider that this approach of combiningbasic, routinely and immediately available para-meters is clinically valuable~ Other investigatorshave previously demonstrated that a combinationof parameters is much more helpful than anysingle parameter (2, 12). The multivariate modelprovides a valu able aid in differentiating A VMfrom ABM with both high negative and positivepredictive values and consequently with goodconfidence. The predictor pABM proved accu-rate in excluding bacterial in favor of viralmeningitis. So far, no single CSF or blood para-meter could achieve this goal although thereverse (excluding viral in favor of bacterialmeningitis) was quite possible.

Recently McKinney et al. (14) have also validatedSpanos' model on a retrospective series of 62ABM cases and 98 A VM cases which originatedfrom five hospitals in Dallas, Texas, and Mil-waukee, Wisconsin, USA. They showed that theclinical prediction fuie proved robust when ap-plied to a geographically distinct population ofadults, irrespective of patients' age.

However, both Spanos' model and ours should befurther validated in a prospective appraisal oftheir contribution to therapeutic decisionmaking. We started such an evaluation in our in-stitution, as illustrated in the following case re-port. A 32-year-old male doctor complained offever, headache and vomiting for about 12 hourswhen he was admitted to our hospital on 15 June,1992. Ris neck was stiff. Lumbar tap revealed aslightly cloudy CSF containing: leukocytes 700 x10 /1 (PMN = 75 %), protein 0.43 g/1 glucose

References

1. Geiseler PJ, Nelson KE, Levin S, Reddi KT, MosesVK: Community-acquired purulent meningitis: a re-viewof 1,316 cases during the antibiotic era, 1954-1976. Reviews of Infectious Diseases 1980, 2: 725-745.

2. Landaas S, Von der Lippe B: Chemical analyses forearly differential diagnosis between bacterial and viralmeningitis. Scandinavian Journal of Clinical andLaboratory Investigations 1985, 45: 525-529.

Eur. J. Clin. Microbiol. Infect. Dis.274

3. Briem H: Comparison between cerebrospinal fluidconcentrations of glucose, total protein, chloride, lac-tate, and total ami no acids for the differential diagnosisof patients with meningitis. Scandinavian Journal ofInfectious Diseases 1983, 15: 277-284.

4. BenGershôm E, Briggeman-Mol GJJ, de Zegher F:Cerebrospinal fluid C-reactive protein in meningitis:diagnostic value and pathophysiology. EuropeanJournal of Pediatrics 1986, 145: 246-249.

5. Lindquist L, Linné T, Hansson LO, Kalin M, AxelssonG: Value of cerebrospinal fluid analysis in the differ-ential diagnosis of meningitis: a study in 710 patientswith suspected central nervous system infection.European Journal of Clinical Microbiology & InfectiousDiseases 1988, 7: 374-380.

6. Spanos A, Harrell FE, Durack DT: Differential di ag-nosis of acute meningitis. An analysis of the predictivevalue of initial observations. Joumal of the AmericanMedical Association 1989, 262: 2700-2707.

7. Baker AC, Lenane AM: The predictive value of cere-brospinal fluid differential cytology in meningitis. Pe-diatric Infectious Disease Journal 1989. 8: 329-330.

8. Fletcher RH, Fletcher FW, Wagner EW: Diagnostictest. ln: Clinical epidemiology, the essentials. Williams& Wilkins, Baltimore, 1982, p. 41-58.

9. Hanley JA, McNeil BJ: The meaning and use of thearea under a receiver operating characteristics (ROC)curve. Radiology 1982, 143: 29-36.

10. Hosmer DW, Lemeshow S: Applied logistic regres-sion. John Wiley, New York, 1989, p. 140-145.

11. Hanley JA, McNeil BJ: A method of comparing theareas under receiver operating characteristic curvesderived from the same cases. Radiology 1983,148:839-843.

12. Nye FJ: The value of initial laboratory investigationsin the management of meningitis. Journal of Infection1983, 7: 31-38.

13. Swartz MN, Dodge PR: Bacterial meningitis - a reviewof selected aspects. New England Joumal of Medicine1965,272: 779-787.

14. McKinney WP, Heudebert GR, Harper SA, Young MJ,Mclntire DD: Validation of a clinical prediction rulefor the differential diagnosis of acute meningitis. Jour-nal of General Internai Medicine 1994, 9: 8-12.