Although accurate models for predicting acute bacterial meningitis exist, most have narrow application because of the specific variables selected for them. In this study, we estimate the accuracy of a simple new model with potentially broader applicability.
On the basis of previous reports, we created a reduced multivariable logistic regression model for predicting bacterial meningitis that relies on age (years) (AGE), cerebrospinal fluid (CSF), total protein (TP) and total neutrophil count (TNC) alone. Data were from children ages 1 month–18 years diagnosed with acute enteroviral or bacterial meningitis whose initial CSF revealed >7 white blood cells/mm3. A fractional polynomial model was specified and validated internally by the bootstrap procedure. The area under the receiver operating characteristic curve (discrimination: criterion standard, >0.7), the Hosmer-Lemeshow deciles-of-risk statistic (calibration: criterion standard, P > 0.05) and sensitivity-specificity pairs at prespecified probability thresholds of the model were computed.
We identified 60 children with bacterial meningitis and 82 with enteroviral meningitis. At an area under the receiver operating characteristic curve of 0.97, our model represented by the equation: log odds of bacterial meningitis = 0.343 − 0.003 TNC − 34.802 TP0.5 + 21.991 TP − 0.345 AGE, was highly accurate when differentiating between bacterial and enteroviral meningitis. The model fit the data well (Hosmer-Lemeshow statistic; P =[r] 0.53). At probability cutoffs between 0.1 and 0.4, the model had sensitivity values between 98 and 92% and specificity values between 62 and 94%.
Among children with CSF pleocytosis, a prediction model based exclusively on age, CSF total protein and CSF neutrophils differentiates accurately between acute bacterial and viral meningitis.