Development and verification of a discriminate algorithm for diagnosing post‐neurosurgical bacterial meningitis—A multicenter observational study

Abstract Objective To evaluate the diagnostic accuracy of cerebrospinal fluid (CSF)–based routine clinical examinations for post‐neurosurgical bacterial meningitis (PNBM) in multicenter post‐neurosurgical patients. Methods The diagnostic accuracies of routine examinations to distinguish between PNBM and post‐neurosurgical aseptic meningitis (PNAM) were evaluated by determining the values of the area under the curve (AUC) of the receiver operating characteristic curve in a retrospective analysis of post‐neurosurgical patients in four centers. Results An algorithm was constructed using the logistic analysis as a classical method to maximize the capacity for differentiating the two classes by integrating the measurements of five variables. The AUC value of this algorithm was 0.907, which was significantly higher than those of individual routine blood/CSF examinations. The predicted value from 70 PNBM patients was greater than the cutoff value, and the diagnostic accuracy rate was 75.3%. The results of 181 patients with PNAM showed that 172 patients could be correctly identified with specificity of 95.3%, while the overall correctness rate of the algorithm was 88.6%. Conclusions Routine biomarkers such as CSF/blood glucose ratio (C/B‐Glu), CSF lactate (C‐Lac), CSF glucose concentration (C‐Glu), CSF leukocyte count (C‐Leu), and blood glucose concentration (B‐Glu) can be used for auxiliary diagnosis of PNBM. The multicenter retrospective research revealed that the combination of the five abovementioned biomarkers can effectively improve the efficacy of the PNBM diagnosis.


| INTRODUC TI ON
Post-neurosurgical meningitis (PNM), a common complication of neurosurgical procedures, has an incidence rate of 0.3%-25%. 1 PNM significantly affects the length of hospital stay (LOS), increases the cost, decreases the success rate of neurosurgery, and increases patients' mortality rate. PNM can cause approximately 35% mortality. 2 PNM has been mainly divided into post-neurosurgical bacterial meningitis (PNBM) and post-neurosurgical aseptic meningitis (PNAM). 3 The pathogenesis and treatment for PNBM and PNAM are completely different. The main cause of PNBM is infections induced by pathogenic bacteria; the onset of PNBM is rapid, and antibiotics are required for treatment. Scattered bone fragments or tumor antigens produced during neurosurgery may be the cause of PNAM, and no antibiotics are required for its treatment. 4 PNBM and PNAM always share some physical signs and clinical symptoms, including mental status changes, neck stiffness, headache, fever, and vomiting, which do not sufficiently specify the clinical manifestations of both meningitis forms to diagnose PNBM. 5 Given these influencing factors, rapid diagnostic methods are needed urgently to provide more diagnostic options and evaluate the efficacy of antibiotic drugs.
Presently, regardless of the Infectious Diseases Society of America (IDSA) 6 and Centres for Disease Control and Prevention (CDC) 7 criteria, the diagnostic strategy for PNBM is based on the combination of clinical symptoms and laboratory tests of the patients. The gold diagnosis standard for bacterial meningitis is the cerebrospinal fluid (CSF) bacterial culture. Culture is, however, unreliable due to time consumption and extensive antibiotic prophylaxis ahead of neurosurgical operations, and the results are available only after a couple of days. Several novel biomolecular approaches to the rapid diagnosis of PNBM have been applied in clinical laboratories, such as PCR and next-generation sequencing (NGS). However, these new approaches have some specific weaknesses. PCR has a higher false-positive feature and does not meet the need. On the other hand, although NGS has high-throughput properties, the cost, complexity, and time consumption of NGS are not suitable for routine clinical application. 8   Continuous data were expressed as mean ± SD or median (25%, 75%), whereas categorical data were expressed as numbers and percentages. Continuous variables were analyzed by nonparametric Mann-Whitney U test or Student's t test when appropriate, and chi-square or Fisher's exact test was performed for the categorical data. Univariate analysis was employed to calculate the P values for all variables; a multivariate algorithm was performed to take into account differences between the two groups by using a logistic regression. All the variables whose P was < .05 were embedded into the one fitting variable. 12

| Univariate analysis of the parameters
We evaluated the diagnostic accuracy of these variables with a statistically significant difference. The result of the univariate analysis for each variable is shown in Table 2. Among these tests, in general, CSF variables performed better than blood variables. All the CSF biomarkers, including C-Cell (10 6 /L), C-Leu (10 6 /L), C-Neu (%), C-Glu (mmol/L), C-Pro (mg/dL), C-Cl -(mmol/L), and B-Glu (mmol/L), showed a statistically significant difference between the PNBM group and the PNAM group.

| Multivariate analysis of the parameters
We included all the factors with P value < .2 in the univariate analysis into the logistic multivariate analysis. C-Leu, C-Glu, B-Glu, C/B-Glu, and C-Lac were determined as the parameters for the algorithm's construction (Table 3). Simultaneously, an algorithm was constructed with the analysis, which is a classification method that integrated the five abovementioned biomarkers to maximize the capacity for differentiating between the PNBM and PNAM groups. The algorithm of PNBM is as follows: F = −0.176*C-Glu + 0.191*B-Glu + 0.515*C-Lac-3.351*C/B-Glu + 0.0000173*C-Leu-1.584.

| Verification of the multivariate algorithm
The fitted variable and the five biomarkers above, including C-Leu   .738 Note: P < .05 was considered to be significant.
verification (Table 4 and Figure 1). Among these biomarkers, C-Leu at a cutoff value of 577.5*10 6 /L had the highest specificity (83.6%), and C-Lac had the best sensitivity (82.4%) and PPV (78.0%). The results of the ROC curve of the fitted biomarkers, which formed in the multivariate algorithm, are shown in Table 5. The AUC of the fitted variable was 0.907, and the sensitivity, specificity, PPV, and NPV were greater than 80.0%. The cutoff of the fitted variable is 0.505. The AUC indicated that the fitted variable was a much better biomarker for distinguishing PNBM from PNAM.

| Verification by the hypothetical cohort of 1000 patients
In  Table 6.

| D ISCUSS I ON
PNBM can be diagnosed through clinical symptoms, CSF laboratory biomarker analysis, and Gram staining with some uncertainty, and ideally through CSF bacterial culture. The analysis of CSF laboratory biomarkers is one of the most pivotal methods to diagnose PNBM. 13 However, many studies have reported that routine infection markers  perience an inflammatory reaction; thus, the difference between the two is actually significant. 16 In addition to routine markers, C-Lac is the second variable for diagnostic efficacy (AUC = 0.791).
In a prospective study, C-Lac can effectively distinguish PNBM from PNAM when its concentration is greater than 6 μmol/L, and the value of C-Lac was determined to be 4-6 μmol/L for PNBM treatment period, and <2 μmol/L for PNAM. 17 Although some biomarkers recommend a good diagnosis ability to discriminate PNBM from PNAM, however, the specificity of the majority monobiomarkers was low (eg, C/B-Glu, 74.30%; C-Glu, 69.50%; C-Leu, 56.00%; B-Glu, 58.80%); thus, the diagnosis of meningitis cannot be performed alone and multiple indicators with combined diagnosis are needed.
Multivariable integrated diagnosis can effectively improve the diagnostic performance with enhanced sensitivity and specificity 18 ; hence, we constructed a multivariable algorithm of PNBM. Through the logistic regression analysis, we obtained five biomarkers (C/B-Glu, C-Lac, C-Glu, C-Leu, and B-Glu) to find the best compositions and choose them as the candidate to construct algorithm. The main limitation of this study is that it comprised routine examinations, but clinical symptoms such as body temperature and age were not considered. In addition, infection markers such as procalcitonin, interleukin-6, and C-reactive protein were not discussed in this study. In our future studies, new biomarkers will be gradually included as variables of the algorithm in order to further increase the sensitivity and specificity of this method.

| CON CLUS ION
To summarize, routine biomarkers such as C/B-Glu, C-Lac, C-Glu, C-Leu, and B-Glu can be used for the auxiliary diagnosis of PNBM.
In this study, multicenter research was performed to show that by combining the five abovementioned biomarkers, we can effectively improve the efficacy of PNBM diagnosis and can speed up the entire diagnostic procedure.

CO N FLI C T O F I NTE R E S T
The authors declare no conflict of interest.

E TH I C A L A PPROVA L
No ethical approval was needed in this study.