Accurate prognostic stratification at presentation facilitates interpretation of individual treatment results and of benefits of therapeutic interventions in clinical trials. Unfortunately, methodologic limitations of existing studies prevented the development of a clinically useful prognostic model for persons with bacterial meningitis [1, 5-26]. To create a prognostic model and risk stratification system, we conducted a retrospective, observational cohort study of persons with community-acquired, microbiologically proven bacterial meningitis.

Methods

Top Methods Results Discussion Author & Article Info References

Patients were identified through review of medical records and microbiology laboratories at four Connecticut hospitals. To develop and test the prognostic model, we divided patients in the total cohort into derivation and validation cohorts. The derivation cohort (n = 176) was assembled from persons at least 16 years of age in whom community-acquired bacterial meningitis was diagnosed at the two hospitals serving New Haven, Connecticut (Yale-New Haven Hospital and Hospital of Saint Raphael), from January 1970 through December 1995. The validation cohort (n = 93) was assembled similarly during the same period for patients given a diagnosis at the two hospitals serving a separate geographic location in Waterbury, Connecticut (Waterbury Hospital Health Center and St. Mary's Hospital). Patients with community-acquired meningitis (initial lumbar puncture done within 24 hours of presentation in the emergency department) and a microbiologically identified bacterial cause were included. Bacterial meningitis was diagnosed on the basis of a compatible clinical picture and a positive cerebrospinal fluid culture [n = 226] or a negative cerebrospinal fluid culture with one or more of the following: 1) a positive result on a cerebrospinal fluid bacterial antigen or Quellung test [n = 10], 2) a positive blood culture in the presence of cerebrospinal fluid pleocytosis (defined as a leukocyte count 10 cells/mL of cerebrospinal fluid) [n = 39], or 3) identification of gram-negative diplococci after Gram staining of cerebrospinal fluid (n = 7). Patients with meningitis caused by Mycobacterium species, Treponema pallidum, or Borrelia burgdorferi and patients with intracranial devices were excluded. For patients who had multiple episodes of meningitis during the study period, only the first episode was analyzed. From medical records, we extracted information on sociodemographic data, comorbid conditions, immunocompetence, exposures, presenting symptoms, physical signs, laboratory and cranial imaging results, and clinical outcome. Precise times were recorded for arrival in the emergency department, administration of the first dose of effective antibiotics (defined as antibiotic therapy to which the identified bacteria were susceptible in vitro), and performance of lumbar puncture. Baseline characteristics were assessed at a specified "zero time," defined as the time at which the first dose of effective antibiotics was administered. Comorbidity was assessed by using the Charlson comorbidity scale [27]. For the derivation cohort, bivariate analyses were done with the presence or absence of an adverse clinical outcome (defined as neurologic deficit at discharge or death during hospitalization) as the target end point. All patient episodes were followed from admission date to discharge or death. Baseline characteristics associated with adverse clinical outcome in the bivariate analyses were analyzed in a multivariable context by using multiple logistic regression modeling. Using these independently prognostic baseline variables, we developed a prognostic model that stratified patients into three levels of risk (low, intermediate, and high) for adverse clinical outcome. The prediction accuracy of the prognostic model was assessed in both the derivation and validation cohorts.

The effect of antibiotic timing was evaluated for the total cohort and was analyzed for episodes in which the patient remained in the same prognostic stage as well as those in which the patient's prognostic stage advanced from the time of arrival in the emergency department to the time of initiation of antibiotic therapy.

Differences in proportions were tested by using the chi-square test or the Fisher exact test. For contrasts of dimensional variables, the Student t-test and the Wilcoxon rank-sum test were used. We used multiple logistic regression given the binary target outcome. To avoid overfitting, we entered no more than one variable per 10 outcome events in our logistic regression models [28-30]. The goodness of fit of the final model was examined by using the Hosmer-Lemeshow test. Prediction accuracy was calculated by using the concordance index (c-index) [31].

Results

Top Methods Results Discussion Author & Article Info References

Cohort Assembly

A total of 269 persons with community-acquired, microbiologically proven bacterial meningitis were identified at the four hospitals [176 patients in the derivation cohort and 93 patients in the validation cohort]. After initial screening of 1024 medical records, 755 episodes were excluded for the following reasons: 1) lumbar puncture was never done [n = 155], 2) lumbar puncture was done more than 24 hours after presentation [n = 185], 3) an incorrect International Classification of Diseases, Ninth Revision, diagnosis code was applied [n = 10], 4) the case definition was not fulfilled [n = 317], 5) the medical record was incomplete [n = 80], or 6) the patient's meningitis episode was not the first episode in the study period (n = 8).

Baseline Features and Microbial Causes

Baseline patient characteristics, laboratory results, and microbial causes are shown in Table 1. The derivation and validation cohorts were similar with respect to age, sex, ethnicity, and baseline laboratory features. Baseline clinical features were also similar, but significantly more patients in the derivation cohort than in the validation cohort were immunocompromised (P = 0.03) and had focal neurologic examinations (P = 0.03). For the total cohort, fever and abnormal mental status were the most common presenting clinical features. Streptococcus pneumoniae caused nearly 50% of all cases, and Neisseria meningitidis, other streptococci, Staphylococcus aureus, and Listeria monocytogenes caused most of the remaining cases. The derivation cohort had a greater proportion of patients whose meningitis was caused by Staphylococcus aureus than did the validation cohort (P = 0.02).

Clinical Outcome

The clinical outcome for all patients is shown in Table 1. The hospital mortality rate for the total cohort was 27%, and the median number of hospital days before death was 6 (range, 0 to 65). Of the 269 patients, 56 (28% of survivors; 21% of the total cohort) had a neurologic deficit and 25 (13% of survivors; 9% of the total cohort) had a neurologic deficit that persisted at discharge. The most common persisting neurologic deficit in the total cohort was hearing loss. Clinical outcome for the total cohort did not differ significantly when analyzed by time period (P > 0.2 when adverse outcome were compared between 5-year intervals). However, episodes of Streptococcus pneumoniae meningitis in the total cohort were associated with a greater proportion of adverse clinical outcomes than were episodes that were not caused by Streptococcus pneumoniae (P = 0.005).

Bivariate analyses relating baseline characteristics to adverse clinical outcome in the derivation cohort are shown in Table 2 and Table 3. Hypotension (defined as systolic blood pressure 90 mm Hg or a 40 mm Hg decrease in systolic blood pressure), altered mental status (defined as lethargy, disorientation, or coma), seizures, and age all showed a bivariate association with adverse clinical outcome. To determine independent relations with adverse clinical outcome, these variables were entered into a logistic regression model with three other variables that had a biologically plausible relation to clinical outcome (immunocompetence, comorbidity, and leukocyte count in cerebrospinal fluid). In this model, only hypotension (adjusted odds ratio, 2.75 [95% CI, 1.22 to 6.18]; P < 0.001), altered mental status (adjusted odds ratio, 6.56 [CI, 1.71 to 25.2]; P = 0.006), and seizures (adjusted odds ratio, 4.42 [CI, 1.56 to 12.48]; P = 0.005) remained independently associated with adverse clinical outcome after adjustment for age, comorbidity, immunocompetence, and leukocyte count in cerebrospinal fluid.

Development and Validation of the Prognostic Model

Using the three independently predictive baseline variables from the logistic regression model (hypotension, altered mental status, and seizures), we created prognostic stages by stratifying patients who had 0, 1, and at least 2 of the variables into low-, intermediate-, and high-risk subgroups, respectively (Table 4). For the derivation cohort, the observed proportions of patients with adverse clinical outcome were 5% in stage I (low risk), 37% in stage II (intermediate risk), and 63% in stage III (high risk) (P = 0.001). For the validation cohort, the observed proportions of patients with adverse clinical outcome were 14% in stage I, 28% in stage II, and 43% in stage III (P = 0.06). The predicted probabilities of the model for each stage closely match the observed outcome event rates in both the derivation and validation cohorts. The Hosmer-Leme-show test yielded a P value of 0.08 for the derivation cohort and a P value greater than 0.2 for the validation cohort. This shows that no strong evidence supports lack of fit of the model. Calculation of prediction accuracy by using the c-index showed a value of 0.73 (CI, 0.65 to 0.81) for the derivation cohort; the c-index for the validation cohort was 0.81 (CI, 0.71 to 0.92). Because the CIs were greater than 0.50, the model predicted significantly better than chance for both cohorts.

Timing of Antibiotic Therapy and Clinical Outcome

The effect of antibiotic timing on clinical outcome was analyzed in two ways. For patients who remained in the same prognostic stage from arrival in the emergency department until their first dose of antibiotics, the time of delay in initiation of therapy was compared for those with and those without adverse clinical outcomes (Table 5). As shown for the total cohort of 227 patients who remained in a given prognostic stage, the median delay in initiation of antibiotic therapy was 4.0 hours and did not significantly differ between patients with and those without an adverse outcome (4.5 hours compared with 3.9 hours; P > 0.2). Similarly, for patients in each prognostic subgroup, delay in initiation of antibiotic therapy did not significantly differ between those with and those without an adverse outcome. Patients in the low-risk subgroup (stage I) with an adverse outcome showed a trend toward a longer median delay in antibiotic therapy (13.6 hours compared with 6 hours for those without an adverse outcome), but the difference was not statistically significant (P = 0.08) because of the wide range of values and the relatively small numbers in this subgroup.

The results for patients whose prognostic stage advanced from arrival in the emergency department until the initiation of antibiotic therapy are shown in Table 6. Of the 45 patients who were in stage I at the time of their arrival in the emergency department, 35 were in stage I at the time of initial antibiotic therapy; however, 6 advanced to stage II and 4 advanced to stage III. As noted, patients who advanced from stage I to stage III before initiation of antibiotic therapy had a greater proportion of adverse outcomes (3 of 4 [75%]) than those who remained in stage I at the initiation of antibiotic therapy (3 of 35 [9%]) (P = 0.008). Similarly, of the 191 patients who were in stage II at the time of arrival in the emergency department, 159 were in stage II at the time of initial antibiotic therapy but 32 advanced to stage III. Patients who advanced from stage II to stage III before initiation of antibiotic therapy had a significantly greater proportion of adverse outcomes (20 of 32 [63%]) than those who remained in stage II at the initiation of antibiotic therapy (56 of 163 [34%]) (P = 0.003). Additional analyses revealed no relation between time spent in the advanced prognostic stage and clinical outcome. Specifically, the time spent in the advanced stage was not associated with adverse outcome in patients who advanced from stage II to stage III (P > 0.2). Similarly, no relation existed between time spent in the advanced stage and clinical outcome for patients who advanced from stage I to stage III (P > 0.2), but there were too few patients in this subgroup to allow a meaningful comparison.

Discussion

Top Methods Results Discussion Author & Article Info References

This study analyzed the largest cohort to date of adults with community-acquired, microbiologically proven bacterial meningitis. Three baseline clinical features predicted adverse clinical outcome and stratified patients into distinct prognostic stages. Adverse clinical outcome was associated with a delay in initiation of antibiotic therapy that allowed prognostic stage to advance from the stage at arrival in the emergency department (stage I or stage II) to stage III.

Baseline clinical features, microbial causes, and mortality rates in our study were similar to those in another large study of bacterial meningitis in adults [1]. Although 28% of episodes of community-acquired bacterial meningitis in that study resulted in neurologic illness, the authors included both transient and persistent neurologic deficits. In our total cohort, 21% of patients (57 of 277) developed a neurologic deficit, but the deficit persisted in only 9% (26 of 277) at discharge. For development of the prognostic model, we included only persistent neurologic deficits in our definition of adverse clinical outcome to minimize detection bias and better reflect clinically significant morbidity in our outcome analyses.

Our study design provided several advantages and avoided the methodologic limitations of previous work. First, we specified a "zero time" at which baseline clinical variables were assessed for each patient. Specifying a zero time is particularly important in baseline prognostic assessment of acute infectious diseases because clinical features of disease severity can evolve rapidly. Lack of a specified zero time can cause prognostic features to be identified in one cohort that are not predictive of clinical outcome in other cohorts. Because the time of initial therapy for any disease is the moment at which a prognostic estimate is most useful in management [32], we selected the time of initial antibiotic therapy as the zero time. A second advantage of our study was the use of a validated comorbidity index [27] to search for an association between comorbidity and adverse outcome. Although we hypothesized that patients with comorbid disease would have worse clinical outcome than those who did not, we could not show this relation in our multivariable analyses when including variables that reflected severity of illness (that is, hypotension, altered mental status, and seizures). Third, our prognostic model was developed by using baseline clinical features that were identifiable by the physician when treatment was initiated (zero time); this resulted in a practically useful prognostic tool for the bedside clinician [31, 33]. Fourth, we recognized the hazards of multiple statistical contrasts and the need for a sufficient number of outcome events in multivariable analyses. Therefore, we analyzed biologically plausible baseline variables in our bivariate analyses and entered no more than one variable per 10 outcome events into our logistic regression models [29, 30, 33, 34]. Finally, we tested the validity of the prognostic model in an independent validation cohort.

The three baseline characteristics (hypotension, altered mental status, and seizures) independently associated with adverse clinical outcome represent features of meningitis severity at the time of initial antibiotic therapy. Therefore, they are not merely statistical associations but biologically plausible clinical features reflecting pathophysiologic changes in cytokine generation [35-37], local nitric oxide production [38], and cerebrovascular function [39-41] seen in experimental bacterial meningitis. When creating the prognostic staging system, we stratified patients on the basis of the number of independent prognostic features present at zero time. We could have "weighted" the individual prognostic features in our staging system [42], but the goal was to create a valid yet simple stratification scheme for clinical use.

Although the creation of a prognostic model for bacterial meningitis represented a unique contribution of this study, it was important to test its prediction accuracy. To do this, we calculated the c-index. The c-index provides a numeric summary of the prediction accuracy of the model in discriminating between patients who will have an adverse clinical outcome and those who will not. Specifically, the c-index considers all pairs of patients (one with an adverse clinical outcome and the other without). It represents the percentage of pairs for which the model correctly predicts a higher probability of adverse clinical outcome. As shown in Table 4, the c-index was 0.73 (CI, 0.65 to 0.81) for the derivation cohort; this finding indicates that 73% of the time, the model correctly assigned a higher probability of adverse clinical outcome to the patient in the pair in whom it occurred. To test the validity of the model, the c-index was calculated for the independent validation cohort. The resulting c-index was 0.81 (CI, 0.71 to 0.92). Because the entire CI for the c-index in both cohorts was greater than 0.5 (chance prediction), the prediction accuracy of the model was significantly greater than chance.

One distinct value of an accurate prognostic model is its ability to test the effect of therapeutic interventions in different subgroups of disease severity. We used this opportunity to analyze the effect of delayed administration of antibiotics on clinical outcome in our cohort of patients with bacterial meningitis. Standard reference sources [43, 44] have recommended that patients with bacterial meningitis be given antibiotics within 30 minutes of arrival in the emergency department. Although this is a laudable goal, it can rarely be achieved in the reality of clinical practice [2-4, 45], is unsubstantiated in published literature, and oversimplifies the effect of antibiotic timing on a disease with heterogeneous presentations. Surrogates of antibiotic delay (for example, duration of symptoms before presentation and time to sterilization of cerebrospinal fluid) have been associated with adverse outcome in some studies, but no study has identified delay in initiation of antibiotic therapy as an independent risk factor after adjustment for other variables that affect clinical outcome [23, 26, 46, 47].

In the context of our prognostic staging system, we used two strategies to analyze the effect of antibiotic timing on clinical outcome. First, we analyzed patients who remained in the same prognostic stage from the time of arrival in the emergency department until the initiation of antibiotic therapy. For these patients (n = 227), we compared the time of delay in antibiotic administration with clinical outcome. For this entire group, as well as within each prognostic stage, antibiotic delay was not statistically significantly associated with adverse clinical outcome (Table 5). A trend in stage I patients suggested an association with longer delay and adverse clinical outcome, but this trend did not reach statistical significance (P = 0.08).

A second strategy for assessing the effect of antibiotic timing was to analyze clinical outcome in patients whose prognostic stage advanced from arrival in the emergency department until administration of the initial antibiotic dose. In this analysis, a clear association between antibiotic timing and clinical outcome emerged. Patients who arrived in the emergency department in stage I or II but advanced to stage III at the time of initial antibiotic therapy had significantly more adverse clinical outcomes (regardless of the time spent in stage III) than those who remained in the original prognostic stage (Table 6).

These observations are important in several respects. First, they provide quantitative, supportive data for an intuitive clinical assumption: Treatment of bacterial meningitis before it advances to a high level of clinical severity improves clinical outcome. Second, for patients who arrive in the emergency department with the highest level of clinical severity (that is, stage III), the risk for adverse outcome is influenced more by the severity of illness than the timing of initial antibiotic therapy. Third, the effect of antibiotic timing on clinical outcome is complex, varies with different levels of disease severity, and has little to do with an arbitrary time interval.

Despite the methodologic advantages of our study, there were limitations. First, our cohort was assembled retrospectively, raising the possibility of detection bias in assessment of baseline clinical features and clinical outcome. However, our rigorous definitions of zero time, baseline state, and adverse clinical outcome minimized this potential bias. Second, although the c-index revealed a prediction accuracy of the model that was statistically better than chance for both the derivation and validation cohorts, the observed gradient of adverse outcome was only significant at the P = 0.06 level in the validation cohort. In addition, there was a wide confidence interval estimate for stage III patients (CI, 0.26 to 0.63). These observations probably result from the comparatively small sample size of the validation cohort and suggest limitations in the generalizability of the model. Clearly, a large, geographically independent cohort would be helpful in proving the validity of the model. Third, despite a large total cohort, analyses within small subgroups may detect only high-prevalence variables that have a substantial effect on clinical outcome. This may explain our inability to detect an association between delay in initiation of antibiotic therapy and adverse outcome in the subgroup of stage III patients because therapy was initiated very rapidly (that is, within 30 minutes of arrival in the emergency department) in so few patients. In addition, although our data showed that patients who progressed from stage I to stage III before antibiotic treatment had a significantly worse outcome, the number of episodes in the subgroup was small and should be interpreted with caution.

Clinicians must recognize that our prognostic model applies only to patients with confirmed bacterial meningitis in whom lumbar puncture is performed within 24 hours of presentation, not to all patients with symptoms of meningitis. For patients with confirmed bacterial meningitis, our model will have immediate and future utility for clinical practice. First, it provides a simple and accurate estimate of clinical outcome for the individual episode of bacterial meningitis. This allows more accurate counseling to concerned families and can help physicians decide on the intensity of hospital care (for example, stage II and stage III patients might best be observed in an intensive care unit). Second, the staging system will assist future clinical trials in stratifying the effect of adjunctive therapies (for example, steroids and cytokine or nitric oxide inhibitors) that may benefit outcome in only one prognostic subgroup. Because disease severity is the most important predictor of adverse outcome, adjunctive therapies targeting the pathophysiologic cascade may offer the greatest hope to improving clinical outcome. Until then, initiating antibiotic therapy before advancement of disease severity should be the major therapeutic goal for physicians treating patients who have bacterial meningitis.