Methods

Site, study design, and patients. We conducted a retrospective study at 4 tertiary care teaching hospitals: Beth Israel Deaconess Medical Center, a 740-bed hospital in Boston, Massachusetts; Duke University Hospital, a 750-bed hospital in Durham, North Carolina; Geneva University Hospital, a 1200-bed hospital in Geneva, Switzerland; and Tel Aviv Sourasky Medical Center, a 1200-bed hospital in Tel Aviv, Israel.

The occurrence of P. aeruginosa bacteremia upon hospital admission was determined using electronic microbiology databases. Each center examined a 3–5-year period during 2000–2005 to identify all cases of bacteremia due to gram-negative bacilli that occurred ⩽48 h after hospital admission (defined as bacteremia upon hospital admission). Each patient admission was included once, and all gram-negative bacilli isolates were included in the analysis. In all centers, specimens were processed in accordance with the CLSI guidelines [17], and isolates were identified using the Vitek II system (bioMérieux).

A retrospective case-control study was conducted to characterize patients affected with P. aeruginosa bacteremia upon hospital admission. Medical records of a random sample of 151 patients with P. aeruginosa bacteremia (defined as case patients) and 152 patients with bacteremia caused by Enterobacteriaceae (defined as control subjects) were reviewed for specific underlying conditions and contacts with the health care system. The following data were collected: age, sex, activities of daily living status [18], nursing home residency, specific comorbidities, severe immunodeficiency (defined as receipt of a solid organ or bone marrow transplant, presence of neutropenia [neutrophil count, <1000 cells/µL], receipt of recent chemotherapy [⩽30 days before bacteremia], AIDS [CD4 cell count, <200 cells/µL, or other evidence of AIDS], or receipt of treatment with high-dose corticosteroids [⩾20 mg/day prednisone equivalent for >5 days], azathioprine sodium, or cyclosporine), admitting syndrome and source of bacteremia, recent hospitalization, antimicrobial use during the previous 30 days, invasive procedures performed during the previous 2 weeks, and invasive devices present at the time of admission.

Statistical analysis. Continuous variables were compared using the unpaired t test. Categorical variables were compared using the Pearson χ² test or Fisher's exact test, as appropriate. Differences in distribution of populations were compared using Kruskal-Wallis equality-of-populations rank test. P values ⩽.05 were considered to indicate a statistically significant difference between groups. To find discriminative parameters between patients with P. aeruginosa bacteremia and patients with other gram-negative bacteremia, stepwise multivariable logistic regression was used. All variables with a P value ⩽.10 in the univariate analysis were considered for inclusion in the multivariable logistic regression analysis, and variables for which P<.05 in the multivariable analysis were retained in the final model. The statistical software Stata, version 9, (Stata), was used for analysis.

In addition to logistic regression, we used decision tree analysis for classification of patients with P. aeruginosa bacteremia as a function of a set of predictors that were statistically significant in the multivariate model with logistic regression [19, 20]. We randomly divided the data into in-sample (training) data with 180 cases and out-sample (testing) data with the rest of the cases. We fit the decision tree over the in-sample data using the procedure “treefit” with Matlab (Mathwork) and evaluated the prediction of the decision tree model obtained over the out-sample data using the procedure “treeval.” In addition, we used the “treeprune” procedure to avoid overfitting. We repeated this procedure 100 times and evaluated the best model for classification using area under the receiver-operating-characteristics curves and the log likelihood. To compare the logistic regression model with decision tree models, we used, in addition, 2 recently published criteria [21]: the net reclassification improvement and integrated discrimination improvement.

Among all evaluated criteria, logistic regression revealed similar classification capability. Thus, in our next step, we used logistic regression modeling to discriminate between patients with P. aeruginosa bacteremia and patients with other gram-negative bacteremia. For each case, with use of the logistic regression estimations, we calculated the probability of having P. aeruginosa bacteremia.

Results

Study population. In the 4 participating centers, a total of 4114 unique patient gram-negative rods (GNRs) were isolated from cultures of blood specimens obtained at admission (i.e., ⩽48 h after hospital admission). Escherichia coli was the most common isolate (51%; range, 41%-63%), followed by Klebsiella species (16%; range, 12%-21%), P. aeruginosa (6.8%; range, 5.3%-8.4%), Enterobacter species (4.9%; range, 2.7%-7.8%), and Proteus species (5%; range, 2.7%-6.2%). Serratia, Morganella, and Citrobacter species together contributed 5% of the cases. The incidence of P. aeruginosa bacteremia was calculated to be 5 cases per 10,000 hospital admissions.

For the case-control study, 151 patients with P. aeruginosa bacteremia upon hospital admission (case patients) and 152 subjects with bacteremia due to Enterobacteriaceae upon hospital admission (control subjects) were randomly chosen for inclusion. Blood isolates recovered from control subjects included E. coli, 110 subjects (72%); Klebsiella species, 22 subjects (14%); Enterobacter species, 9 subjects (6%); and Proteus species, 4 subjects (3%). Citrobacter, Serratia, and Providentia species together accounted for isolates in 6 patients (4%). Poly1enta conditional-hyphen>microbial bacteremia (E. coli and K. pneumoniae) was documented in only 1 patient.

Demographic characteristics and comorbidities. The mean age of subjects was 67 years and did not differ between the 2 groups (table 1); however, age >90 years tended to be more prevalent among case patients than control subjects (14 of 151 vs. 6 of 152; OR 2.49; P=.062). Case patients were more often male (55.6% vs. 44.4%; P=.034), and more frequently had poor functional status, as evidenced by dependence on assistance with the activities of daily living (31% vs. 17%; P=.004). Case patients had more comorbidities than did control subjects (mean number of comorbidities ± SD, 2±1.2 vs.1.6±1.1; P=.002). Malignancy was the most common underlying illness among case patients: 47% of case patients had a malignant disease (of which 40% were hematological), compared with 24% of control subjects (P<.001). Nephrolithiasis was the only comorbidity that was significantly more prevalent among control subjects (5.9% vs. 0.7%; P=.02).

Clinical characteristics. Case patients were more often immunocompromised at presentation than were control subjects, mostly because of recent chemotherapy and neutropenia (table 1). Case patients also presented more often with febrile neutropenia than did control subjects (18.6% vs. 2.6%; P<.001).

The source of bacteremia differed between the 2 groups (P<.001) (table 1). Among the control subjects, the urinary tract was the most common focus, with a urinary tract infection documented in 65% of patients. Intra-abdominal infections were next most common focus of infection (14% of control patients), and 10.5% of control subjects had no known focus of infection. Among the case patients, the distribution of sources differed; urinary tract infections were less predominant (23% of patients) as the source of infection than in control subjects. Other foci of infection in case patients included lower respiratory tract infection and intravenous catheter-related infection (in 19% and 15% of patients, respectively). In 22% of case patients, no source of bacteremia was identified. Invasive devices at presentation—specifically, central venous catheters and urinary devices—were more often found among case patients.

Contact with the health care system. Few patients in either group (19 case patients [13%] and 15 control subjects [10%]) underwent an invasive procedure during the 2 weeks preceding the index hospitalization (table 1). Cystoscopy tended to be more prevalent among case patients, and prostatic biopsy tended to be more prevalent among control subjects; however, statistical significance was not reached in group-to-group comparisons for either procedure (P=.06).

Recent hospitalization and prior antimicrobial therapy were more common characteristics of case patients than control subjects. Forty-four percent of case patients had been hospitalized during the preceding 30 days, and 65% had been hospitalized during the preceding 90 days, whereas these figures were 25% and 36.2%, respectively, among control subjects (P<.001). Case patients more often received antimicrobial therapy (mostly β-lactam agents) during the month before hospitalization than did control subjects (36% vs 13.6%; P<.001).

Discriminative model. An initial multivariate logistic regression analysis identified 5 variables as independent predictors of P. aeruginosa bacteremia: severe immunodeficiency (including solid organ or bone marrow transplantation, neutropenia, recent chemotherapy, or treatment with high-dose corticosteroids for >5 days, azathioprine sodium, or cyclosporine), age >90 years, receipt of antimicrobial therapy during the prior 30 days, having a central venous catheter, or having a urinary device.

Because empirical coverage of P. aeruginosa is the standard of care in severely immunocompromised patients with signs of sepsis, we excluded patients with severe immunodeficiency from further analysis. After exclusion of patients with severe immunodeficiency, logistic regression analysis revealed that the same 4 variables (age >90 years, receipt of antimicrobial therapy during the previous 30 days, having a central venous catheter, and having a urinary device) were independently associated with being a case patient. Accordingly, the diagnostic accuracy of the logistic regression model, as given by the area under the receiver-operating-characteristics curves, was 0.726 (moderately good prediction) (table 2).

Among the 250 patients without severe immunodeficiency (109 case patients and 141 control subjects), a prediction model was constructed, and the probability of P. aeruginosa bacteremia was calculated. Subclassification of patients and probabilities based on the total number of independent predictors found for each patient is presented in table 3, including the number of total patients needed to treat to effectively treat a single case of P. aeruginosa bacteremia in each category. We found that the probability of bacteremia due to P. aeruginosa was 2.3% if none of the predictors was documented, 8.84% (range, 7%-14%) in the presence of 1 predictor, 28.1% (range, 21%-47%) in the presence of 2 predictors, and ⩾59% in the presence of 3 or 4 predictors. Having none of the predictors above was associated with a likelihood of P. aeruginosa bacteremia of 1:42, whereas having ⩾2 predictors increased the likelihood to nearly 1:3.

Discussion

Prudent use of antibiotics requires that clinical decision-making take into account both the risk of the individual patient if not covered appropriately for a certain pathogen and the risk associated with overuse of antibiotics—in the context of this study, overuse of anti-pseudomonal agents. Thus, a clinician must consider both the number of patients not covered effectively by a certain antibiotic regimen and the number of patients who are treated but do not have the condition.

In the present study, we found P. aeruginosa bacteremia upon hospital admission to be a rare condition, occurring in 5 of 10,000 hospital admissions. Yet it was the third-leading GNR isolate to cause bacteremia at the time of admission (6.8% of 4114 unique patient cases of GNR bacteremia). Affected patients often had severe immunodeficiency, a malignant condition, recent hospitalization, prior antibiotic treatment, and invasive devices at presentation. Kang et al. [22] reported similar results. In their study, P. aeruginosa caused 4.4% of all community-acquired cases of gram-negative bacteremia, and clinical features of affected patients were alike; in particular, there was a high proportion of patients with cancer (solid tumors, 41%; and hematological malignancy, 18%).

To provide a simple clinical decision-making tool, a discriminative model was created. We intended this tool to be applied when bacteremia is suspected at hospital admission, to help the clinician decide whether to treat low-risk patients (i.e., those without severe immunodeficiency) with an anti-pseudomonal agent. A previous prediction model for P. aeruginosa bacteremia included mostly hospital-acquired infections (91%), and most predictors were related to a long duration of hospitalization [23). Thus, this model is not applicable to patients with bacteremia at the time of admission. We omitted patients with severe immunodeficiency from the model, because anti-pseudomonal coverage for these patients is accepted as standard medical practice [24].

We based the model on the occurrence of 4 predictors of P. aeruginosa bacteremia among patients without severe immunodeficiency: age >90 years, receipt of antimicrobial therapy ⩽30 days before hospital admission, having a central venous catheter, and having a urinary device. We found that the probability of P. aeruginosa bacteremia was 2.3% if none of the predictors was present, almost 9% if only 1 predictor was present, and >28% if ⩾2 predictors were present.

Table 3 provides estimates of risk of P. aeruginosa bacteremia generated by the discriminative model. These may guide the clinician regarding the implications of the choice of empirical therapy. Two scenarios are presented: (1) treatment of only patients in the same risk category as the index patient, and (2) treatment of patients on the basis of the threshold level (i.e., if we administer an anti-pseudomonal agent to a patient in a specific risk category, we should also treat patients in a higher risk category, and conversely for not treating). Treatment of every patient with GNR bacteremia upon hospital admission with an anti-pseudomonal agent will result in unnecessary treatment of 20 patients to cover 1 case of P. aeruginosa bacteremia (number needed to treat, 20). According to the risk stratification model, if no predictor is present, the number needed to treat is 42; if 1 predictor is present, the number needed to treat is 11; and if ⩾2 predictors are present, the number needed to treat is <4. We must remember that, when empirical therapy is instituted, the presence of GNR bacteremia is suspected, but in only a minority of cases (likely <10%) does GNR bacteremia truly exist. Thus, the numbers needed to treat for suspected GNR bacteremia are in reality much higher than those reported by us. Nevertheless, as shown in table 3, patients can be divided into 3 risk groups with respect of to P. aeruginosa bacteremia: patients with ⩾2 predictors are at high risk, patients with no predictors are at extremely low risk, and those with 1 predictor are at intermediate risk. In our study, 96% of the total population of patients with gram-negative bacteremia belonged to the lower-risk groups (i.e., they had <2 predictors). Because the number needed to treat for this group of patients is 25, not treating them will spare anti-pseudomonal treatment for 96% of patients who are admitted to the hospital with GNR bacteremia while taking the risk of 1:25 that 1 case of P. aeruginosa bacteremia will be missed.

In several studies, inadequate empirical treatment of P. aeruginosa bacteremia has been associated with increased mortality [10, 25]. However, others have not found this association [26]. The difference in study results may be related, at least in part, to differences in the patients studied; delay in effective therapy was linked to increased mortality, mostly in patients with hematological malignancies and neutropenia. Overuse of anti-pseudomonal agents also has unwarranted consequences—mainly, the emergence of resistance in P. aeruginosa isolates with subsequent limited treatment options and adverse outcomes [15, 27]. The question of when to cover and when not to cover P. aeruginosa remains a clinical decision that has to take into account various parameters, including the severity of the infection, the risk imparted to the patient by delaying appropriate therapy, and other medical and social parameters. Such decisions go beyond the consideration of P. aeruginosa and include the need to cover gram-positive bacteria, various drug-resistant organisms, and nonbacterial pathogens. However, we believe that the model presented here provides support for the decision not to administer anti-pseudomonal agents to patients with no predictors and to treat those with ⩾2 predictors. For patients with 1 predictor, we recommend consideration of the risks and benefits on an individual basis.

Our study has several limitations: first, it is a retrospective study; thus, some of the data recorded in the medical charts may have not been complete. We do not believe that this is a major problem, because the data included in the study are routinely recorded in the patient chart. Second, the study represents 4 tertiary care centers, and the results may not be generalizable to non-tertiary care centers. On the other hand, the international character of this study is a clear strength that makes it more easily generalizable. We recommend additional study of this topic and validation of the model that we have suggested in additional settings. Third, and most importantly, our study refers to patients with documented GNR bacteremia, and the true question that the clinicians face regards all patients with suspected bacteremia: who is at risk of having P. aeruginosa bacteremia. We believe that future studies should focus on this question and that our model will be validated in a prospective group of patients who present with infection.

On the basis of the model we have created, the clinician can estimate the probability of P. aeruginosa bacteremia for each clinical case of suspected gram-negative bacteremia, and make an educated decision whether an anti-pseudomonal therapy should be included in the empirical therapy.