Methods

Study population. All patients who underwent coronary artery bypass procedures (alone or with another cardiac procedure) at Barnes-Jewish Hospital (St. Louis, MO) from 1 January 1996 through 31 December 2004 and survived until the second postoperative day were included in the study population. Demographic characteristics, comorbidities, preadmission medication use, and operative variables were obtained from the Washington University Adult Cardiac Society of Thoracic Surgeons database. Laboratory results and microbiologic data were obtained from the Medical Informatics database. This study was approved by the Washington University School of Medicine Institutional Review Board.

Ascertainment of BSI and mortality. The surgical admission and readmission medical records were reviewed for all patients with positive blood culture results ⩽90 days after undergoing a surgical procedure, to determine whether the patient met criteria for having a BSI. The designation of BSI required ⩾1 blood culture positive for a known bacterial or fungal pathogen. Diagnosis of BSI caused by a potential skin contaminant (coagulase-negative staphylococci or Streptococcus viridans group) required isolation of the same organism from ⩾2 blood cultures or 1 positive blood culture result for a patient for whom only 1 blood culture was performed. Diagnosis of BSI caused by potential skin contaminants also required a temperature >38°C or <36°C, chills, or hypotension (systolic blood pressure <90 mm Hg) and antimicrobial therapy temporally associated with the positive blood culture result [7, 8]. Polymicrobial BSI was defined as isolation of pathogens from ⩾2 different etiologic categories in the same blood culture(s) within 48 h or isolation of 1 known pathogen and ⩾1 potential skin contaminant from ⩾2 blood cultures.

The Society of Thoracic Surgeons coordinator tracked mortality ⩽30 days after surgery by reviewing all Barnes-Jewish Hospital data and calling the referring clinics to obtain 30-day follow-up data. We developed a 2-step algorithm for 90-day mortality. If there was no mortality date available, patients with any readmission to Barnes-Jewish Hospital >90 days after surgery were considered to be alive at 90 days. If no readmissions to the hospital >90 days after surgery were identified, the Social Security Death Index was used to determine whether the patient died ⩽90 days after surgery, using their name and Social Security number for identification [9].

Statistical analyses. All analyses were performed using SPSS, version 14.0 (SPSS), and SAS, version 9.0 (SAS). Multivariable Cox proportional hazards models were developed in forward stepwise fashion, including all variables with P<.20 in the univariate analyses or related to mortality associated with coronary artery bypass surgery in the literature. All relevant 2-way interactions were tested after selection of the main effects [10]. Interactions were included in the final models only if they were statistically significant (P<.05). Any BSI occurring ⩽90 days after surgery was included in all multivariable models as a time-dependent variable, using the interval from the date of surgery to the date of collection of the first blood sample with a positive culture result as the time to BSI [11, 12]. In this method, for each patient who developed a BSI, the dichotomous BSI variable was equal to 0 before diagnosis of BSI and was changed to 1 the day of diagnosis of BSI.

Propensity scores were also used to test the association between patients having ⩾1 episode of BSI and mortality. A logistic regression model was created to predict the likelihood of BSI developing ⩽90 days after surgery, including all variables with clinical relevance, regardless of statistical significance. The predicted probabilities of developing BSI from the logistic regression model were categorized into quintiles, with each quintile having an equal number of patients with BSI but a varying number of control subjects. The groups were categorized with equal numbers of patients with BSI, because quintiles based on the predicted probabilities within the entire population resulted in a highly skewed distribution of patients with BSI across groups. An extended Cox model was then created for each propensity score quintile, with the time-dependent BSI variable and unequally distributed covariates as the only predictors in the models.

Results

Demographic characteristics of the 4561 persons who underwent coronary artery bypass procedures at Barnes-Jewish Hospital from 1996 through 2004 are shown in table 1. The number of surgical procedures performed per year decreased steadily during the 9-year period, in keeping with national trends. A total of 4515 patients survived at least until the second day after the surgical procedure was performed and were included in subsequent analyses. Because identification of BSI depended on recognition of signs and symptoms of infection that resulted in collection of blood samples for culture, the 46 excluded patients who died <48 h after surgery would likely not have had the opportunity to develop BSI and have a diagnostic examination performed.

During the 9-year study period, 151 patients (3.3%) developed 184 episodes of BSI ⩽90 days after surgery. One hundred twenty-eight (85%) of the patients who had a BSI had 1 episode, with a maximum of 5 episodes of BSI in 1 patient. The number of episodes, according to etiologic agent(s), is shown in table 2. The incidence of BSI ⩽90 days after surgery varied from 2.2% to 4.9%, without apparent trend.

Univariate analyses of mortality. Mortality 2–90 days after surgery varied from 8.8% in 1996 to 5.0% in 1997 (P=.566, by χ² test for trend). A total of 303 patients (6.7%) were lost to long-term follow-up. Patients lost to follow-up had a lower body mass index (calculated as the weight in kilograms divided by the square of the height in meters), were more likely to have congestive heart failure and arrhythmia, and were more likely to be white, compared with patients who received follow-up care at Barnes-Jewish Hospital or had a recorded death date (for all variables, P<.05). Patients lost to follow-up were also marginally less likely to die 2–90 days after surgery than were patients who were not lost to follow-up (12 [4.0%] of 303 patients vs. 269 [6.4%] of 4212 patients; P=.091).

Variables associated with mortality in univariate extended Cox models are shown in tables 3 and 4. Women had a higher risk of mortality and were significantly older at the time of surgery than men (mean age, 66.9 years vs. 64.8 years; P<.001). There was a backwards J-shaped relationship between body mass index and mortality, with the highest risk of mortality among persons with a body mass index <18.5 (table 3). Mortality was significantly higher after valve procedures, repeat operations, and nonelective surgical procedures. There was a trend toward increasing mortality associated with lower intraoperative core temperature (table 4).

The univariate results for BSI and 2–90-day mortality are shown in table 3. At any given time, the hazard for a person who developed a single episode of BSI was 18.5 times the hazard for a person who had not developed a BSI by that time. Likewise, the hazard for a person who had multiple episodes of BSI was 11.4 times the hazard for a person who had not developed a BSI by that time.

BSIs were also categorized according to etiology, with the categories consisting of single episodes of BSI due to gram-negative bacteria, S. aureus (methicillin sensitive and methicillin resistant), gram-positive bacteria other than S. aureus, yeast, and multiple pathogens (i.e., polymicrobial BSI) and multiple episodes of BSI. The hazard for mortality at any given time varied from 10-fold for S. aureus BSI to >45-fold for BSI due to yeast (table 3). The crude mortality rate associated with single episodes of BSI was 19% for BSI due to S. aureus (7 of 37 patients died), 25% (8 of 32) for BSI due to gram-positive bacteria other than S. aureus, 42% for BSI due to gram-negative bacteria (14 of 33), 75% for BSI due to yeast (9 of 12), and 50% (7 of 14) for a single episode of polymicrobial BSI. The crude mortality rate among patients who had multiple episodes of BSI ⩽90 days after surgery was 26% (6 of 23 patients died).

Multivariable analyses of mortality within 90 days after coronary artery bypass surgery. Parsimonious models were developed to determine the hazard of mortality associated with having any BSI (⩾1 episode), with BSI included as a time-dependent covariate. BSI was associated with a significantly increased risk of mortality (hazard ratio at any given time, 4.19; 95% CI, 2.97–5.92; P<.001). A separate model was developed with BSI categorized in 4 groups, as follows: no BSI ⩽90 days after surgery, 1 episode of gram-negative bacteremia, 1 episode of S. aureus bacteremia, or 1 episode of BSI caused by gram-positive bacteria other than S. aureus. The categories of yeast BSI, polymicrobial BSI, and multiple episodes of BSI were not included because of the heterogeneity and relatively small numbers in each group. Gram-negative and S. aureus bacteremia remained significantly associated with an increased risk of mortality (table 5). The hazard ratio for mortality associated with BSI caused by gram-positive bacteria other than S. aureus was much lower after controlling for other risk factors of mortality. Of the 37 patients who had 1 episode of S. aureus BSI, 18 (48.6%) also had mediastinitis. Female sex did not remain statistically significantly associated with mortality in the model, after adjusting for the BSI categorical variable and other covariates (P=.954). The results did not change after excluding patients who were lost to long-term follow-up (table 5).

Propensity score analysis. A propensity score model was created to test the association between a patient having ⩾1 episode of BSI and mortality. The predicted probabilities of developing a BSI ⩽90 days after surgery were categorized into quintiles with equally distributed numbers of patients with BSI within the quintiles. Comparisons of all relevant covariates and BSI were performed to verify that the covariates were balanced, with few exceptions (table 6). The results, stratified by quintile of propensity score risk to develop BSI, are presented in table 6. The hazard ratio for mortality at any given time was highest within quintile 1 (lowest propensity to develop BSI) and was lowest within quintile 5 (highest propensity to develop BSI). The hazard ratio for mortality was similar within the middle 3 quintiles. The models were reanalyzed after exclusion of the patients who were lost to long-term follow-up, with no appreciable difference in results (table 6).

The distribution of the microbial etiology of BSI was examined according to propensity score quintiles. Of the 37 patients with S. aureus BSI, 15 had calculated probabilities of developing BSI that placed them in the lowest propensity group, significantly more than would be expected if the patients with S. aureus BSI were evenly distributed across quintiles (P=.009, by χ² test). In contrast, the distribution of patients with gram-negative BSI, BSI caused by gram-positive bacteria other than S. aureus, yeast BSI, and multiple episodes of BSI did not differ from the expected 20% in each quintile.

Discussion

In this cohort of patients who underwent coronary artery bypass procedures, crude mortality rates were significantly higher among patients who developed BSI than among patients who did not develop BSI. BSI due to gram-negative bacteria and S. aureus BSI were associated with a higher risk of mortality, compared with BSI caused by other gram-positive bacteria. Using propensity scores and extended Cox proportional hazards models, the risk of mortality 2–90 days after surgery was highest among patients who developed BSI but had the lowest risk of developing BSI. In contrast, the mortality associated with BSI was lowest among patients who had a very high risk of developing BSI.

In the multivariable Cox model, BSI caused by gram-negative bacteria and BSI caused by S. aureus were each associated with a 7-fold increased risk of mortality 2–90 days after surgery, compared with no BSI. In contrast, BSI caused by gram-positive bacteria other than S. aureus (predominantly vancomycin-resistant enterococci and coagulase-negative staphylococci) was associated with only a 2.2-fold increased risk of mortality, after adjusting for other covariates. This finding is consistent with reports in the literature of increased mortality associated with BSI due to gram-negative bacteria [5, 13] and lower attributable mortality associated with BSI due to gram-positive bacteria, such as vancomycin-resistant enterococci and coagulase-negative staphylococci [5, 14, 15]. The relatively high adjusted risk of mortality associated with S. aureus BSI in our study differs from the results of the study by Blot et al. [16]. In their study, the mortality associated with methicillin-susceptible S. aureus BSI was only 6%, whereas the attributable mortality of methicillin-resistant S. aureus BSI was 23%. In our study, 65% of the monomicrobial S. aureus BSIs were due to methicillin-susceptible S. aureus. One possible explanation for the disparity in results is that the study by Blot et al. [16] was conducted in an intensive care unit. Almost one-half of the patients in our study who had S. aureus BSI also had mediastinitis, which increases the risk of mortality [17, 18].

Most published studies that have described mortality associated with BSI have focused on intensive care unit populations. Although the incidence of BSI is high among these severely ill patients, the overall mortality rate also tends to be high, making it difficult to determine attributable mortality. In this study, patients at lowest risk of developing BSI had the highest risk of mortality if a BSI occurred ⩽90 days after surgery. Only 2% of uninfected control patients at lowest risk of developing BSI died, compared with 13% of such patients who developed postoperative BSI. A disproportionate percentage of patients in this lowest propensity score quintile had S. aureus BSI, consistent with the finding in the parsimonious model of increased risk of mortality associated with S. aureus BSI. More than 85% of persons with S. aureus BSI in the lowest risk quintile had a chest surgical site infection as the source of their positive blood culture result. Thus, development of BSI associated with surgical site infection and BSI in less severely ill patients may be particularly hazardous.

In contrast to the very high hazard ratio for mortality among patients in the lowest propensity score quintile, patients at highest risk of BSI who developed BSI ⩽90 days after surgery had a much lower attributable mortality. In this highest risk quintile, patients were almost equally likely to die, regardless of whether they developed BSI. This finding is consistent with previous reports of little or no attributable mortality associated with BSI in intensive care unit patients [2, 4, 19]. Our finding is also consistent with the outcomes of nosocomial BSI in intensive care unit patients that were reported by Kim et al. [20]. In their study, patients were stratified into 2 groups on the basis of Acute Physiology and Chronic Health Evaluation II scores at intensive care unit admission. BSI was associated with a 2.4-fold increased risk of mortality among patients with a low severity of illness but was not associated with mortality among patients with high severity of illness at intensive care unit admission. We found a 10-fold higher risk of BSI–associated mortality among patients who underwent coronary artery bypass procedures and had a low risk of infection and a 2-fold increased risk of BSI–associated mortality among the most severely ill patients. More than 60% of the BSIs reported in the study by Kim et al. [20] were catheter related, compared with 40% in our study. Most investigators have failed to demonstrate significant attributable mortality associated with catheter-related BSI [2, 5, 21, 22]. Thus, the lower attributable mortality in the study by Kim et al. [20] may, in part, be attributable to the predominance of catheter-associated BSI among their population.

Previous studies of mortality after coronary artery bypass surgery that used the Society of Thoracic Surgeons database have focused on preoperative or operative risk factors rather than on postoperative complications. In addition, the focus in the literature has been on operative or 30-day mortality rather than longer-term mortality. Nonetheless, many of the noninfectious risk factors for mortality that we identified, such as peripheral vascular disease, dialysis, use of an intra-aortic balloon pump, and congestive heart failure, have also been reported to increase the risk of operative death [23, 24]. Interestingly, female sex was not associated with increased risk of mortality in the multivariable Cox model, although this variable was independently associated with mortality in several previous reports [25–27]. The increased risk of mortality among women has been attributed to higher severity of illness at the time of surgery [28], increased prevalence of infection [29], or receipt of blood transfusions [30]. Consistent with the findings of Rogers et al. [30], we found that female sex was no longer significantly associated with mortality, after controlling for perioperative blood transfusions. Women in our study were more likely to receive blood transfusions and received a greater mean number of blood units, compared with men. There was also a clear dose response between the number of transfused units and increased risk of mortality. Thus, our results are in agreement with the finding of Rogers et al. [30] that female sex is not independently associated with mortality after coronary artery bypass surgery when the excess transfusion rates among women are taken into account.

As with all hospital-based observational studies, we could only identify BSI in patients who developed infection at our hospital. Patients who did not have an identified BSI may have developed BSI after hospital discharge and been readmitted to another institution for therapy. We identified a small number of patients lost to long-term follow-up for BSI status, because they had no additional visits to our hospital. Exclusion of these patients had no statistically significant effect on the results. Therefore, it is unlikely that the inability to exclude diagnosis and treatment of BSI at other institutions had a significant impact on our findings.

This study has significant strengths, including rigorous confirmation and validation of BSI using standard criteria and verification of mortality. This study was strengthened by the robust Society of Thoracic Surgeons database, which includes a large quantity of information on potential risk factors for mortality. We performed extensive logic checks with manual medical record review to verify the Society of Thoracic Surgeons data and supplemented such data with laboratory data, such as serum albumin level, which is an important measure of malnutrition and severity of illness.

We used rigorous statistical methods to determine the hazard of mortality associated with BSI, including extended proportional hazards models. In most previous studies, BSI had been treated as a dichotomous variable, without taking into account time to infection. Treating BSI as a time-dependent covariate allows for the divergence of survival curves once BSI has occurred (analogous to the risk of mortality increasing after—but not before—diagnosis of BSI). This type of model is more relevant clinically and results in more accurate hazard ratio estimates.

Using 2 statistical methods, we determined that BSI after coronary artery bypass surgery is associated with a significantly increased risk of mortality. The risk of mortality was highest among patients with the lowest likelihood of developing BSI and was lowest among severely ill patients. Mortality was higher among patients with S. aureus or gram-negative bacteremia and lower among patients with BSI caused by gram-positive bacteria other than S. aureus. Interventions focused on preventing postoperative infections could significantly reduce long-term mortality, especially among the healthiest patients.