INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Primary nosocomial bloodstream infections (BSI) are associated with excess mortality and hospital costs. The crude mortality for nosocomial bloodstream infections, estimated from a recent review of 3,077 patients, appears to have decreased from 51% in 1981 to 29% in 1992 (1). This estimate included all hospitalized patients, not just those in the ICU. Specific studies investigating the crude mortality for bloodstream infections in the ICU have estimated the crude mortality to be approximately 56%, ranging from 31.5 to 82.4% (2). Other studies have estimated that there are close to 4,500 deaths directly caused by these infections yearly (6). Furthermore, primary bloodstream infections are felt to increase length of stay (LOS) by 7.4 d, and increase charges by $3,517 per episode (1992 dollars) (6).

A limitation of studies investigating the attributable mortality of BSI is that a major risk factor for BSI is severity of illness. Severity of illness is a risk factor for mortality as well as longer LOS. To address this problem, Haley (7) has suggested that measures of severity be used to avoid confounding.

We planned to more definitively measure the impact of BSI on mortality and LOS after controlling for severity of illness using Acute Physiology and Chronic Health Evaluation (APACHE) III. APACHE III has been developed using a large number of patients (17,440) and provides predictions concerning ICU and hospital mortality on admission to the ICU and on a daily basis (8). It provides a prediction of hospital and ICU length of stay on admission. It has been validated in a number of different ICU settings for a number of different admitting diagnoses (9). The large number of admissions in the APACHE database provides accurate predictions for a wider variety of diagnoses and daily predictions of mortality. After matching for predicted ICU mortality and controlling for other confounders, we could then truly ask whether BSI affects mortality and LOS in the ICU.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Study Population

The University of Michigan Hospital is an 886-bed hospital that serves as a primary care and referral facility. The medical center contains a 20-bed medical intensive care unit (MICU) that receives approximately 1,000 admissions per year. This facility is staffed by full-time University faculty. The MICU has been equipped with APACHE III since January 1, 1994.

During the study period (January 1, 1994 to December 31, 1996), all patients admitted to the MICU had daily APACHE III scores generated for them. These data were available to clinicians to assist in management of patients at the time of their admissions. The database contains age, sex, race, date of admission to the hospital and ICU, date of discharge from the ICU and hospital, ICU and hospital discharge status (alive or dead), and any chronic health comorbidities that have been shown to affect mortality (these include: history of cancer, immunosuppression, AIDS, and liver disease). The database also contains the daily acute physiology score, the daily APACHE III score, the daily predicted ICU and hospital mortality, and the predicted ICU and hospital length of stay on admission.

The hospital also has an active surveillance program for primary nosocomial bloodstream infections in the ICU. Centers for Disease Control and Prevention (CDC) definitions were used to identify all cases of primary nosocomial bloodstream infections in the MICU between January 1, 1994 and December 31, 1996 (15). These definitions specify that a laboratory-confirmed bloodstream infection must meet one of the following criteria.

1. 1. Recognized pathogen isolated from blood culture and pathogen is not related to infection at another site. OR
2. 2. One of the following: fever (> 38° C), chills, or hypotension AND any of the following: (1) common skin contaminant isolated from two blood cultures drawn on separate occasions AND the organism is not related to infection at another site; (2) common skin contaminant isolated from blood culture from patient with intravascular access device AND the physician institutes appropriate antimicrobial therapy; (3) positive antigen test on blood AND organism is not related to infection at another site.

Process of Matching

Matched control patients were chosen as close as possible to the date of infection so as to avoid intervening changes in severity, which can confound the relationship between BSI and mortality. To determine if the patient would begin to manifest changes in severity because of the infection prior to the day that the positive blood cultures were drawn, we investigated the trend of the acute physiology scores (APS) for the 4 d preceding the day of infection for each patient. Using this data (shown in RESULTS), patients were matched based upon the predicted mortality on the day immediately prior to infection (Day 1).

Each infected patient was matched to a noninfected control patient, defined as a patient who had no evidence of primary nosocomial bacteremia anytime during his ICU stay. Matching was done for the following criteria: predicted mortality on Day 1 (± 10%), sex, age (± 10 yr), race, length of stay prior to the day of matching (± 1 d or 33% of the case's length of stay), admission during the study period, admitting diagnosis (or diagnostic group), and chronic health. The first priority was to match each patient according to predicted mortality, then to match on as many of the other criteria as possible. When multiple matches were available, the uninfected patient who was admitted closest in time to the infected patient was used. We derived a scoring system to assess the quality of matching. Because we felt that matching for severity of illness was the highest priority, we scored a match on this variable as worth five points. Matches on all other variables were worth one point.

Methods of Analysis

We used conditional logistic regression and paired t tests to investigate ICU mortality and LOS, respectively (16). In terms of death in the ICU, conditional logistic regression was used to test whether BSI was associated with increased mortality in the ICU. As all major potential confounders were matched for, they were not included in the model unless they were significantly different between the two groups. We did similar analyses using hospital discharge status (alive or dead) as the outcome measure.

Expenditures in the ICU were assessed using both excess length of stay and direct costs. Length of stay was defined as the time from the infection (or the equivalent day in the uninfected group) until discharge from the ICU or the hospital. Direct costs were obtained from a decision support information system from Transition Systems, Inc. (TSI, Boston, MA). The TSI system uses a weighting system (relative value units) that allows costs to be allocated to each charge code/service within a department. The costs are based on the actual expenses incurred by each department for the time period of the service.

The TSI database does not contain itemized daily costs for each patient, but does include total admission costs ("admcost") and ICU nursing costs ("rncost"). We therefore calculated a ratio of "rncost" to "admcost" for a subset of patients in our study who spent their entire stay in an ICU. This ratio allowed an estimate of total ICU costs to total ICU nursing costs. We then estimated the total direct costs in the ICU for the entire stay for each patient by multiplying the patient's "rncost" by the ratio generated above.

As death can artificially decrease length of stay, the analyses of length of stay and cost were done with all patients and also among the 35 pairs in which both the infected and the uninfected patients survived to ICU and/or hospital discharge. Also, for the cost analysis, some of the patients were excluded because they were in the ICU prior to the time we began collecting and storing TSI cost data (July 1, 1994).

The length of stay data and the cost data were not normally distributed, and, therefore, Wilcoxon's signed rank statistic was used to evaluate whether these variables were different from zero (17). All statistical tests were done using an alpha level of 0.05. The analyses were done using the Statistical Analysis Software Package (SAS Institute, Cary, NC).

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Between January 1, 1994 and December 31, 1996, there were 3,003 MICU admissions. Sixty-eight patients developed 72 episodes of primary nosocomial bloodstream infections. Four of the patients had two episodes of BSI. The second episode for each patient was excluded from study, and 68 cases formed our "infected" group. This group included eight patients with polymicrobial bacteremia. The infections occurred on average on Day 10.4 of ICU care (median, 5.5 d; range, 2 to 61 d). The most common pathogens are shown in Table 1. Note that the data are compared with national data obtained through the National Nosocomial Infections Surveillance (NNIS) report from the CDC (18).

A preliminary analysis of the trend of the APS for the 4 d preceding the day of infection was performed for each patient. As can be seen in Figure 1, there is no change in the mean APS until the day of infection. The lack of evidence for a change in severity before the day of infection allowed us to match the patients based upon the predicted mortality on the day immediately prior to infection (Day 1).

View larger version (51K): [in this window] [in a new window]   Figure 1.   Acute Physiology Scores for patients with NB for the 4 d prior to and the day of infection. The graph shows that the severity of illness did not appreciably change prior to the day the positive blood cultures were drawn.

The demographic variables of the study populations and the success of matching on severity are shown in Tables 2 and 3, respectively. The infected and uninfected groups are closely matched for age, sex, comorbidities, and admitting diagnoses. The average matching score was 11.4, with 28 of 68 pairs having a perfect score, 37 of 68 pairs having a score of 11, and three of 68 pairs with a score of 10. Some mismatch for race occurred, but it did not confound the relationship between BSI and ICU or hospital outcomes. Despite the fact that 64 of 68 of the pairs were matched on prior length of stay, there was a significant difference in length of stay (9.4 versus 7.6 d, p = 0.004), as shown in Table 3. However, there was no difference between the groups in terms of APS, APACHE III score or predicted ICU mortality on the day of matching. Furthermore, although the groups were not matched on the day of admission, it turned out that the measures obtained upon admission (APS, APACHE III, predicted ICU and hospital mortality, and predicted ICU and hospital LOS) were also all closely matched between the two groups. To evaluate the trend in severity of illness further, we compared the acute physiology scores for both cases and control subjects on the first 3 d of ICU admission as well as on the 5 d leading up to the day of infection. It can be seen in Figure 2 that these scores were similar and were not statistically different on any day (p > 0.2).

View larger version (46K): [in this window] [in a new window]   Figure 2.   Acute Physiology Scores for all patients on the first 3 d in the ICU and for the 5 d prior to and the day of infection. Data are included only when information was available for both members of a pair. Therefore, the number of pairs included for each day is shown below the figure. The graph shows that the severity of illness is not significantly different between the two groups on any day (p > 0.2).

The crude ICU mortality data are shown in Table 4A. The crude mortality among the patients with BSI was 35.3%, and the crude mortality among the matched patients without BSI was 30.9% (risk ratio, 1.33; 95% CI, 0.56 to 3.16; p value = 0.51). As noted above, there was discrepancy in the prior length of stay between the two groups. Therefore, further analysis was done including length of stay in the model. When length of stay was included in the model, the risk ratio (RR) of death associated with primary bacteremia changed from 1.33 to 0.77 (95% CI, 0.26 to 2.27). However, length of stay was not a significant variable in the regression (p = 0.12) and therefore not included in the final model. For hospital mortality (which was not a primary end point), the overall RR was 1.0 with a 95% CI of 0.43 to 2.31, p > 0.99. Death rates varied according to the organism causing the BSI (Figure 3). Patients infected with gram-positive cocci had the lowest ICU mortality at 25.5%, and patients infected with yeast had the highest mortality at 72.7%. Differences between the three groups were statistically significant (p = 0.013), but there was no difference in mortality between the infected and uninfected groups when stratified by type of organism (Table 4B-E).

View larger version (50K): [in this window] [in a new window]   Figure 3.   Predicted and actual ICU mortality for the three organism types. Note the change in predicted mortality for the group with yeast infections between the time of admission until the day prior to the infection.

The ICU LOS, hospital LOS, and costs were compared among all 68 pairs as well as the 35 pairs in which both the infected and the uninfected patients survived the ICU stay. The results are shown in Table 5. There was significantly longer LOS and cost in the infected group, both when all patients were analyzed as well as when only the survivors were analyzed. Further analysis, including LOS prior to matching, did not appreciably change these outcomes and therefore is not shown. Overall, BSI results in 5 d of excess stay in the ICU resulting in at least $16,000 of excess costs per episode.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Several studies show an attributable mortality associated with BSI in both hospitals and intensive care units (2, 4, 19). Only three prior studies have attempted to measure the attributable mortality caused by these infections in the ICU. Forgacs and colleagues (5) reviewed the experience in an ICU over a 15-yr period. The mortality was 60.4% in patients with bloodstream infections compared with 13.1% in those without detectable bloodstream infections. Smith and colleagues (4) attempted to control for severity of illness by matching patients based upon predicted mortality. These investigators matched the 34 patients who developed bloodstream infections with another group with the same predicted mortality on admission, based upon their APACHE II score. Although both groups had a predicted mortality of 53% upon admission to the ICU, the bacteremic group had an actual mortality of 82.4%, whereas the nonbacteremic group had an actual mortality of 52.9%, yielding an attributable mortality of approximately 29% (4). Finally, Pittet and colleagues (2) performed a matched cohort study for a group of surgical ICU patients. In this study, the mortality rate in the patients with bloodstream infections was 50%, whereas in the control subjects it was 15%, yielding an attributable mortality of 35% (2).

Studies attempting to assess the attributable mortality of BSI can be confounded by severity of illness. To attempt to remove this confounding, Pittet and colleagues (2), in their study in a surgical ICU, instituted individual matching. Specifically, they matched on admitting diagnosis (DRG group), the number of discharge diagnoses, sex, age, and length of stay prior to infection. They state that the matching on admitting diagnosis and number of discharge diagnoses has been shown to be an adequate surrogate measure of severity of illness (7). Although this may be true, the episode of bacteremia could influence the number of discharge diagnoses, thereby confounding the relationship.

It has been suggested that systems such as APACHE are better suited to control for severity of illness in studies assessing the attributable mortality of BSI (7). In this light, Smith and colleagues (4) used the APACHE II predicted mortality to attempt to control for confounding. Their study is limited in two respects. First, the investigators made no attempt to match their patients on any other criteria such as sex, race, or age. Although it is not clear that this is entirely necessary, it would limit the face validity of the study if, for example, the cases were all older than the control subjects. Our main critique of this study is that the patients were matched on admission to the ICU. Given that the infection occurred anywhere from Day 1 to Day 87 (median, Day 8), the severity of illness could have changed prior to the infection, as it did for the group of patients with fungemia in our study. Therefore, the development of BSI could have been a marker of severity rather than an independent risk factor for mortality.

Our study has shown that, after appropriate matching, we could not detect an association between the development of BSI and increased ICU mortality. However, these infections are associated with an increased length of stay and the expenditure of more resources. The first conclusion, that BSI may not have an attributable mortality, is important in that it calls into question the assertion that as many as 4,500 people are dying yearly because of bloodstream infections (6). Trying to gauge the public health implications of nosocomial infections requires sound evidence concerning their attributable mortality. This study contradicts prior conclusions that BSI is associated with high attributable mortality. Our study does strongly support the public health importance of BSI by delineating the large costs associated with this infection. Strikingly, it does so in a study that was tightly controlled for severity of illness.

We believe that the attributable mortality found in prior studies may have been influenced by study design and that the null result in our study is due to more appropriate matching. Nevertheless, there are some potential limitations of our study that may have influenced our results.

One potential limitation would be that the groups were overmatched. Overmatching can limit one's ability to recognize a true difference between groups by either reducing the validity or the statistical efficiency of a case-control comparison (25). The main threat to validity in this study would be if the APACHE III score that was used for matching was changed because of the presence of BSI. By evaluating the trend in APACHE III scores during the 4 d prior to the day of infection (Figure 1), we were able to show that the episode of BSI did not change the APACHE III score until the day of infection. Also, there is the problem of reduced statistical efficiency. Overmatching can decrease power without improving accuracy. However, we do not think that a decrease in statistical efficacy was the main reason that we failed to find a difference between the two groups. The point estimate for relative risk for ICU mortality in this study was 1.33 and the "attributable" mortality was only 4.4%. This small difference is not in line with differences that were estimated in prior studies of ICU mortality secondary to BSI (attributable mortalities of 29 and 35%) (2, 4). Therefore, we do not feel that decreased power was a major factor in the null result of this study.

Another potential limitation of this study could be that the patients with BSI in this study were not representative of ICU patients with BSI in the general population or in prior studies. In our study, we limited the cases to patients with primary bacteremia. That is, we did not include as cases patients who developed bacteremia from another source such as pneumonia or urinary tract infection. Both the study by Pittet and colleagues (2) and the study by Forgacs and associates (5) included patients with secondary bacteremia. Limiting our study to primary bacteremias could decrease the likelihood of finding a difference between the two groups as other researchers have shown that secondary bacteremias have a higher mortality than primary bacteremias (26, 27). However, we feel that primary and secondary bacteremias are distinct clinical syndromes and should be studied separately. Otherwise, data that are related to secondary bacteremias will be used to support initiatives to decrease the incidence of primary bacteremias. In fact, recent studies evaluating the utility of antibiotic impregnated catheters have used data from the study of Pittet and colleagues to support the cost-effectiveness of these catheters (28, 29). Certainly, the use of special catheters will not decrease the incidence of secondary bacteremias, and therefore the utility of those catheters can only be judged using studies assessing the cost of primary bacteremias. Hence, the rationale for our study design.

Finally, our study could be limited by inadequately adjusting for confounders. A case-matching approach will never be able to match perfectly on more than a few characteristics. It can be seen in Tables 2 and 3 that the groups were well matched on all factors expect for prior length of stay. When accounting for this increased prior length of stay in the cases, we find that there is a lower mortality rate in the cases. However, prior length of stay is not statistically significant in this model and, therefore, we have presented our results primarily in an unadjusted form.

We attempted to include all the relevant variables in the matching criteria. However, it is possible that lack of matching on some unmeasured characteristic may have biased the results of our study. The bias could have caused us to either overestimate or underestimate the attributable mortality of primary bacteremia. This limitation is unavoidable given our research methods.

It is important to stress that the finding of a low attributable mortality in primary bacteremias does not seem to be due to a high prevalence of coagulase-negative staphylococci. Prior studies have shown that the coagulase-negative staphylococcal bacteremia has a relatively low attributable mortality, whereas fungemia is associated with an attributable mortality of 38% (23, 24). However, in our study, there does not seem to be much difference in the attributable mortality when broken down by causative organism (Table 4). Therefore, we feel that our results are no less valid for fungemia than they are for coagulase-negative staphylococcus, and we can conclude that all primary nosocomial bloodstream infections have little, if any, attributable mortality, but they do have a high cost associated with them.