Hospital-associated infection is well recognized as a patient safety concern requiring preventive interventions. However, hospitals are closely monitoring expenditures and need accurate estimates of potential cost savings from such prevention programs. We used a retrospective cohort design and economic modeling to determine the excess cost from the hospital perspective for hospital-associated infection in a random sample of adult medical patients. Study patients were classified as being not infected (n = 139), having suspected infection (n = 8), or having confirmed infection (n = 17). Severity of illness and intensive unit care use were both independently associated with increased cost. After controlling for these confounding effects, we found an excess cost of $6767 for suspected infection and $15,275 for confirmed hospital-acquired infection. The economic model explained 56% of the total variability in cost among patients. Hospitals can use these data when evaluating potential cost savings from effective infection-control measures.
Nosocomial, or hospital-associated, infections (HAIs) figured prominently in the 2000 Institute of Medicine patient safety report . Approximately 5% of hospitalized patients experience an HAI, and hospitals are important reservoirs for strains of bacteria resistant to antimicrobial drugs [2–11. There are effective interventions for reducing the occurrence of HAI, but like many events in complex systems, they require changes throughout the hospital and come with substantial costs [7, 11–22]. At a time when health care systems must monitor expenditures closely, the demonstration of immediate health benefits or cost savings is often a prerequisite for convincing administrators to support control measures [13, 17, 20, 23–25].
Important progress has been made in measuring how the development of HAIs affects the cost of patient care [8, 15, 26–28]. The cost of caring for patients with HAI has been compared with the cost of caring for control subjects without HAI who were matched for factors (such as severity of illness, diagnosis, and comorbidities) that may confound the measurement of cost [9, 29–33. However, a strategy that matches patients requires careful selection of clinical data to establish parity between patients with HAI and control subjects. Patients with indeterminate or nonmatching criteria are often excluded. Other studies have used concurrent analyses to measure the excess resource use and cost of treating HAIs [2, 25, 26, 34, 35]. The excess cost of treatment of HAIs is compared with the hypothetical cost for the same patients had they not developed HAIs. Many studies have focused on patients in intensive care units (ICUs), a population that represents the upper extremes in both severity of illness and cost [9, 28, 30–32, 36]. Others have concentrated on specific organisms or infection sites [9, 25, 30, 31, 35–37].
As part of the Chicago Antimicrobial Resistance Project (CARP), a 5-year demonstration program to evaluate interventions for control of antimicrobial drug—resistant infections [38–42], we compared the total cost of caring for patients with HAI with the cost of caring for those who did not develop HAI. Because CARP is a hospitalwide intervention, we used cost comparison methods that can estimate hospitalwide cost savings . Comparative, concurrent, and site- and organism-specific HAI methodologies exclude groups of hospital patients that might influence the estimate of excess cost and so were not suited to the needs of this project. We conducted a retrospective cohort study that used economic modeling to control for cost confounders and measured cost from the hospital's perspective.
Patients and Methods
Sample selection. The study population included patients admitted to Cook County Hospital, an urban public teaching hospital in Chicago, Illinois, from 1 January through 31 December 1998. Previous studies have shown that a patient's risk of HAI increased with that patient's number of comorbidities [8, 15, 43–45]. In addition, the administrative database used to generate our random sample showed that patients with fewer than 6 International Classification of Diseases, Ninth Revision (ICD-9), codes at discharge had an average length of stay that was often too short to permit the patient to manifest HAI. Therefore, a sample of 246 patients was randomly selected from among those discharged with ≥6 ICD-9 codes to ensure that the number of patients with HAI was adequate for the cost analysis. Patients aged <18 years or those hospitalized in surgical, obstetrical, or trauma wards were excluded from the study. This 2-step process was needed because it was not feasible to exclude these patients when we used the available administrative database containing 26,834 patients.
Measurements. We expected that the cost of HAI would be confounded by both severity of illness and by treatment in the ICU [2, 5, 6, 15, 18, 28, 46–55]. To control for the increased cost due to severity of illness, the APACHE III score was used because it can be measured in non-ICU patients, was useful for controlling for cost, was widely used in the United States, and was applicable across all diagnostic groups . We also tested the Simplified Acute Physiology Score II (SAPS-2) and Charlson scores for this purpose, but we found that SAPS-2 was not as useful for controlling for costs, and the Charlson score was originally developed to predict outcomes at 1 year [57, 58].
To classify patients as having an HAI or not, we used the Center for Disease Control and Prevention's National Nosocomial Infection Surveillance (NNIS) system definitions, with slight modification for use in a retrospective study [8, 9, 18, 59, 60]. The NNIS definitions were developed for prospective hospital surveillance and are designed to be quite specific. Because clinical decisions are often not made on the basis of surveillance definitions, we believed that some cases of clinically suspected infection would meet most but not all of the NNIS criteria and thus be classified as non-HAI, especially on a retrospective chart review [28, 49, 61–63]. Some of these cases could have relatively higher costs for diagnostic services and treatment. If these patients were classified as not having an HAI, we might underestimate the cost for HAI, but excluding them might artificially increase the cost for HAI as well as limit the applicability of the cost results. Therefore, we classified patients as follows: patients who were not infected, those with suspected HAI, and those with confirmed HAI. In general, patients with suspected HAI included those who were provided antimicrobial therapy for a condition that appeared >48 h after hospital admission and who met all but one clinical criteria for a confirmed infection. Definitions for confirmed HAI were the same as those used by the NNIS, except that receipt of appropriate antimicrobial therapy was excluded as a criterion for a confirmed infection. These criteria were all finalized before chart data abstraction began.
The economic perspective used for measuring cost was that of the hospital, because the hospital administration is currently the decision maker for instituting and financing infection-control programs .
Data collection. All data were abstracted from patient medical records by trained abstractors. Interrater reliability was not measured, because each abstractor focused on recording a single element of data for each patient, similar to an assembly line. All were directly supervised by one author (R.R.R.). Only data collected during the first 24 h of hospitalization were used for the physiologic portion of the APACHE III. A single study physician, who was not involved in the cost or HAI data collection or analysis, assigned all APACHE III physiologic and chronic condition scores on the basis of information abstracted from the medical record. Patients with suspected or confirmed HAI were identified on the basis of their vital signs, laboratory and microbiology data, and clinical findings documented in physician progress and consultation notes.
The first step in measuring the cost of patient care was abstracting patient resource use from medical records. This included length of stay in all locations of care (medical wards, intermediate-care wards, and ICUs) and the numbers and types of laboratory and radiographic tests, procedures, consultations, and medications administered.
We did not use charges or cost-to-charge ratios. Instead, unit costs were calculated using Cook County Hospital's 1998 annual expenditure report, which included all costs related to building, utility, equipment, and labor, as well as variable costs for medication, food, test reagents, and supplies. Departmental and finance directors were interviewed to obtain work outputs. The total cost for operating each hospital department that provided patient service was calculated by using the multiple distribution method to allocate building overhead and support costs to the patient service departments [64, 65]. These allocations were based on the number of full-time equivalent employees and square footage of space occupied. The total service department costs were divided by the total work outputs (e.g., total prescriptions filled or radiographs performed) to calculate the unit costs. The total cost for each patient was calculated by multiplying the quantity of each resource used by the unit cost of that resource, then summing all resource costs. This method has been described elsewhere [66–68].
Analysis. Ordinary least-squares (OLS) regression was used to test for linear relationships between per-patient hospital costs (the dependent variable) and the hypothesized predictors of hospital resource use (APACHE III; HAI, both suspected and confirmed; and ICU care as the independent variables). The APACHE III score is a continuous variable. Suspected HAI, confirmed HAI, and admission to ICU were coded as dummy variables, with the values of 1 assigned for patients with the attribute and 0 for those without it. When present, these dichotomous variables act as intercept shifters but do not change the slope of the estimated regression line.
Given that these regression lines represent total per-patient hospital costs, the estimated constant terms can be interpreted as measures of per-patient fixed costs. However, caution must be exercised when making inferences based on constant terms . If independent variables with a small statistical impact on per-patient cost are omitted, their marginal effects are attributed to the constant term in OLS regression, leading to a biased estimate of the constant term. For this reason, the raw data are presented as a scattergram to show the relationships among APACHE III, HAI, and cost.
Three economic models were tested sequentially to illustrate cost relationships. Model 1 tested the relationship between per-patient costs and APACHE III scores. Models 2 and 3 incorporated additional cost predictors sequentially: suspected and confirmed HAI and ICU treatment. Care in the ICU was introduced last to differentiate more clearly its effect on cost from that of developing HAI. Student's t test was used to determine the statistical significance of each independent variable's relationship to cost. The statistical significance for each model was determined using the F statistic. The a priori α for statistical significance for the individual variables and the overall models was ≤.05. The predictive power of each model was evaluated by comparing the overall fit (the amount of statistical variation in cost explained by the independent variables) using the adjusted coefficient of determination (R2). The null hypothesis tested was that the parameter in question has no linear relationship to cost. Therefore, the P values for the model variables quantify the chance that the parameter tested had no relationship to cost. After estimating the economic models, we plotted the residuals to confirm that our OLS assumption of linearity was appropriate. Comparisons included all patients, even outliers, and were generated using SAS statistical software (SAS Institute) . Data are presented as mean ± SE, unless otherwise indicated.
A total of 26,834 patients were admitted to Cook County Hospital in 1998. Of these, 13,233 (49%) had ≥6 ICD-9 codes and were eligible for the study. The randomly selected sample included 246 (2%) of the eligible group. Of those, 82 were excluded because they were <18 years of age or were hospitalized in surgical, obstetrical, or trauma services. HAI developed in 25 (15.2%) of 164 total study patients; 8 (32%) of 25 met our definition for suspected HAI, and 17 (68%) met our definition for confirmed HAI (table 1). The most common site of infection was pulmonary (11 [44%] of 25 patients), followed by bloodstream (8 [32%] of 25). Urinary tract and skin infection each accounted for 2 (8%) of 25 infections. Gastrointestinal and CNS each accounted for 1 (4%) of 25 infections. One of the patients categorized as having bloodstream infection had concurrent urine and gastrointestinal infections. For the measured comorbidities, patients with cancer, HIV infection, and renal disease more commonly developed HAI (table 1). High APACHE III scores (table 2) and ICU care (table 3) were also significantly associated with increased rates of HAI and thus were potential cost confounders.
The mean total cost of care for patients without HAI ($7338 per patient) was significantly lower than that for patients with HAI ($25,638; P < .001; table 1). However, both high APACHE III scores and ICU treatment were associated with increased cost and increased HAI rates (table 2 and table 3), thus confirming our expectation of confounding. The data shown in figure 1 demonstrate the relationship between APACHE III scores and cost of care for those with and without HAI. The sequential economic models shown in table 4 control for this confounding.
In model 1 (only APACHE III scores included), the base cost of caring for a patient was $3341. This cost increased by $195 ± $41 for each APACHE III score point. Although APACHE III scores were significantly associated with cost of care (P < .001), this model explained only 12% of the total variability in cost. In model 2 (APACHE III and both HAI variables included), the explained variability in total cost was 40%. The base cost decreased slightly, and controlling for the effect of severity resulted in an excess cost of $18,223 ± $2225 for confirmed HAI (P < .0001) and $13,236 ± $3145 for suspected HAI (P < .0001). Finally, by adding the ICU variable, model 3 was able to explain 56% of the total per-patient cost variability. Care in the ICU contributed $14,075 ± $1860 to the total base cost (P < .0001). The excess cost for HAI remained significant, but it decreased to $15,275 ± $1953 for confirmed HAI (P < .0001) and $6767 ± $2837 for suspected HAI (P < .0183). The relative cost contribution for each point in the APACHE III score decreased in each regression model as the site of care and presence of HAI explained more of the total cost.
The results of this study indicate that HAI is associated with significantly increased cost from the hospital perspective, even after controlling for the confounding effects of initial severity of illness and ICU care on cost. The estimated cost of caring for patients at our hospital increased by $6767 for patients with suspected HAI and $15,275 for those with confirmed HAI. This is similar to estimates in earlier reports [2, 25, 27, 29–31, 33, 34, 36, 37]. Our findings also confirm the hypothesis that severity of illness and ICU care independently increase cost and thus need to be identified as confounders in future studies. It is of note that more than one-half of the patients in our cohort who developed HAI did so outside of an ICU. Confining studies to only those patients treated in the ICU may underestimate the prevalence of HAI and result in cost estimates that are not applicable to the entire facility. Additional studies to evaluate the cost-effectiveness of more rigorous hospitalwide infection-control plans may be justified .
Our methods introduce unconventional uses of existing measurement tools. Like most severity of illness indices, APACHE III was designed for use in ICUs to predict the risk of death [18, 56, 71–75]. In this project, the APACHE III scores were used to predict cost and to control for potential confounding due to severity of illness [28, 76]. A major benefit of APACHE III is that scores can be calculated even when particular laboratory values are not obtained during the usual course of care . This study also demonstrates the usefulness of APACHE III scores as indicators of severity of illness in patients outside of the ICU. In addition, the physiologic measures for APACHE III are usually measured, as in our study, in the first 24 h after admission to the ICU . A future refinement in our methods would be to evaluate severity trends over time or to measure the APACHE III score in the 24 h just before the first evidence of HAI was noted, which is similar to methods used by others [30, 32, 48, 50, 77].
We also modified the NNIS definitions to allow classification of patients as having either suspected or confirmed HAI. For surveillance, the specificity of HAI definitions must necessarily be high for valid comparisons between hospitals and over time [28, 49, 62, 78]. However, because surveillance definitions do not always reflect clinical decision making, this high specificity could have resulted in underestimated HAI-associated cost if patients who were evaluated and treated for HAI were classified as not having HAI. The regression model results confirmed our hypothesis that patients who met most but not all of the NNIS criteria for infection would be intermediate in cost between those with confirmed HAI and those without evidence of HAI. This supports the argument that the use of dichotomous outcome variables for conditions that are difficult to diagnose may not fully describe cost relationships .
Another challenge was to increase feasibility by selecting a patient sample with a high HAI rate. Others have found that a greater number of comorbidities correlates with higher risk of HAI [8, 44, 45]. Our surrogate for numbers of comorbidities was the hospital discharge ICD-9 rate. This approach met our need to increase the number of HAIs in our study sample and supports previous reports [8, 43–45]. A potential flaw with this method is that the development of HAI may have contributed ≥1 additional ICD-9 code. However, if true, it would favor the null hypothesis—that is, patients without HAI might have more preexisting comorbidities than those in the HAI group, and their treatment would have been more costly. Selection of the sample from the more severely ill subset of hospitalized patients could overstate both baseline and excess HAI-related costs, but this subset represented one-half of the hospitalized population.
Our results indicate that economic modeling techniques can generate estimates of the contribution of HAI to the cost of providing care that are consistent with those of other investigators [17, 19, 25–27, 29, 30, 33, 34, 36, 37]. The total variability in cost explained by our third economic model was 56%. Compared with previous studies, this is a high overall fit for a random sample of medical patients with many comorbidities and in different diagnosis-related groups (DRGs) [24, 79–84]. Seldom have others shown this degree of correlation between cost and any single predictive system, even within isolated diagnostic groups. There are several possible explanations for the ability of our model to reliably predict cost. This study was performed at a single hospital in 1 year. Rather than using an average daily cost or total charge, we laboriously measured the costs for resources used by each patient. Sampling patients with the most ICD-9 codes and confining eligibility to medical service patients also contributed to the strength of our model. However, we postulate that the major reason our third regression model had good predictive power was due to the inclusion of HAI and its separation into confirmed and suspected groups, because HAI is both expensive and an uncommon event.
This study has several limitations. Our methods did not control for potential confounding by length of hospital stay [2, 7, 28, 62]. The length of hospital stay before manifestation of HAI was measured, but there was no corresponding “preinfection” length of stay for those without infection. In those with HAI, the pre-HAI length of stay was highly variable, and our sample had an insufficient number of subjects to generate any predictive power when this variable was used. Our ability to generate data for specific infection sites was also limited by the small number of patients. Larger studies addressing these limitations are clearly warranted.
Any study that measures cost must choose one of a variety of economic perspectives [28, 67, 85]. Some cost analyses use charges or reimbursements to hospitals by third-party payors. From the hospital perspective, costs are for labor, building space, utilities, equipment, and supplies that the hospital purchases to operate the health care facility . For this study, we reported cost from the hospital perspective for several reasons. Haley and others have reported that the occurrence of HAI does not greatly increase DRG-based reimbursement for hospitals by third-party payors [15, 86]. There are also no specific DRG codes for HAI . This suggests that much of the excess cost for HAI is borne by the health care facility rather than third-party payors [15, 28, 86]. In addition, it is unlikely that hospitals incorporating HAI-prevention programs will be able to directly bill for them [88, 89]. To encourage patient safety programs that prevent HAI, payments similar to those used by Medicare in reimbursing hospitals for graduate medical education are worth considering, but would require further data and discussion before national policy implementation. Until hospitals can negotiate higher reimbursement rates for maintenance of effective infection-control interventions and low HAI rates, the decision on how much to spend for infection control is still a hospital decision [13, 86].
The most important elements in deciding whether to incorporate new infection-control interventions is how feasible and effective they are, how much they cost, and, finally, how much they might save in prevented nonreimbursed HAI costs [15, 23, 28, 89]. Our results begin the process of providing costs from the hospital perspective to help make these important decisions.