Discussion
The most important finding in this work is that adherence to the IDSA/ATS guidelines regarding empiric antibiotic therapy for elderly patients hospitalized with CAP was cost effective compared to non-adherence in non-ICU patients, but was not the most cost effective strategy in ICU patients. Elderly patients admitted to the ICU with CAP who were treated with an antibiotic regimen in excess of the IDSA/ATS guidelines had lower costs and increased quality of life compared to those treated with an adherent regimen.
The cost effectiveness implications of adherence to treatment guidelines has been evaluated in a limited number of settings, finding that in general, adherence to treatment guidelines is cost effective. Adherence with guidelines for sarcoma treatment was cost effective in a recent study from two European regions [
12]. Adherence to guidelines has also been shown to be cost-effective in such diverse fields as intrauterine insemination and hepatitis B therapy [
10,
11]. A study from Japan has shown the adherence to guidelines for gastric ulcer therapy is cost effective in that country [
9]. Consensus treatment and practice guidelines have the potential to not only standardize care and disseminate best practice measures throughout complex healthcare systems, but also may offer cost effective strategies that are of particular importance in the current era of health care utilization reform.
Adherence to guidelines for the treatment of CAP has been associated with improved outcomes. An analysis of the same CAPO data used in this study found adherence to guidelines improved mortality, LOS, and time to clinical stability in elderly patients hospitalized with CAP [
8]. However, the previous study did not separate patients on the basis of admission to the ICU or ward. Nonadherence to guidelines was associated with in-hospital mortality in studies of both community hospitals and tertiary teaching hospitals [
25,
26]. These findings of improved outcomes in both hospitalized and outpatient settings with adherence to antibiotic guidelines have been replicated [
27‐
29]. Nonadherence rates were higher in patients treated by non-pulmonary specialists in one study [
30]. The cost implications of adherence to guidelines in hospitalized patients with CAP have not been studied as extensively. One study from Europe evaluated the cost-effectiveness of adherence to Spanish guidelines for the treatment of hospitalized CAP in non-ICU patients of any age [
13]. This study found that adherence to the guidelines was the dominant cost-effective strategy compared to non-adherence. To our knowledge, this study is the first to evaluate the cost-effectiveness of adherence to guidelines in an exclusively elderly population with CAP in a multi-institutional setting, stratified by admission to the ICU or ward. Additionally, this study accounts for the stochastic nature of a hospital stay by estimating conditional transition probabilities directly from the data using the Aalen-Johansen estimator. This is an improved method by which to estimate cost and utility changes during a hospitalization course, as one is able to estimate daily, conditional changes in the probability of transitioning between states rather than averaging the transition probabilities over the course of an entire hospitalization, thus providing more precise estimates of cost and utility changes.
While the finding that adherence to IDSA/ATS guidelines was the dominant strategy in ward (non-ICU) patients was expected and is consistent with a previous report, it was surprising to find that over-treatment was the dominant strategy in ICU patients [
13]. A close analysis of the state occupation probabilities used to build the model suggests the likely explanation is the accelerated time to clinical stability in this cohort of patients, which under our assumptions led to a reduction in ICU LOS for these patients. Daily ICU costs were the dominant driver of total costs in this model. Drug costs, which are a direct consequence of empiric antibiotic therapy choice, are miniscule compared to the daily cost of a stay in the ICU, which is measured in thousands of dollars per day. Thus, any large cost benefit from an antibiotic choice in relationship with the IDSA/ATS guidelines would likely be driven by a reduction in ICU and overall LOS rather than a reduction in daily drug costs. In the CAPO cohort, ICU LOS was shorter in patients who were treated with an over-treated antibiotic regimen compared to an adherent strategy. By hospital day 5, only 43 % of over-treated patients initially admitted to the ICU remained there, while the percentage of patients treated with an adherent regimen remaining in the ICU was nearly twice that (71 %) (Fig.
2). This led to dramatic reductions in daily and cumulative costs, and subsequent improvements in utility for the over-treated group. The marginal (unadjusted) 14-day mortality rate was higher in the over-treated group (18 %) compared to the adherent group (14 %) (Table
1), though differences in adjusted mortality rates were smaller (4.1 % vs. 2.4 %, respectively, Fig.
2). Overall, the lower costs associated with quicker transfer from the ICU to the ward made the over-adherent treatment strategy the most cost effective.
From this analysis, one cannot draw a conclusion of direct causation regarding the implementation of an overly-broad antibiotic regimen and the reduction in ICU LOS, although there certainly seems to be a strong association between an over-treatment strategy and a decreased ICU LOS that leads to reduced hospital costs. One potential explanation is the possibility that the patients in the over-treatment cohort were over-triaged to the ICU compared to adherent and under-treated patients and thus were more likely to leave the ICU earlier, independent of initial empiric antibiotic coverage. However, the initial demographics of our ICU patients suggest that the three antibiotic groups were relatively similar (Table
1). Pulmonary severity indices were not significantly different across the antibiotic groups, nor were the differences in comorbid conditions such as cancer, congestive heart failure, COPD, renal, or liver disease. Though the sample size for the ICU was small, over-treated patients did have a significantly higher probability to reach clinical stability by day 7 (Table
1), and in particular patients that received both a macrolide and a quinolone in addition to a β-lactam. Interestingly, another recent study of hospitalized CAP patients from three world regions found an elevated risk of mortality associated with ICU patients prescribed flouroquinolones relative to macrolides [
31]. However, this study did not investigate whether the combination of a flouroquinolone with a macrolide provided any additional benefit, nor did it investigate time to clinical stability or length of hospital stay. These findings suggest that strategies to reduce ICU LOS may include aggressive empiric antibiotic regimens that can be later tailored to more targeted therapy guided by clinical response and microbial laboratory findings.
The findings in this study were robust over a wide range of sensitivity analyses performed. Sensitivity analyses of model parameters in multi-state Markov models are a critical component of the evaluation of model findings, as the model predictions are only as good as the parameters used to build the model. Antibiotic costs were estimated from a single local institution (University of Louisville Hospital), thus a valid criticism would be the external validity of these cost estimates to other hospitals. Sensitivity analyses found that the dominant strategies were robust over a wide range of daily antibiotic costs, thus suggesting that these findings have external validity to other institutions. Utility estimates are also quite subjective. We attempted to control for this using an expert panel of CAPO investigators who are familiar with the care of elderly patients with CAP. Modeling the utility estimates as a probability helps to account for this variability in the probabilistic sensitivity analysis. State occupation probabilities were available for the entire 30 day hospital course for our study cohort. We limited our model to an analysis of only the first 14 days, reasoning that hospitalizations beyond 14 days for CAP in the elderly likely were the result of complex, confounding factors and associated comorbidities that were unaffected by the initial empiric antibiotic coverage. Sensitivity analysis, however, did confirm the cost dominant strategies remained the same for models analyzed > 14 days.
The results reported here are based on the adjusted transition probability estimates based on the parametric models. However, conclusions based on the non-parametric estimates were similar, with adherence being the dominant strategy for the ward and over-treated the dominant strategy for the ICU. If the setting of the Markov model was extended to include post-discharge utility, then the difference in unadjusted mortality between the adherent and other antiobiotic groups might have a greater impact on the cost-effectiveness analysis. The most salient difference between the adjusted and unadjusted models was the reduced probability of mortality and increased probability of discharge and reaching clinical stability in the ICU for the adjusted model relative to the unadjusted estimates (Additional file
5).
The findings in this study must be considered with consideration of the shortcomings. Decision analysis models are only as strong as the parameters used to build them. While every effort was made to develop the model with realistic parameters, limitations arise in estimating transition probabilities, cost and utility estimates, and perspective of the model. Baseline differences in the clinical risk factors among the three groups, including PSI score and age, clearly have an impact on estimates of in-hospital mortality and discharge rates. While attempts were made to limit their effect on the model using adjusted transition probability estimates, differences in clinical pathways, practice, and resource utilization among the participating institutions and their effects on the clinical outcomes measured in this study cannot be discerned. An additional limitation is the assumption that patients transitioned out of the ICU once they reached clinical stability. This may not necessarily be the case, as patients may require longer stays due to deterioration of other comorbidities or other reasons. However, on average patients who reach clinical stability sooner should also leave the ICU earlier, so the observed differences between antibiotic treatment classes should be relatively robust. Finally, the estimated transition probabilities for over-treated patients in the ICU were based on only 28 patients, and require independent validation of the finding that this treatment regime was most cost-effective for these patients in the hospital setting.
A second limitation of the model is how patient utility estimates were obtained. The sample size of expert opinions (six) was limited and only from US institutions. While the experts were independently polled and we have no reason to believe there responses were subject to bias, an alternative method for obtaining a consensus (e.g., the Delphi method) might be preferable. Co-morbidities were not factored into the utility estimates both to simplify the data collection and because we felt this was not directly related to CAP/CAP therapy. Finally, using expert opinions to model patient-centric utilities is less desirable than patient-based estimates. While expert judgments concerning utilities should be used sparingly, it became pragmatically difficult to elicit utilities from patients and a data source does not exist in which the specific utilities we needed were present. A superior approach would be to estimate utilities based on appropriate patient questionnaires, perhaps as part of a prospective study to include post-discharge outcomes. While we did not have the resources to obtain such data as part of this study, future research will definitely take this into consideration.
Third, the CAPO data used in this study does not contain cost data, and external references and data were used to model both hospitalization and antibiotic regimen costs. While it may prima facie seem inconsistent to model costs from two different sources, we elected to use data from Dasta et al. [
22] in addition to Kaplan et al. [
23] for two reasons. First, the data from Dasta et al. suggests that ICU costs are not constant, rather they change over time. Hence these data are uniquely appropriate for use in a Markov model, in which one can capture these differences and make more accurate inferences about differences in costs based on how many days are spent in the ICU. Second, the data from Kaplan et al. only provides estimates of the total hospitalization costs for patients initially treated for CAP in the ICU. Thus these estimates capture both costs in the ICU and in the ward, and an average cost per day calculation from the Kaplan data would be “contaminated” with time spent on the ward. Hence it is a less fit estimate for daily ICU costs compared to the Dasta et al. data. Lastly, costs were estimated based on US institutional rates but applied to international data. However, since the results were consistent across sensitivity analyses for costs and willingness to pay thresholds, we feel the conclusions are relatively robust to any regional variation in costs.
Finally, the conclusions here are based on the short-term utility and cost-effectiveness for CAP patients during their hospital stay. Since we did not have direct data concerning post-discharge mortality and quality of life for this cohort of patients, we did not attempt to model long-term cost-effectiveness. While we have no evidence to suggest that post-hospitalization outcomes are different among the treatment cohorts in this study, we acknowledge that it is a limitation of the model that it incorporates only a fraction of the life-expectancy for these patients. However, our multi-state modeling approach is easily extendable to prospectively collected data on patients discharged after hospitalization due to CAP [
32‐
35]. For example, one could model the post-discharge outcomes as additional states within the Markov model, tracking transitions between occurrence of various co-morbidities, re-hospitalizations, and mortality. The multi-state model can also be coupled with causal inference approaches to assess the effect of various interventions/treatment regimens on outcomes of interest. For example, such models have been fit in other disciplines to evaluate whether partial versus full time sick leave and having a cooperation agreement is helpful in reducing sickness absence from work [
36]. In a similar fashion, a multi-state model coupled with causal inference approaches may form the basis of active, prospective evaluation of antibiotic choices and the cost-effectiveness of adherence to recommended guidelines. Patient quality of life would then be ideally collected using appropriate designed patient questionnaires. As a final thought, such studies would also need to take into consideration the effect of antibiotic regimen on antimicrobial resistance. The societal costs of antibiotic over-treatment pressures on the development of antimicrobial resistance are hidden costs to the over-treatment strategy that must be considered.
Competing interests
The authors declare that they have no competing interests
Authors’ contributions
JAM and GNB conceived of the project and supervised the research. MEE, GNB, and JAM conducted the statistical and cost-effectiveness analysis. FWA and LAP compiled the antibiotic usage and daily cost data. JAR helped supervise the research and contributed to the clinical interpretation of the results. MEE drafted the manuscript, and all authors read and approved the final version of the manuscript.