ABSTRACT
We sought to define trends in and predictors of carbapenem consumption across community, teaching, and university-affiliated hospitals in the United States and Canada. We conducted a retrospective multicenter survey of carbapenem and broad-spectrum noncarbapenem beta-lactam consumption between January 2011 and December 2013. Consumption was tabulated as defined daily doses (DDD) or as days of therapy (DOT) per 1,000 patient days (PD). Multivariate mixed-effects models were explored, and final model goodness of fit was assessed by regressions of observed versus predicted values and residual distributions. A total of 20 acute-care hospitals responded. The centers treated adult patients (n = 19/20) and pediatric/neonatal patients (n = 17/20). The majority of the centers were nonprofit (n = 17/20) and not affiliated with medical/teaching institutions (n = 11/20). The median (interquartile range [IQR]) carbapenem consumption rates were 38.8 (17.4 to 95.7) DDD/1,000 PD and 29.7 (19.2 to 40.1) DOT/1,000 PD overall. Carbapenem consumption was well described by a multivariate linear mixed-effects model (fixed effects, R2 = 0.792; fixed plus random effects, R2 = 0.974). Carbapenem consumption increased by 1.91-fold/quarter from 48.6 DDD/1,000 PD (P = 0.004) and by 0.056-fold/quarter from 45.7 DOT/1,000 PD (P = 0.93) over the study period. Noncarbapenem consumption was independently related to increasing carbapenem consumption (beta = 0.31 for increasing noncarbapenem beta-lactam consumption; P < 0.001). Regular antibiogram publication and promotion of conversion from intravenous (i.v.) to oral (p.o.) administration independently affected carbapenem consumption rates. In the final model, 58.5% of the observed variance in consumption was attributable to between-hospital differences. Rates of carbapenem consumption across 20 North American hospitals differed greatly, and the observed differences were correlated with hospital-specific demographics. Additional studies focusing on the drivers of hospital-specific carbapenem consumption are needed to determine whether these rates are justifiable.
INTRODUCTION
Resistance to last-line antimicrobials, including carbapenems (C), among Gram-negative bacteria has been labeled as an urgent public health threat (1). Economic modeling suggests that the current U.S. rates of infections caused by carbapenem-resistant Enterobacteriaceae produce excess hospital and third-party payer costs of over $422 million combined, in addition to costing society over $553 million as a whole (2). Decreases in carbapenem effectiveness have been driven by increasing resistance and higher MICs (3–5). Carbapenem resistance is associated with higher morbidity, reductions in initial active antibiotic treatment, and worse patient outcomes (6–10). Recent studies found that Gram-negative bloodstream infections caused by carbapenem-resistant bacteria are associated with crude mortality rates of between 20% and 50% (11–15). As a result, to address the current crisis of carbapenem resistance, antibiotic stewardship programs have focused on limiting carbapenem use to cases in which more-narrow therapy is not feasible.
Antibiotic stewardship programs are ideally positioned to address carbapenem resistance by developing and implementing interventions to improve carbapenem use and by monitoring carbapenem consumption. Previous investigations have demonstrated that higher rates of carbapenem consumption correlate with higher rates of carbapenem resistance (16–18). Limiting unnecessary carbapenem use has been shown to produce favorable effects on the resistance rates of problematic nosocomial pathogens such as Pseudomonas aeruginosa (19). Importantly, antibiotic consumption is one of the few drivers of antimicrobial resistance that antibiotic stewards can influence. Therefore, measuring consumption is a critical component of successful antibiotic stewardship efforts.
Determination of how the rate of carbapenem consumption in a given hospital correlates to rates in other hospitals is a necessary first step in benchmarking use. Improved understanding of the drivers of carbapenem use is important because carbapenems represent the fourth most commonly used antibiotic class in within hospitals (20) and because unnecessary carbapenem use increases the risk of acquiring carbapenem-resistant pathogens (21). Therefore, we conducted a cross-sectional survey to understand potential drivers and contemporaneous rates of carbapenem consumption across small and large hospitals from the Making a Difference in Infectious Diseases (MAD-ID) network. The ability to compare the carbapenem consumption rate of a given hospital to that of similar institutions is an essential element in determining whether local consumption is above or below the level in benchmark hospitals; however, this task complicated by between-institution differences (e.g., hospital size, number of intensive care unit [ICU] beds, etc.) and the choice of consumption metrics (22, 23). Here, we sought to evaluate trends in and predictors of carbapenem consumption across 20 demographically and geographically diverse North American hospitals.
(These findings were presented, in part, as a poster at the American Society of Health-System Pharmacists Midyear Clinical Meeting in December 2014, at the Making a Difference in Infectious Diseases [MAD-ID] Annual Meeting in May 2015, and as a platform presentation at the Interscience Conference of Antimicrobial Agents and Chemotherapy [ICAAC/ICC] Meeting in September 2015.)
RESULTS
Demographics of participating centers.Among the 181 network sites, 20 participating centers contributed antimicrobial consumption and demographic data. A total of 9 centers reported consumption in defined daily doses (DDDs), and 11 reported consumption in days of therapy (DOTs). Over 12 quarters spanning January 2011 to December 2013, participating centers provided a total of 228 consumption observations from 240 possible observations. Consumption data were not reported from quarters 1 and 2 by 3 centers, quarters 3 and 4 by 2 centers, or quarters 5 and 6 by 1 center. Data corresponding to administrations were available from 6 centers, while the remaining 14 centers relied on dispensing records. The demographics of the participating centers are shown in Table 1 and stratified according to DDD or DOT reporting status. The majority of participating centers (n = 17/20) were not-for-profit institutions. Among the not-for-profit institutions, 59% reported DOTs (n = 10/17) and 41% reported DDDs. Most centers served both adults (95%) and pediatric (70%) populations. A minority of centers served neonatal patients (15%) or a long-term-care population (10%). Numbers of beds ranged from between 100 and 249 to ≥750 among the participating centers. Numbers of beds were fairly evenly distributed across DDD-reporting hospitals (Table 1). DOT-reporting hospitals commonly had between 250 and 499 beds (45.5%, n = 5/11), which was also the most common number of beds overall (35%, n = 7/20). DOT-reporting hospitals were numerically more likely to have between 30 and 59 ICU beds compared to DDD-reporting hospitals (36.4 versus 0%; P = 0.09). Overall, facilities within the response sample represented a collection of demographically diverse hospitals from our network.
Institution demographics according to DDD or DOT reporting capabilitiesa
Antimicrobial stewardship characteristics.Participating centers employed a variety of self-reported antimicrobial stewardship strategies to control carbapenem use at each site, as shown in Table 1. The most common self-described antimicrobial stewardship strategy employed was annual antibiogram dissemination (95%, n = 19/20). Most participating centers utilized combination approaches that included the following: antimicrobial prescribing guidelines (85%), prospective audit and feedback for prescribers (80%), active policies promoting conversion from intravenous (i.v.) to oral (p.o.) administration (75%), and antimicrobial formulary restrictions (70%). Only a minority of participating centers reported using rapid diagnostic testing for early identification of pathogens at the time of the survey (20%). Likewise, only 40% of centers reported using specialized medication order forms to control carbapenem use. There were no significant differences in the types of self-reported antimicrobial stewardship strategies used to control carbapenem use between DDD-reporting and DOT-reporting centers (Table 1). However, numerically more DDD-reporting centers than DOT-reporting centers utilized carbapenem order forms (66.7% [n = 6/9] versus 18.2% [n = 2/11], respectively; P = 0.07).
Variability in antimicrobial consumption across demographics.The rates of consumption of carbapenems (i.e., doripenem, ertapenem, imipenem-cilastatin, and meropenem) and noncarbapenem (NC) beta-lactam (i.e., cefepime and piperacillin-tazobactam) in DDD/1,000 patient days (PD) are shown in Table 2. Meropenem was the most commonly utilized carbapenem (44% of carbapenem DDDs) followed by ertapenem (40.3% of DDDs), while imipenem-cilastatin and doripenem were infrequently utilized (8.9 and 6.8% of DDDs, respectively). Carbapenem DDDs were highly variable. Among the centers reporting DDDs, the overall median (interquartile range [IQR]) consumption rate of carbapenems across the entire 3-year study period was 38.8 (17.4 to 95.7) DDD/1,000 PD. The corresponding 3-year noncarbapenem beta-lactam consumption rate was 77.7 (42.2–210.1) DDD/1,000 PD. The percent coefficients of variation (i.e., standard deviation [SD]/mean × 100) for carbapenem and noncarbapenem consumption were 94.3 and 100.3%, respectively. Across all hospital demographics, the highest levels of carbapenem consumption were noted at smaller facilities (median [IQR], 104 [60.2 to 119]; n = 2) and facilities with 9 or fewer ICU beds (median [IQR], 118.7 [106.2 to 125]; n = 1), while the lowest levels of consumption were observed at governmental facilities and facilities that provide long-term care (median [IQR], 11.3 [10.4 to 15.7]; n = 1).
Carbapenem versus noncarbapenem DDDs/1,000 PD according to hospital demographicsa
Rates of consumption of carbapenem and noncarbapenem beta-lactams are also compared according to DDD reporting status in Table 2. Noncarbapenem consumption was generally 1.1-fold to 5-fold greater than carbapenem consumption in DDD/1,000 PD across demographic categories. Exceptions to this trend were observed at facilities that provide long-term (noncarbapenem-to-carbapenem median consumption [NC/C] ratio, 0.1; n = 1), at governmental facilities (NC/C ratio, 0.1; n = 1), at medium-sized (i.e., 250 to 499 licensed beds) facilities (NC/C ratio, 0.2; n = 2), at centers with 29 or fewer ICU beds (NC/C ratio, 0.3 to 0.4; n = 5), and at facilities that utilized rapid diagnostics in their antimicrobial stewardship efforts (NC/C ratio, 0.7; n = 3). Notably, smaller facilities (i.e., 100 to 249 licensed beds; n = 2) and centers that utilized order forms to limit carbapenem use (n = 2) exhibited carbapenem consumption rates in DDDs that were similar to the noncarbapenem consumption rates (NC/C ratio, 1.1; n = 2).
The rates of carbapenem and noncarbapenem beta-lactam consumption in DOT/1,000 PD are shown in Table 3. Meropenem was the most commonly utilized carbapenem (49% of carbapenem DOTs), while imipenem-cilastatin, doripenem, and ertapenem were all utilized less frequently (20.1, 15.6, and 15.3% of DOTs, respectively). Carbapenem consumption rates in DOT/1,000 PD were also highly variable, similarly to the levels of variability observed in the centers reporting consumption in DDD/1,000 PD. Among the centers reporting DOTs, the overall median (IQR) consumption rate of carbapenems during the 3-year study period was 29.7 (19.2 to 40.1) DOT/1,000 PD. The corresponding 3-year median (IQR) consumption rate of noncarbapenem beta-lactams was 102 (69.4 to 131) DOT/1,000 PD. The percent coefficients of variation for carbapenem and noncarbapenem beta-lactam consumption were 90.1 and 123%, respectively. Across all hospital demographics, the highest levels of carbapenem consumption in DOTs were noted within the centers that used order forms to restrict carbapenem use (median [IQR], 180 [12.4 to 216]; n = 2) as well as in the governmental facilities (median [IQR], 38.2 [29.1 to 48.3]; n = 2) and in the facilities that provide long-term care (median [IQR], 35.5 [34.1 to 39.9]; n = 1), while the lowest levels of consumption were observed at the university teaching hospitals and the centers with 60 or more ICU beds (median [IQR], 21.8 [16.9 to 35.4]; n = 6).
Carbapenem versus noncarbapenem DOTs/1,000 PD according to hospital demographicsa
Rates of consumption of carbapenems and noncarbapenem beta-lactams are compared according to DOT reporting status in Table 3. In contrast to DDD-reporting centers, DOT-reporting centers experienced noncarbapenem consumption rates that were consistently higher than the rates of carbapenem consumption. Consumption of noncarbapenem beta-lactams in DOT/1,000 PD was 1.7-fold to 6.1-fold greater than carbapenem consumption across all evaluated demographics (Table 3).
Variability in antimicrobial consumption across institutions.The relationship between carbapenem consumption and noncarbapenem beta-lactam consumption is displayed in Fig. 1. Carbapenem consumption appeared to be moderately to highly positively correlated with noncarbapenem beta-lactam consumption for both DDD-reporting and DOT-reporting centers in the univariate analysis (r = 0.65 and 0.94, respectively; P < 0.001 for each). However, the carbapenem consumption rates were not uniform across centers. The influence of high-consumption centers was assessed among DDD-reporting centers. Two DDD-reporting centers demonstrated carbapenem consumption rates that were, on average, 4.5-fold higher than those of all other DDD-reporting centers combined. The overall correlation between carbapenem and noncarbapenem beta-lactam consumption appeared to be influenced by the data from these two centers. When the data from these centers were removed, the correlation coefficient reflecting carbapenem and noncarbapenem consumption in DDD/1,000 PD fell from 0.65 to 0.43 (P = 0.001) with a corresponding decrease in the R2 value from 0.427 to 0.187 (Fig. 1A and C). Likewise, among the DOT-reporting centers, a single center demonstrated carbapenem consumption rates that were, on average, 7.2-fold higher than those of all the other centers combined. The overall correlation between carbapenem and noncarbapenem beta-lactam consumption appeared to be heavily influenced by the data from a single center. When the data from this center were removed, the correlation coefficient reflecting carbapenem and noncarbapenem beta-lactam consumption in DOT/1,000 PD fell from 0.94 to 0.11 (P = 0.25) with a corresponding decrease in the R2 value from 0.877 to 0.012 (Fig. 1B and D).
Correlations between carbapenem and noncarbapenem consumption across 20 North American hospitals. Data represent correlations between carbapenem and noncarbapenem consumption with and without outlier removal. (A) Correlation of carbapenem and noncarbapenem DDDs/1,000 PD across all centers and quarters. r = 0.65; R2 = 0.427. (B) Correlation between carbapenem and noncarbapenem DOTs/1,000 PD across all centers and quarters. r = 0.94; R2 = 0.877. (C) Reanalysis of data presented in panel A with 24 observations from two outlier institutions removed from analysis. r = 0.43; R2 = 0.187. (D) Reanalysis of data presented in panel B with 12 observations from a single outlier institution removed from analysis. r = 0.11; R2 = 0.012.
Univariate and multivariate models of antimicrobial consumption over time.After observing that antimicrobial consumption rates varied by DDD or DOT reporting status and by hospital demographic characteristics and at the individual hospital level, we next sought to identify whether any time trends existed with respect to antimicrobial consumption. Time-dependent linear trends in carbapenem consumption were observed at the DDD-reporting centers (see Fig. S1 in the supplemental material), whereas DOT-reporting centers demonstrated relatively time-invariant rates of consumption (Fig. S2). The univariate linear regression over time revealed that carbapenem consumption in DDD/1,000 PD increased by a factor of 2.2 each quarter over the 3-year study period from an average baseline of 46.1 DDD/1,000 PD (P < 0.001 for linear trend). In contrast, the univariate linear regression of carbapenem consumption in DOT/1,000 PD over time revealed that consumption decreased each quarter by a factor of −0.47 from an average baseline of 49.1 DOT/1,000 PD (P = 0.26 for linear trend). On the other hand, univariate linear regressions in DDD-reporting centers revealed that noncarbapenem beta-lactam consumption increased by a factor of 0.63 each quarter from an average baseline of 121 DDD/1,000 PD (P = 0.39). Likewise, univariate linear regressions in DOT-reporting centers revealed that noncarbapenem beta-lactam consumption increased by a factor of 0.38 each quarter from an average baseline of 133 DOT/1,000 PD (P = 0.22).
Because consumption rates were highly clustered within hospitals, mixed-effects models were used to identify population-level predictors of consumption. Parameter estimates and standard errors for the final multivariate linear mixed-effects model of carbapenem consumption are shown in Table 4. A within-center relationship was observed between the rate of change in carbapenem consumption over time and the average baseline consumption rate: thus, these variables appeared nonindependent. Carbapenem consumption in DDD or DOT/1,000 PD was best and most parsimoniously represented by a model consisting of 11 fixed-effects parameters. The goodness of fit of the final model, as judged by regressions of predictions from the fixed-effects model (i.e., population-level predictions) and the fixed-effects-plus-random-effects model (i.e., individual hospital-level predictions) for the observed carbapenem consumption rates (classified per 1 unit carbapenem consumed/1,000 PD in the combined model), is graphically displayed in Fig. 2. The final fixed-effects model was parameterized with 5 independent predictors (5 df), 1 categorical predictor of ICU bed number (the reference group corresponded to 0 to 9 beds) with three levels (3 df), and an interaction between consumption metric reported (i.e., DDD or DOT reporting status) and time (3 df). The final linear mixed-effects model was significantly more parsimonious and explanatory than the next most explanatory model with 10 parameters (Akaike’s information criterion [AIC] = 1,795 versus 1798.1; P = 0.02 by likelihood ratio testing). Overall, the fixed-effects predictions fit the observed consumption data reasonably well (Fig. 2A) (R2 = 0.792; bias = 0.03 units carbapenem consumed/1,000 PD; imprecision = 7 units carbapenem consumed2/1,000 PD2). The final random-effects model accounted for the dependency of the rate of change in carbapenem consumption (i.e., slope) on the average baseline consumption rate (i.e., intercept) at the individual hospital level. The covariance of within-hospital random effects was estimated using an unstructured variance-covariance matrix. The time-dependent slope and intercept were highly correlated for each hospital (working correlation, r = 1), indicating a positive association. In combination, the fixed-effects-plus-random-effects predictions fit the observed consumption data very well (Fig. 2B) (R2 = 0.974; bias = 0.0031 units carbapenem consumed/1,000 PD; imprecision = 0.873 units carbapenem consumed2/1,000 PD2). The final mixed-effects model accounted for between-hospital differences in consumption rates. The interclass correlation coefficient for the final model was 0.585. That is, holding the covariates constant, 58.5% of the observed variance in carbapenem consumption was due to hospital-level differences.
Final linear mixed-effects model parameter estimates of carbapenem consumptiona
Population-level and individual hospital-level observed and model-predicted carbapenem consumption rates. (A) Fixed-effects model, predicted versus observed carbapenem consumption. (B) Fixed- plus random-effects model, predicted versus observed carbapenem consumption.
Once the final model was selected, the impact of model covariates could be assessed. We observed a significant time-dependent trend in carbapenem consumption across the study period, holding constant all other covariates and within-hospital differences. On average, carbapenem consumption at the DDD-reporting centers increased by an adjusted factor of 1.91 each quarter from a mean baseline of 48.6 DDD/1,000 PD (P = 0.037). On the other hand, carbapenem consumption increased on average by a factor of 0.056 each quarter from a mean baseline of 45.7 DOT/1,000 PD (P = 0.93). Irrespective of DDD or DOT reporting status, carbapenem consumption increased, on average, by 0.31 for every 1-unit increase in noncarbapenem beta-lactam consumption. Other covariates that independently influenced carbapenem consumption rates included the following: medium-sized centers with 250 to 499 hospital beds (the carbapenem consumption rate increased 28-fold), centers with increasing numbers of ICU beds (the carbapenem consumption rate decreased 90.7-fold to 112-fold compared to centers with 0 to 9 ICU beds for each categorical increase in numbers of ICU beds), centers that self-identified an antimicrobial stewardship strategy of active i.v. to p.o. interventions (the carbapenem consumption rate decreased 28-fold), and centers that utilized antibiogram publication as a self-identified antimicrobial stewardship strategy (the carbapenem consumption rate increased 29-fold). The overall average model predictions did not substantially change in the sensitivity analyses, excluding ertapenem, potential outlier sites, and antibiogram publication as a self-identified antimicrobial stewardship (Fig. S3).
DISCUSSION
We assessed carbapenem and noncarbapenem beta-lactam consumption data from a group of 20 North American MAD-ID Research Network hospitals in a multicenter, retrospective study. Our analysis was unique in that we utilized mixed-effects models to identify predictors of consumption across demographically diverse centers. These models were able to accurately predict carbapenem consumption in a sample of hospitals that exhibited highly variable consumption. After adjusting for between-hospital demographic differences and consumption of noncarbapenem beta-lactams, we observed significant time trends in consumption. Carbapenem DDD/1,000 PD increased by a factor of 1.91 each quarter (P = 0.004) in DDD-reporting centers. On the other hand, carbapenem DOT/1,000 PD increased modestly, though not significantly (P = 0.93), by roughly 5.6% each quarter. Our findings suggest that centers that utilize DDDs or DOTs exhibited different consumption trajectories even after adjusting for facility demographics.
We identified several independent predictors of carbapenem use. Consumption of noncarbapenem beta-lactams was independently and positively correlated with carbapenem consumption across all participating centers. Carbapenem consumption increased on average by 0.31 for every 1-unit increase in noncarbapenem beta-lactam consumption in the adjusted analysis, irrespective of DDD or DOT reporting status. We also found an influence of i.v. to p.o. programs on reducing carbapenem consumption (by an average of −28-fold) and an influence of disseminating antibiogram data as a self-reported stewardship strategy on increasing carbapenem consumption (by an average of 29-fold). The latter was an unexpected finding and may have been related to a variety of factors, including unmeasured confounders, the method of presentation of the antibiogram, and other stewardship strategies. Additional studies are needed to clarify optimal stewardship strategies for controlling carbapenem use.
Previous investigations have evaluated differences in DDD and DOT consumption. Polk and colleagues found that DDD correlated poorly with DOT for an array of antibiotics across 130 U.S. hospitals (24). This finding was echoed by findings reported by Dalton and colleagues from a study performed in Canada (25). However, time trends were not explicitly evaluated. More recently, Baggs and colleagues evaluated time trends in DOT-based consumption across 383 U.S. hospitals as a function of facility demographics and populations served, using a methodology similar to ours. While they did not observe time trends in overall consumption, they noted mean increases in DOTs/1,000 PD for carbapenems, carbapenem alternatives (3rd and 4th generation cephalosporins and beta-lactam/beta-lactamase inhibitor combinations), and glycopeptides (26). In our study, we also observed higher rates of carbapenem consumption associated with facility characteristics. Facilities that were for-profit institutions (n = 1) or smaller institutions (i.e., 100 to 249 licensed beds) (n = 2) exhibited higher DDDs/1,000 PD. Likewise, we found higher rates of consumption in facilities that were government-associated institutions (n = 2), institutions that provided long-term care (n = 1), or medium-sized institutions (i.e., 250 to 499 licensed beds) (n = 5) among DOT-reporting centers. These findings underscore the need for population-adjusted consumption benchmarks, which may ultimately be addressed through nationwide benchmarking tools, such as the CDC National Healthcare Safety Network (NHSN) standardized antimicrobial administration ratio (SAAR) (27).
Our report has implications for benchmarking efforts. Notably, our methodology allowed us to account for within-hospital correlations between the rate of change in carbapenem consumption and the baseline level of carbapenem consumption: revealing that the institutions that consume carbapenems at a higher baseline rate also tended to consume them at an increasing rate over time. The implication here is that, in the absence of external interventions, hospitals are more likely to continue to regress toward their own mean consumption rates rather than toward a population mean rate. We chose not to arbitrarily convert DDDs to DOTs in order to allow the “real world” differences between these centers to become visible. Future studies are needed to explore this autoregressive tendency and to determine if hospital-specific and nationwide benchmarking can be leveraged to improve carbapenem use. In the interim, our approach can provide to institutions without access to robust benchmarking tools (such as the NHSN antimicrobial use module, which gives participating facilities access to SAAR data [28]) an initial framework to (i) make preliminary assessments regarding their own carbapenem use and (i) determine if interventions targeting carbapenems are warranted.
Limitations to our analysis should be considered. First, our study was retrospective and unmeasured confounders are possible. Second, DDD or DOT reporting status may have been directly linked to the observed trends in consumption rates. Centers that have the ability to quantitate DOTs may have more-advanced clinical decision support systems and stewardship programs in place. However, we adjusted for DDD or DOT reporting status along with several other facility-level demographics in our analysis. Third, DDDs and DOTs can exhibit imperfect correlations for a variety of antimicrobial agents (24); however, as measurements, DDDs and DOTs are highly related, and no measure is completely free from bias (22). We included an interaction term to estimate the impact of either DDD reporting status or DOT reporting status over time in our mixed-effects model. Thus, our findings can be readily translated to centers reporting DDDs or DOTs. Third, the stewardship interventions used to control carbapenem use were self-reported and may have been inconsistently applied over the study period. However, variability in the quality and consistency of stewardship interventions was likely a reality for many centers prior to publication of the Joint Commission’s Antimicrobial Stewardship Standard (29). Fourth, our study captured data from only 20 hospitals, a relatively small number considering the total number of hospitals in North America. However, our sample may be instructive to others seeking to evaluate their own consumption rates. Fifth, longitudinal bacterial susceptibility data were not evaluated as part of this study. Therefore, we are unable to discern the degree to which any of the model covariates could be surrogates for resistance. Future prospective studies evaluating both resistance and consumption across centers are needed.
In summary, we identified trends in and predictors of carbapenem consumption across 20 North American hospitals. Increasing rates of carbapenem consumption in DDD/1,000 PD were observed over the study period and were independently predicted by increasing noncarbapenem beta-lactam consumption, hospital and ICU bed size, and aspects of antibiotic stewardship interventions. Additional studies are needed to define the impact of uniformly applied stewardship interventions on these metrics and their ultimate effects on carbapenem resistance across a wider range of diverse health care institutions.
MATERIALS AND METHODS
A multicenter, retrospective, cross-sectional survey of antimicrobial usage was conducted among participating hospitals within the Making a Difference in Infectious Diseases (MAD-ID) Research Network (http://mad-id.org/the-mad-id-research-network/). This network is composed of 181 institutions, 31% of which are university-affiliated institutions and 52% of which are not-for-profit, nonuniversity hospitals (30–32). Nearly all (94%) of the network institutions treat adult populations, while 47% care for pediatric populations and 25% serve the long-term-care population (30–32). This study was approved by the institutional review board (IRB) at each participating institution. The IRB of Wayne State University (Detroit, MI) served as the coordinating IRB of record and assisted with all data use agreements (DUA).
Data elements and definitions.Data elements were collected at participating hospitals for the study period of January 2011 through December 2013. Each center reported antimicrobial use for a variety of beta-lactam antibiotics in defined daily doses (DDDs) or days of therapy (DOTs) in addition to hospital or health system demographic information. Defined daily doses were defined according to the World Health Organization anatomical therapeutic chemical (ATC) standard as grams of drug at a given facility summated over each quarter (33). Days of therapy data were defined as the tally of all administrations or orders for a given facility summated over each quarter (22). For this analysis, orders and administrations were combined for DOTs. All data elements were recorded by volunteer survey respondents at each study site. Antimicrobial consumption data were collected as quarterly repeated measures from pharmacy purchasing records (for DDDs) or from the electronic medical record (for DOTs). Hospital-specific demographics were self-reported by each participating site.
Demographics and stewardship program characteristics.Self-reported hospital demographic data included primary financial designation (i.e., for profit or not for profit), academic or government affiliation, patient populations served (e.g., adult, pediatric, neonatal, or long-term care), number of licensed general ward beds, number of licensed intensive care unit (ICU) beds, the self-reported antimicrobial stewardship strategies employed (e.g., use of restrictive order forms, circulation of antibiograms), classification of the antimicrobial stewardship strategies employed (e.g., formulary restriction of carbapenems, prospective audit and feedback of carbapenem use, computer-assisted clinical decision support, institutional guidelines, order forms, etc.), and level of restrictions placed on carbapenem use (e.g., on formulary, not restricted, on restricted formulary, or nonformulary).
Consumption of antimicrobial agents.Participating hospitals reported their rates of consumption of the following beta-lactams: cefepime, doripenem, ertapenem, imipenem-cilastatin, meropenem, and piperacillin-tazobactam. The consumption metrics that were reported (e.g., DDDs and DOTs) are commonly used to describe antimicrobial utilization at the hospital level and are standardized to patient days (PD) of hospitalization (e.g., metric/1,000 PD) to adjust for differences in hospital occupancy, which contribute to each center’s at-risk denominator (34). Patient days were defined as the tally of patients admitted to each participating facility between census-taking times summated over each quarter, which is distinct from the more modern NHSN days present (34). Carbapenem consumption rates were considered to represent a composite of all rates of consumption of any of the following agents: ertapenem, doripenem, imipenem-cilastatin, and meropenem. Ertapenem was considered together with the group 2 carbapenems, as it is often used as a step-down agent. Two commonly utilized broad-spectrum agents, piperacillin-tazobactam and cefepime, were classified as noncarbapenem beta-lactams.
Statistical analysis. (i) Univariate analyses.Demographic predictors of increasing consumption in DDDs and DOTs were analyzed separately at the univariate level. Descriptive statistics were calculated for hospital demographic variables. Categorical variables were analyzed using chi-square testing or the Fisher’s exact test, as appropriate. Differences in carbapenem and noncarbapenem beta-lactam consumption rates were summarized according to hospital demographic characteristics as a ratio of the median noncarbapenem (NC) consumption rates to the carbapenem (C) consumption rates (i.e., NC/C ratio). Variability in consumption measures over the study period was quantified as percent coefficient of variation (i.e., SD/mean × 100). Missing data were left as missing in all analyses.
(ii) Multivariate analyses.Multivariate models were fitted to the observed consumption data with time (in quarters) forced into all regression models, as this was used to determine the presence of temporal linear trends in consumption rates. Repeated time-dependent consumption rates (i.e., Y = DDD or DOT/1,000 PD) were modeled using a linear mixed-effects approach. In this approach, consumption data from all instituions were modeled using fixed-effects (e.g., global intercept and slope estimates across all hospitals) and random-effects (e.g., random intercept and slope estimates for each hospital) parameter estimation, as previously described (35). To account for nonindependence of observations across centers, the following variance-covariance structures were evaluated: (i) single variance with no covariance (identity), (ii) equal variances with no covariance (independent), (iii) equal variances with compound symmetrical covariance (exchangable), and (iv) direct covariance estimation (unstructured). A random slope and intercept mixed-effects variance-covariance matrix, for example, would estimate each of the parameters shown in Table 5. Noncarbapenem consumption and hospital demographics were iteratively entered into the base model using a forward-stepwise method for covariate evaluation. To evaluate inherent differences that might exist between DDD-reporting and DOT-reporting centers over time, a time-by-metric interaction was evaluated a priori in all models. Stepwise improvements in multivariate model fitness were evaluated using minimization of Akaike’s information criterion (AIC).
Random slope and intercept mixed-effects variance-covariance matrix
The final covariate-adjusted multivariate linear mixed-effects models were selected by comparing the final competing models for each metric using likelihood ratio testing. Competing models were assessed by taking the difference in −2× log-likelihood function values between a model with n + 1 predictors and a model with n predictors (i.e., the same model with one fewer predictor). Goodness of fit for the final models was evaluated by analysis of regressions of observed and fitted consumption values, mean weighted prediction error (bias) values, and bias-adjusted mean weighted squared prediction error (imprecision) values (weighted according to the standard error of the prediction) and by visual inspection of the homoscedasticity of standardized residuals. Intercooled Stata, version 14.2 (Statacorp, College Station, TX), was utilized for all model fitting procedures and diagnostics.
ACKNOWLEDGMENTS
S.L.D. and M.H.S. disclose that they have received grant support for this study from Merck & Co., Inc. (Antimicrobial Stewardship Research Grant Program). S.L.D. has also received research support from Allergan. N.J.R., J.L.W., J.A.B., D.A.G., and M.J.R. disclose that they have received no financial support related to this study.
N.J.R. reports receipt of honoraria from the American Society of Health-System Pharmacists and research funds from Hartford Hospital. J.L.W. reports receipt of research funds from MAD-ID outside the current study. J.A.B. reports that he has received grant support from Merck & Co., Inc., and has served on advisory boards for Nabriva, QPEX, Shionogi, and Tetraphase. D.A.G. reports that she has served as an Advisory Board attendee for OpGen and Curetis and has received grant support from Merck & Co., Inc., and from Pfizer. M.J.R. reports that he received grant support and consulted from Merck & Co., Inc., and Allergan. M.H.S. reports that grant support from Merck & Co., Inc., and has consulted for Bayer and Paratek. S.L.D. reports she serves as a consultant for Merck & Co., Inc., and Allergan.
We thank the following participants from the MAD-ID Research Network for collecting and reporting patient-level data: Drew Zimmer (Northeast Regional Medical Center), Jonathan Edwards (Huntsville Hospital), Vicki Barnes and Christy Rooks (Stormont Vail Healthcare), Tanea Womack and Daniel Chastain (Phoebe Putney Memorial Hospital), Jesse Jacob (Emory University Hospital Midtown), Jordan Wong and Sheetal Kandiah (Grady Memorial Hospital), Rachel Kenney (Henry Ford Hospital), Lynn Wardlow (University of Louisville Hospital), Sonal Patel (Georgia Regents Medical Center), Carrie Sorenson and Joan Galbraith (St. Alexius Medical Center), Melinda Monteforte (Stony Brook University Hospital), Annette Brady and Debra Boswell (West Georgia Health System), Kevin Chapple (Shore Regional Health), Edward Cubick (Good Samaritan Hospital), Kazumi Morita (Temple University Hospital), Yasser Mohamed (North Bay Regional Health Center), and Jamie Kisgen (Sarasota Memorial Hospital).
FOOTNOTES
- Received 12 February 2019.
- Returned for modification 5 March 2019.
- Accepted 29 April 2019.
- Accepted manuscript posted online 6 May 2019.
Supplemental material for this article may be found at https://doi.org/10.1128/AAC.00327-19.
- Copyright © 2019 American Society for Microbiology.
