Abstract
Background There is conflicting evidence of the role of antibiotic in driving antimicrobial resistance. Existing studies asses use-resistance relationships typically only for few pathogens, for narrow time window, or only in one geographic space, and/or for few pathogen-antibiotic combinations and often only one pathogen and antibiotic (class) at a time. No systematic evidence of whether some antibiotics are more efficient at selecting for resistance than others. Here, we tested the hypothesis that each antibiotic class exerts selection for the antimicrobial susceptibility patterns resistant to it, using data on four major pathogens in the US Veterans Affairs Healthcare System.
Methods We analysed clinical microbiology data from electronic health records from patients admitted to 138 Veterans Affairs Medical Centers with acute care wards across the USA from Feb 1, 2007, to Dec 31, 2021. We used clinical isolates classified as hospital-onset Staphylococcus aureus, Escherichia coli, Klebsiella pneumoniae, and Pseudomonas aeruginosa. We quantified inpatient antibiotic prescribing as days of therapy (DOT) per 1000 patient-days and the antibiogram incidence as the number of incident isolates with that antibiogram per 1,000 admissions. We performed time trend analyses for antibiogram incidence and antibiotic prescribing using generalized estimating equations and reported average annual percentage changes (AAPC). We performed hierarchical multinomial logistic regressions to estimate the effect of inpatient facility-level prescribing on antimicrobial susceptibility patterns. We included calendar time, facility characteristics, and community prevalence of a respective antibiogram as covariates to account for ecological confounding.
Results From 2007 to 2021, the incidence of hospital-onset isolates declined for all four pathogens except E. coli isolates resistant to third-generation cephalosporins (3GC) and co-resistant to either fluoroquinolones or beta-lactam/beta-lactamase inhibitors (BL/BLI), or both. Over the same period, antibiotic prescribing generally remained stable or decreased, with the notable exception of 3GC prescribing, which increased from 2007 to 2019 (AAPC: 2.4%, 95% CI: 1.3%-3.5%). In general, higher facility-level use of a given antibiotic was associated with increased resistance to that antibiotic, although exceptions were observed. Fluoroquinolones consistently selected for resistance across the four pathogens. In S aureus, each additional 14-day course of fluoroquinolones (per 1,000 patient-days) was linked to a 5.9% (95% CI: 0.8%-11.2%) increase in the odds of isolating a fluoroquinolone-resistant, macrolide-susceptible, methicillin-resistant phenotype. Anti-staphylococcal beta-lactams were not associated with methicillin-resistant S aureus. In Enterobacterales, each additional 14-day course of 3GC treatment increased the odds of isolating 3GC- and BL/BLI-resistant E coli by 7.4% (95% CI: 1.8%-13.4%) and K pneumoniae by 4.1% (95% CI: 0.0%-8.7%). For P aeruginosa, antipseudomonal carbapenem use selected for carbapenem-resistant isolates; the largest effect was observed in carbapenem-resistant phenotypes susceptible to fluoroquinolones and 3GC, where each additional 14-day course increased the odds by 22.6% (95% CI: 13.4%-32.7%).
Conclusions Our findings demonstrate that fluoroquinolones play a key role in selecting for resistance across multiple pathogens, underscoring their significance as a driver of AMR. Although these results highlight the importance of cautious fluoroquinolone use and targeted stewardship strategies, we also observed that reductions in fluoroquinolone prescribing were offset by increased 3GC prescribing and co-resistance in E coli. Consequently, effective AMR mitigation requires comprehensive stewardship approaches that address multiple antibiotic classes in tandem rather than in isolation.
Introduction
Antimicrobial use has long been recognized as a major driver of antimicrobial resistance (AMR).1 However, our understanding of the complex relationship between antibiotic use and resistance remains incomplete, partly because of methodological challenges in assessing this association.2 Many studies have demonstrated ecological correlations for single pathogen-single antibiotic combinations at regional, national, or international scales.3–11 In these studies, the evidence linking antibiotic use to resistance is often contradictory, with some studies showing significant correlations for certain pathogens, while others reveal no clear association, particularly for commensal bacteria.3–6 A landmark study of 26 European countries used data from 1997 to 2002 and correlated the use of β-lactams and fluoroquinolones with resistance in Streptococcus pneumoniae, S. pyogenes, and Escherichia coli isolates.7 The study showed that nations with the highest outpatient antibiotic consumption had the greatest prevalence of antibiotic-resistant bacteria, whereas low-consumption countries had much lower resistance rate. In the United States, cross-sectional analyses have similarly examined multiple pathogen-antibiotic pairs.3,11 Olesen and colleagues described the landscape of correlations for 72 pathogen-antibiotic combinations across U.S. states using outpatient antibiotic prescription data covering 2011-2014.3 The most positive significant correlations involved macrolides, fluoroquinolones, and cephalosporins. An analysis of five pathogen-antibiotic combinations in eight U.S. hospitals showed that high antimicrobial use was not necessarily associated with high resistance.11 For example, the use of anti-staphylococcal agents and percentage of methicillin-resistant Staphylococcus aureus (MRSA) were not correlated.
Few studies have investigated the potential impact of co-selection, i.e., the process that the use of one antimicrobial agent promotes the selection and persistence of resistance not only to that agent but also to other antimicrobials, on the relationship between antibiotic use and resistance.12–14 In multivariate models for S pneumoniae, macrolide consumption explained as much variation in penicillin resistance as penicillin use, implying that macrolide exposure co-selected penicillin-resistant strains.15 Pouwels and colleagues found that increased prescribing of amoxicillin in the community was associated not only with more amoxicillin-resistant E coli in urine isolates, but also with higher ciprofloxacin and trimethoprim resistance in those E coli.13 These studies typically aggregate data at a high level (e.g., states, regions, nations) and examine the relationship between total antibiotic consumption and overall resistance prevalence. Such analyses provide a broad view of the association but can mask within-group variability, may suffer from ecological fallacies, and are often subject to reverse confounding by not capturing the precise time lags between antibiotic exposure and the development of resistance.2
Several other studies investigated the relationship between antibiotic use and resistance on an individual-level.16 In a case-control study, Weber and colleagues found that exposure to fluoroquinolones, specifically levofloxacin and ciprofloxacin, was a significant risk factor for the subsequent isolation of MRSA in hospitalized patients, but not for methicillin-susceptible S. aureus (MSSA), highlighting the role of fluoroquinolones in co-selecting MRSA.17 However, these do not typically account for the indirect effects on AMR due to transmission and thus fail to capture population-level dynamics – they may even yield associations that have the opposite sign from the effects at the population level.18
To our knowledge, no study has yet examined the impact of antibiotic use on susceptibility patterns across multiple antibiotics (‘antibiograms’) and multiple pathogens accounting for both individual-level and group-level effects. Under the guiding hypothesis that each antibiotic exerts selection for the antibiograms that are resistant to it, we conducted hierarchical multinomial logistic regression analyses for four major community- and hospital-associated pathogens in the Veterans Affairs Healthcare system (VAHS) in the United States (U.S.): S aureus, E coli, K pneumoniae, and P aeruginosa. We included data on antimicrobial susceptibility patterns from individual-level clinical isolates and facility-level antibiotic use from 138 Veterans Affairs Medical Centers (VAMCs) across all U.S. census regions from 2007-2021. By linking patient-level outcomes to facility-level antibiotic use, we aimed to investigate how institutional prescribing practices shape the resistance patterns observed in patients. The scope of our study - the number of species, diversity of geographic sites, and long period of study - provides a unique opportunity to evaluate this question.
Methods
Study design and participants
In this US nationwide retrospective cohort study, we analysed clinical data from all patients admitted to all VA Medical Centers with acute care wards (either medical or surgical wards or intensive care units) in the USA from Feb 1, 2007, to Dec 31, 2021. We define the pre-COVID-19 period as Feb 1, 2007 till Dec 31, 2019 and the COVID-19 period as Jan 1, 2020 till Dec 31, 2021. We excluded facilities from the study if they did not provide acute care or if they did not report data to the VA’s facility complexity assessment (e.g., level and type of care provided) in all eligible years during the study period. For trend analyses of antibiotic prescribing, we included 138 VAMS. For antibiotic use-resistance analyses, we included facilities according to a patchwork approach (see below, Table 1). The patient age was determined from the patient’s date of birth. The biological sex of patients was determined from their birth certificates. Race and ethnicity were collected by self-report.
Patient age was determined from the date of birth. The biological sex of patients was determined from their birth certificates. Race and ethnicity were collected by self-report. Only the most common groups are displayed in this table. A more detailed table on patient demographics can be found in the appendix (pp X).
This study was approved by the University of Utah Institutional Review Board and the VA Salt Lake City Health Care System Research and Development Office, which waived patient consent because the project relied on retrospective analysis of existing patient records. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline for cohort studies.
Procedures
Clinical diagnostic cultures were extracted from electronic health records, as previously described.19 Our analysis only included results from clinically directed cultures, ordered to evaluate suspected infection. We defined a 30-day incident isolate as a clinical culture where no other clinical culture of the same species was identified in the previous 30 days and used it as a proxy measure for clinical infection incidence. When multiple species were recovered from the same sample or from different samples collected during the same hospitalisation, each incident isolate was separately counted in the analysis of a given species.
Isolates were classified as community-onset when the specimen was obtained 3 days or fewer after admission and as hospital-onset if the specimen was obtained more than 3 days after admission and prior to discharge. We included only isolates classified as hospital-onset in our analysis.
Antimicrobial susceptibility test (AST) results were used to identify whether a culture was resistant to a specific antibiotic. Interpretation of ASTs was based on the reported minimum inhibitory concentrations and categorised into susceptible (S), intermediate (I), or resistant (R).
We grouped intermediate and resistant to R in our analyses. Interpretation of AST results adhered to CLSI breakpoint revisions. We defined antibiograms as the susceptibility profile of a bacterial isolate to a panel of key antibiotics chosen based on their clinical relevance for first- and second-line treatment of the target pathogens (Table 1, appendix p.2 Table S2). An overview of pathogen-antibiotic combinations and the respective antibiograms can be found in the appendix (section S4). We refer to isolates susceptible to all three key antibiotic classes as susceptible isolates.
Outcomes
Our main outcomes were antibiogram incidence in the study population, facility-level inpatient antibiotic prescribing rates, and regression coefficients obtained from fitting multinomial logistic regression models to datasets of the target antibiotic-pathogen combinations. We defined antibiogram incidence as the number of incident isolates with a specific antibiogram per 1,000 admissions. We defined inpatient antibiotic prescribing rates as the days of therapy (DOT) per 1,000 patient-days.
Statistical Analysis
For trends in inpatient antibiotic prescribing and antibiogram incidence, we performed robust Poisson regression using generalized estimating equations (GEE).20 Clustering by VAMC was modelled with an autoregressive correlation structure within a facility. We used calendar year as a continuous variable as the main exposure and the number of admissions as offset. We adjusted for major hospital characteristics, including census region, facility complexity and rurality, and patient volume, and therefore, accounted for spatial variation and population density (appendix section S2). In addition, we estimated the multiplicative average annual percentage change (AAPC) based on the estimated marginalized time trend from our GEE analysis.
For antibiotic use-resistance analyses, we performed multinomial logistic regressions where the outcome was the probability that an isolate of that pathogen has a particular antibiogram, and the main exposures of interest are the cumulative facility-level antibiotic days of therapy in the previous 14 days per 1,000 patient days. We adjusted for potential ecological confounding from calendar time, community prevalence by including the incidence rate of antibiograms collected within 3 days of admission during the previous 90 days, and facility characteristics (e.g., census region, facility complexity, facility rurality). For each pathogen, we used a mixed effect multinomial logistic regression model in which the different VA facilities are units within which frequencies of antibiograms change over time. A more detailed description of the model is given in the appendix (section S2)
For some antibiotic classes, such as fluoroquinolones, a substantial proportion of isolates did not have reported AST results, complicating the construction of antibiograms. The incidence and proportion of these missing AST results varied over time, indicating selective reporting.21 To address this issue, we applied a threshold-based patchwork approach in our main analysis. For each antibiotic-pathogen combination and each calendar year, we included only those facilities where the proportion of missing AST results did not exceed 30% in each quarter of that year. By excluding facility–year combinations with excessive missing data, we minimized the likelihood that our findings would be skewed by selective reporting, thus reducing the risk of bias. As a result, the final dataset is a composite of facilities contributing data from different calendar years. We performed sensitivity analyses to evaluate alternative methods for handling missing AST results. More information on the performed approaches can be found in the appendix (section S3).
To assess the impact of variation in facility-level antibiotic prescribing on the ability to recover true regression coefficients, we conducted a simulation analysis for a multinomial logistic regression model, assuming the model was correctly specified (the model used in the simulation was the same used for model fitting). This analysis examined how these factors influence bias, confidence interval coverage, the envelope, and the average confidence interval width. More details can be found in the appendix (section S6).
Role of the funding source
The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.
Results
Overall, antibiogram incidence trends showed a general decline from 2007 to 2021, with only a few notable exceptions (Figure 1, appendix Table S5). For S aureus, K pneumoniae, and P aeruginosa the incidence of isolates resistant to all three key antibiotic classes (R-R-R) declined from 2007 to 2021. Specifically, the incidence of S aureus isolates resistant to fluoroquinolones, ASBL, and macrolides increased with an AAPC of −9.5%, (95%CI: −11.6%, −7.3%; p-value < 0.0001). The incidence of K pneumoniae isolates resistant to fluoroquinolones, 3GC, and BL/BLI increased with an AAPC of −5.1% (95%CI: −9.8%, −0.2%; p-value = 0.0401). The incidence of P aeruginosa resistant to fluoroquinolones, BL/BLI, and carbapenems increased with an AAPC of −8.7% (95%CI: −12.8%, −4.5%; p-value < 0.0001) respectively. However, the incidence of E coli isolates resistant to fluoroquinolones, 3GC, and BL/BLIs increased from 2007 to 2021 with an AAPC of 3.4% (95%CI: 0.5%, 6.5%; p-value = 0.0204). The incidence of E coli isolates resistant 3GC increased. In particular, the incidence of R-R-S increased with an AAPC of 4.6% (95%CI: 0.2%, 9.2%; p-value = 0.0421). The incidence of S-R-R had an increasing trend (AAPC: 2%, 95%CI: −1.3%, 5.5%; p-value = 0.2372). The incidence of E coli resistant to fluoroquinolones and BL/BLI but susceptible to 3GC declined with an AAPC of −5.4% (95%CI: −7.3%, −3.5%; p-value < 0.0001). Fluoroquinolone-resistant E coli susceptible to both 3GC and BL/BLI declined with an AAPC of −8.1% (95%CI: −11.4%, −4.7%; p-value < 0.0001). The incidence of susceptible isolates declined from 2007 till 2021 for S aureus, E coli, and P aeruginosa with an AAPC of 3.8% (95%CI: −6%, −1.5%; p-value = 0.0012), −5.1% (95%CI: −6.8%, −3.4%; p-value < 0.0001), and −2.6% (95%CI: −4.9%, −0.2%; p-value = 0.031), respectively. For K pneumoniae, the incidence of susceptible isolates declined from 2007 to 2021 (AAPC: −5.5%, 95%CI: −7.1%, −3.9%; p-value < 0.0001)
Patchwork data was used for the analysis. (A) Points represent the incidence in infections with the respective antibiogram. Lines and grey bands are linear regression curves and corresponding 95% CIs. (B) AAPC obtained from time trend analyses using generalised estimating equations for the respective antibiogram incidence. The interaction with the COVID-19 period was tested and a separate time trend estimate is reported if the interaction term was significant (p-value < 0.05). An AAPC greater than zero indicates an increasing trend and an AAPC less than zero indicates a declining trend. AAPC=average annual percentage change.
Inpatient antibiotic prescribing declined for most antibiotic classes (Figure 2, appendix Table S18). The largest decline was estimated for fluoroquinolones from 2007 to 2022 (AAPC: −8.4%, 95%CI: −9.1%, −7.7%; p-value < 0.0001). Carbapenem prescribing rates increased from 2007 to 2011 with an AAPC of 5.3% (95%CI: 3.1%, 7.6%; p-value < 0.0001) and subsequently declined from 2012 to 2022 (AAPC: −3.7%, 95%CI: −6.2%, −1.1%; p-value = 0.0062). BL/BLI prescribing rates stayed stable from 2007 till 2014 (AAPC: −0.1%, 95%CI: −0.8%, 0.6%; p-value = 0.7993). After a nationwide shortage of piperacillin/tazobactam in 2015, BL/BLI prescribing rates declined (AAPC: −2.3%, 95%CI: −3.9%, −0.7%; p-value = 0.0046). Prescribing rates increased for 3GC from 2007 to 2019 (AAPC: 2.4%, 95%CI: 1.3%, 3.5%; p-value < 0.0001) but subsequently declined during pandemic years 2020 to 2022 (AAPC: −5%, 95%CI: −7.7%, −2.2%; p-value = 0.0005). Prescribing rates of ASBLs and macrolides declined slightly and remained stable prior to the COVID-19 pandemic (2007-2019) with AAPCs of −1% (95%CI: −1.7%, −0.2%; p-value = 0.0103) and −0.1% (95%CI: −1%, 0.9%; p-value = 0.8488), respectively. ASBL prescribing rates increased during COVID-19 pandemic years (2020-2022) with an AAPC of 4.6% (95%CI: 1.5%, 7.8%; p-value = 0.0038) whereas prescribing of macrolides declined (AAPC:-17.9%, 95%CI: −21%, −14.7%; p-value < 0.0001).
(A) Bar plots show the overall antibiotic prescribing rates, measured as days of therapy per 1,000 patient days, from 2007 till 2022 in 138 VA medical centres for key antibiotic classes. (B) AAPC obtained from time trend analyses using generalised estimating equations for antibiotic classes depicted in A. The interaction with the COVID-19 period was tested and a separate time trend estimate is reported if the interaction term was significant (p-value < 0.05). An AAPC greater than zero indicates an increasing trend and an AAPC less than zero indicates a declining trend. AAPC=average annual percentage change.
Our analysis relating facility-level antibiotic use and antibiogram patterns revealed a consistent positive association between recent fluoroquinolone use and the odds of fluoroquinolone resistance (Figure 3). Across all target pathogens, the rate of fluoroquinolone prescribing in the preceding 14 days was linked to an increase in the odds of obtaining antibiograms showing resistance to fluoroquinolones. For S aureus, the most pronounced association was observed in fluoroquinolone-resistant MRSA isolates (R-R-S) which had an odds ratio of 1.0041 (95% CI: 1.0006, 1.0076). For an extra 14-day course of fluoroquinolone treatment (per 1,000 patient days), we expect about a 5.9% (=1.004114-1) increase in odds of isolating a fluoroquinolone-resistant, macrolide-susceptible MRSA strain. Macrolide prescribing was associated with an increased odds of isolating fluoroquinolone-susceptible, macrolide-resistant MRSA (S-R-R, odds ratio = 1.0050, 95%CI: 1.00, 1.0100). ASBL prescribing did not increase the odds of ASBL resistance for either of the studied antibiograms. In fact, ASBL prescribing was associated with a decrease in odds ratio of fluoroquinolone-resistant MRSA (0.9912, 95% 0.9834, 0.9990).
(A) Coefficient estimates and confidence intervals (80% and 95%) for pathogen-specific key antibiotic classes. S aureus: fluoroquinolones, anti-staphylococcal beta-lactams, macrolides. (B) E coli: fluoroquinolones, 3rd generation cephalosporins, beta-lacatam/beta-lactamase inhibitors. (C) K pneumoniae: fluoroquinolones, 3rd generation cephalosporins, beta-lacatam/beta-lactamase inhibitors. (D) P aeruginosa: fluoroquinolones, beta-lacatam/beta-lactamase inhibitors, carbapenems. Estimates represent the odds ratio with respect to the reference categories. For all analyses, the reference category was chosen as the antibiogram susceptible to all key antibiotic classes.
For E coli, the largest increase in odds was estimated for isolates resistant to fluoroquinolones but susceptible to 3GC and BL/BLIs (Figure 3B, R-S-S) and for isolates resistant to both fluoroquinolones and BL/BLI (Figure 3B, R-S-R). In K pneumoniae, the largest odds ratio was estimated for isolates resistant to both fluoroquinolones and BL/BLIs (Figure 3C, R-S-R, 1.0061, 95%CI: 1.0018, 1.0105). For P aeruginosa, fluoroquinolone use was associated with an odds ratio of 1.0028 (95%CI: 1.0008, 1.0047) for antibiograms resistant only to fluoroquinolones (Figure 3D, R-S-S) and an odds ratio of 1.0021 (95%CI: 0.9997, 1.0045) for antibiogram resistant to all three key antibiotic classes.
For both E coli, the use of 3GC increased the odds of isolates resistant to 3GC and BL/BLI but susceptible to fluoroquinolones (S-R-R, odds ratio: 1.0051, 95%CI: 1.0013, 1.009). We did not find an association between 3GC use and isolates resistant to both 3GC and fluoroquinolones (R-R-R and R-R-S antibiograms). Similarly, for K pneumoniae, previous 3GC use was associated with an increase in odds of MDR isolates resistant to both 3GC and BL/BLI (S-R-R, odds ratio: 1.0029, 95%CI: 0.9999, 1.0060) but not for XDR isolates additionally resistant to fluoroquinolones (R-R-R). For both Enterobacterales species, previous use of BL/BLI was associated with an increase in odds of resistant isolates (S-S-R, S-R-R, R-S-R antibiograms, appendix Table S10, S13), except for isolates resistant to all three key antibiotic classes (R-R-R).
For P aeruginosa, the previous use of antipseudomonal carbapenems was associated with an increased odds of isolates resistant to carbapenems. The largest increase in odds was estimated for carbapenem-resistant isolates susceptible to both fluoroquinolones and 3GC (S-S-R, odds ratio: 1.0147, 95% CI: 1.0090, 1.0204), followed by XDR P aeruginosa isolates (R-R-R, odds ratio: 1.01, 95% CI: 1.0056, 1.0143) and MDR P aeruginosa isolates resistant to fluoroquinolones and carbapenems (R-S-R, odds ratio: 1.0097, 95%CI: 1.0045, 1.0150). For For most pathogen-antibiogram combinations, the importation of antibiograms by community-onset isolates was significantly associated with an increase in odds of isolates with that specific antibiogram (appendix section S4).
To evaluate how dataset characteristics affect regression coefficient estimation in multinomial logistic regression, we conducted a simulation analysis under a correctly specified model (Figure 4, appendix section S6). We focused on how decreased variability in antibiotic prescribing influences bias, coverage, the envelope (range) of estimated coefficients, and confidence interval precision. Although a smaller variance in antibiotic prescribing (Drug 2 and 3) did not affect the median bias, it increased the bias variance (Figure 4A). Furthermore, as the variability in antibiotic prescribing declined, both the envelope and the average width of the 95% confidence intervals increased (Figure 4C), reflecting greater uncertainty in the coefficient estimates. The coverage, the proportion of 95% confidence intervals containing the true coefficient, remained unchanged.
Three prescribing distribution of drug classes were simulated with large (Drug 1), medium (Drug 2), and small variance (Drug 3) based on log-normal distribution with parameters described in the appendix (section S6). The results were summarized across 100 simulated data sets. (A) Boxplot and distribution of bias in regression coefficient estimates, defined as the difference between the estimate and the true coefficient. (B) Barplots of coverage of 95% confidence intervals. This panel illustrates the proportion of simulation runs in which the true regression coefficient is captured by the estimated 95% confidence intervals. (C) Envelope (lines, left y-axis) and Average Confidence Interval Width (bar plots, right y-axis). The envelope (range) of estimated regression coefficients is shown as line intervals (left y-axis) and the average width of the 95% confidence intervals is represented by bar plots (right y-axis).
Discussion
In this nationwide study, we observed notable shifts in phenotypic incidence trends among four major pathogens in the VAHS from 2007 to 2022, in parallel with changes in antibiotic prescribing patterns. Furthermore, we found that facility-level antibiotic use was significantly associated with resistance patterns at the individual-level for most pathogen-antibiotic combinations.
A particularly notable finding was the broad decline in fluoroquinolone resistance across all four pathogens and eleven corresponding phenotypes, except for E coli resistant to both fluoroquinolones and 3GC. These observations were supported by significant associations between prior facility-level fluoroquinolone prescribing and fluoroquinolone resistance for all four target pathogens. Nevertheless, the persistence of co-resistance to fluoroquinolones and 3GC in E coli isolates, despite an overall decline in fluoroquinolone prescribing, underscores the complexity of resistance dynamics. In fact, overall 3GC prescribing increased over the study period, potentially as a replacement for fluoroquinolones, and higher previous facility-level 3GC exposure was generally associated with an increase in odds of antibiograms with 3GC resistance. However, this relationship was less pronounced in isolates already resistant to fluoroquinolones, though the estimates had considerable uncertainty. In addition, while the incidence of E coli resistant to both 3GC and BL/BLI had an increasing trend, the incidence of BL/BLI-resistant E coli (susceptible to 3GC) declined. Similarly, the incidence of isolates resistant to fluoroquinolones but susceptible to 3GC declined, suggesting that co-selection by 3GC may help maintain or even drive fluoroquinolone and BL/BLI resistance in MDR E coli. These findings suggest while reductions in antibiotic use may lead to a decrease in resistance incidence, reductions in prescribing in one antibiotic class may be offset by increased prescribing of another, potentially leading to collateral resistance. Ongoing surveillance is therefore crucial to detect and address evolving co-resistance dynamics. Given that once resistance emerges, large reductions in consumption are required to meaningfully reduce resistance levels,22 early and multifaceted stewardship and infection control interventions are essential to controlling AMR.
Prescribing to carbapenems and BL/BLs was also associated with subsequent resistance in P aeruginosa, with antipseudomonal carbapenems showing higher odds ratios than fluoroquinolones. These results align with earlier reports linking carbapenem use to XDR P aeruginosa infections, and fluoroquinolone, cephalosporin, and carbapenem use to MDR or XDR P aeruginosa infections.23,24 In addition, we observed a decline in the incidence of all studied phenotypes in this study, aligning with the previously described decline in hospital-associated infections in the VAHS and the decline of MDR and XDR P aeruginosa in other regions.21 The declining trends were paralleled with reductions in prescribing to fluoroquinolone, carbapenems and BL/BLIs. While previous studies have linked the restriction of carbapenem use to a decline in carbapenem resistance incidence rates,25 the relative contribution of each antibiotic class to these declines remains unclear, warranting further research to disentangle their individual impacts on resistance selection and pathogen fitness.
Our results for S aureus suggest that MRSA incidence may not be driven by prescribing of anti-staphylococcal beta-lactams but rather by prescribing of fluoroquinolones and macrolides, echoing previous findings that linked fluoroquinolone use but not beta-lactam use to MRSA colonization.17,26–28 This pattern, together with evidence that fluoroquinolones promote the loss of MSSA colonization more than other anti-MSSA agents,29 suggest that fluoroquinolones may promote MRSA selection through ecological disruption, eradicating MSSA strains and creating a niche for resistant strains to proliferate.30 The lack of association between anti-staphylococcal beta-lactams and MRSA incidence in our study may reflect the distinct evolutionary pathway of methicillin resistance. Historically, penicillin use – not methicillin – has been implicated as the primary driver of mecA acquisition in S aureus.31 Our study underscores the nuanced roles different antibiotic classes play in shaping pathogen dynamics.
Our study has several limitations. First, although our model explicitly accounts for prior antibiotic exposure and includes multiple facility-level characteristics to mitigate confounding, we cannot rule out residual confounding, including reverse causation. For example, facilities with high prevalence of MRSA may refrain from the use of antibiotics known to be associated with MRSA (and the other way around) which could paradoxically link higher use of anti-staphylococcal beta-lactams to lower MRSA prevalence. Second, although we incorporated a 14-day lag to address reverse causality, this may not fully capture such dynamics. Third, we included only inpatient antibiotic prescribing into our model. For many pathogens, community-based prescribing may contribute significantly to selection pressure and subsequent hospital importation of resistant organisms. Fourth, we classified isolates as hospital-onset if they were collected at least 72 hours after admission, which can misclassify some community-acquired infections. Fifth, our simulation analysis suggests that variability in previous antibiotic prescribing is crucial for precisely estimating regression coefficients. For antibiotic classes with limited variability in their prescribing, coefficient estimates can carry high uncertainty, potentially explaining some of our null findings. Sixth, while antibiograms can be used as proxies for bacterial lineages, future studies incorporating whole-genome sequencing alongside phenotypic profiles are needed to more clearly elucidate the relationships between antimicrobial use and microbial ecology. Finally, the patient population in our study is limited to enrolled veterans in the USA with access to the VA Healthcare System. The veteran population differs from the general population in their health characteristics, including their demographics, underlying health conditions, exposures, and health-care use patterns. However, our study included data from up to 138 medical facilities with varying complexity levels, from different geographical regions in the USA, and from more than 15 years.
Overall, our findings underscore the complexity of antibiotic selection pressure and its variable effects depending on drug class, pathogen, co-resistance, and resistance mechanisms. We highlight how reductions in one antibiotic class can inadvertently drive reliance on – and potential resistance to – another. To effectively combat AMR, stewardship efforts must be comprehensive, targeting multiple antibiotic classes in tandem rather than in isolation. Such coordinated strategies, coupled with robust surveillance and early interventions, are essential to contain the rise of multidrug-resistance and preserve the efficacy of existing therapies.
Data sharing
Individual-level patient data cannot be provided due to Veteran Affairs privacy practices. Artificial data that represents the original data will be provided along with a data dictionary that defines each field in the dataset and supporting documentation (statistical and analytic code) are published on Github (https://github.com/tm-pham/mind_aim2-2_AMR).
Declaration of interests
YHG serves on the Scientific Advisory Boards of Day Zero Diagnostics and of Decoy Therapeutics and received US National Institutes of Health grants (R01AI32606, R01AI53521, and R21AI72369, all not related to this project). KK received support from Veterans Affairs Health Systems Research (IIR 21–273) and for two studies from BioMerieux Clinical (IRB_00170297; IRB_00140336). All other authors declare no competing interests.
Acknowledgments
This work was supported by the US Centers for Disease Control and Prevention (Modeling and Simulation to Support Epidemiological Decision-Making in Healthcare Settings, Modeling Infectious Diseases in Healthcare Program; U01CK000585). MS was supported by the Agency for Healthcare Research and Quality (R01HS025175) and the Department of Veterans Affairs (1I50HX002731–01). We thank the Department of Veterans Affairs for the supported infrastructure and data resources. The views expressed in this article are those of the authors and do not necessarily represent the position or policy of the Department of Veterans Affairs, the US Centers for Disease Control and Prevention, the US Government, or any of the affiliated institutions.