Patterns, Predictors, and Intercenter Variability in Empiric Gram-Negative Antibiotic Use Across 928 United States Hospitals

Abstract

Background: Empiric antibiotic use among hospitalized adults in the United States (US) is largely undescribed. Identifying factors associated with broad-spectrum empiric therapy may inform antibiotic stewardship interventions and facilitate benchmarking.

Methods: We performed a retrospective cohort study of adults discharged in 2019 from 928 hospitals in the Premier Healthcare Database. "Empiric" gram-negative antibiotics were defined by administration before day 3 of hospitalization. Multivariable logistic regression models with random effects by hospital were used to evaluate associations between patient and hospital characteristics and empiric receipt of broad-spectrum, compared to narrow-spectrum, gram-negative antibiotics.

Results: Of 8 017 740 hospitalized adults, 2 928 657 (37%) received empiric gram-negative antibiotics. Among 1 781 306 who received broad-spectrum therapy, 30% did not have a common infectious syndrome present on admission (pneumonia, urinary tract infection, sepsis, or bacteremia), surgery, or an intensive care unit stay in the empiric window. Holding other factors constant, males were 22% more likely (adjusted odds ratio [aOR], 1.22 [95% confidence interval, 1.22-1.23]), and all non-White racial groups 6%-13% less likely (aOR range, 0.87-0.94), to receive broad-spectrum therapy. There were significant prescribing differences by region, with the highest adjusted odds of broad-spectrum therapy in the US West South Central division. Even after model adjustment, there remained substantial interhospital variability: Among patients receiving empiric therapy, the probability of receiving broad-spectrum antibiotics varied as much as 34+ percentage points due solely to the admitting hospital (95% interval of probabilities: 43%-77%).

Conclusions: Empiric gram-negative antibiotic use is highly variable across US regions, and there is high, unexplained interhospital variability. Sex and racial disparities in the receipt of broad-spectrum therapy warrant further investigation.

Keywords: antibiotic stewardship; empiric therapy; gram-negative antibiotics; inpatient antibiotic utilization.

Figure 1.

Gram-negative antibiotic spectrum-of-activity categories, as classified by a panel of infectious disease clinicians and antimicrobial stewardship experts. Antibiotics were classified using an iterative, multistage process by infectious disease specialists. Because this study focused on gram-negative empiric therapy, the steps and goals were to (1) identify antibiotics with gram-negative activity; (2) refine this list to only those antibiotics that are primarily used for gram-negative, rather than gram-positive, infections (see asterisk below); and (3) classify those antibiotics that remained into broad- vs narrow-spectrum categories based upon activity against gram-negative organisms. Following an antibiotic literature review (by K. E. G., A. D. H., and J. D. B.), steps 1–3 were performed independently by A. D. H. and J. D. B., followed by discussion to adjudicate discrepancies. The proposed classifications were subsequently provided to P. D. T. for review, comment, and proposal of changes. In round 2, A. D. H. and J. D. B. reconvened to discuss P. D. T.’s feedback and to implement agreed-upon modifications, with further outreach to E. L. H. and P. D. T. where necessary to reach consensus. Narrowest-spectrum and narrower-spectrum antibiotics were collectively designated as “narrow-spectrum” gram-negative antibiotics; extended-spectrum and extremely broad-spectrum antibiotics were classified as “broad-spectrum” gram-negative antibiotics. *The following antibiotics with gram-negative activity were excluded from classification, because they are primarily used for gram-positive infections: first-generation cephalosporins, oxacillin, cloxacillin, didoxacillin, trimethoprim-sulfamethoxazole, doxycycline, nafcillin, and penicillins.

Figure 2.

Cohort of hospitalized patients who received empiric gram-negative antibiotics across 928 United States (US) hospitals (2019).

Figure 3.

Characteristics of patients who received broad-spectrum empiric gram-negative therapy (n = 1 781 306) or with invasive infections who received any type of empiric gram-negative therapy (n = 717 217). Infectious syndromes were pneumonia, urinary tract infection, sepsis, and bacteremia. Invasive infection was defined as sepsis or bacteremia. *In the empiric window. Abbreviations: BS, broad-spectrum; ICU, intensive care unit; MV, mechanical ventilation; NS, narrow-spectrum; POA, present on admission.

Figure 4.

A, Distribution of the estimated hospital-specific probabilities of receiving broad-spectrum therapy among patients receiving empiric gram-negative antibiotics, stratified by United States (US) census division, controlling for patient and hospital characteristics. The boxes represent the 25th–75th percentile interquartile range (IQR); the horizontal line in each box reflects the median; the diamond reflects the mean; the whiskers reflect ± 1.5 IQR; and the circles reflect outliers. The estimated probabilities are derived from each hospital’s random intercept from the mixed-effects regression model (ie, empirical Bayes estimates) and are interpretable as the probabilities in each hospital at the reference value of all variables (fixed-effects) in the regression model. As such, the interhospital variability estimate is not influenced by each hospital’s sample size, which could inflate interhospital variability due to chance. There was high interhospital variability within each geographic division, with the widest IQR in the Mountain (IQR, 44%–64%; hospital n = 41), New England (IQR, 40%–60%; hospital n = 15), and Pacific (IQR, 42%–61%; hospital n = 117) divisions. The narrowest interquartile ranges were in the West North Central (IQR, 54%–66%; hospital n = 72) and the Middle Atlantic (IQR, 56%–67%; hospital n = 113) divisions. B, “Baseline” (ie, starting) estimated probability of receiving broad-spectrum therapy for each hospital in the cohort, among patients who received empiric gram-negative antibiotics in 925 US hospitals. The probability is plotted as a red dot that, due to the large sample size, displays as a continuous red line. These starting probabilities for each hospital are calculated from the hospital’s random-effects as well as its fixed-effects coefficient estimates for the 7 hospital characteristics included in the multivariable model (eg, teaching status, case-mix index) and are interpretable as the probabilities in that hospital for a patient with reference category values for each patient characteristic included in the multivariable model (see Table 3). At each hospital, a patient’s probability of receiving broad-spectrum therapy could move up or down from this starting point based upon their specific patient characteristics, as governed by the effect estimates for these characteristics in the multivariable model. We have divided the graph into 4 shaded boxes, each representing a quartile of hospitals, progressing from lowest (left) to highest (right) based upon the hospital’s starting probability. In this cohort, a hospital’s starting probability of a patient receiving broad-spectrum therapy was as low as 14% or as high as 97%.