1. Introduction
Deaths due to resistance are projected to climb to an estimated 10 million people annually by 2050 [
1]. The importance of managing antimicrobial resistance has never been more critical. At a local level, annual reports of cumulative pathogen incidence and antibiotic susceptibility data, known as antibiograms, help guide the selection of empiric antibiotic therapies [
2,
3]. Antibiograms are utilized to surveillance resistance rates and potential patterns to highlight trends over time within an institution. There are many methods for compiling and presenting antimicrobial susceptibility data, a methodology that can be complex and cumbersome. It produces results that can be difficult to interpret. The Clinical and Laboratory Standards Institute (CLSI) has published multiple guidelines with general recommendations aimed at guiding antimicrobial stewardship programs (ASP) in the development of antibiograms that are both accurate and clinically useful [
2,
3]. However, the guidelines also emphasize that some institutions, units, and patient populations may require tailored data stratification beyond the standard recommendations to obtain the most reliable results. Several studies have evaluated and found the utility of data stratification according to the unit, specimen type, and even method of infection acquisition, further termed enhanced antibiograms (EA) [
4,
5,
6,
7,
8,
9].
Rule-based technology (RBT) can ease antibiogram creation by automated inclusion and exclusion of cultures and susceptibilities by triggering specific rules and criteria [
9,
10,
11]. However, optimization of RBT requires understanding the rules and re-investment [
12]. In one of his many addresses pertaining to quality improvement, Dr. Berwick states what he calls the central law of improvement, “Every system is perfectly designed to achieve exactly the results that it achieves.” [
13]. For example, RBT has been successfully utilized within health informatics to identify and stratify adverse drug events (ADEs) [
14]. A study by Jha et al. identified ways to improve positive capture by comparing automated ADE collection to those collected manually through chart review and voluntary reporting [
15]. Thanks to the Centers for Medicare & Medicaid Services Electronic Health Records Incentive Program, health systems are all too familiar with the “Out-of-the-Box” misnomer often tagged to software [
16,
17]. Similarly, the performance of RBT-generated antibiograms should be subject to performance auditing. Antibiograms should be optimized to best suit their patient populations, which may require revisiting the rules involved in creation.
EA broadens clinical utility and has been successfully deployed in various scenarios [
18,
19,
20,
21,
22,
23]. While the most common utility of the annualized susceptibility report is to guide empiric antimicrobial prescribing decisions, there are further strata that may improve application and prescribing precision. Based on the guidance of “First culture, per patient”, RBT has limitations. Based on such rules, a patient acquiring a multi-drug resistant
P. aeruginosa pneumonia on day 30 of admission would likely be included in the RBT-generated antibiogram if it is their first culture of admission. Due to many factors, acute care units treating patients requiring extended lengths of stay are often challenged with exceptionally resistant organisms [
24,
25,
26,
27]. If included, susceptibilities of cultures taken later in admission will heavily skew the antibiograms for these units. These skewed results may overestimate resistance rates for new admissions and lead to overprescribing broad-spectrum empiric antibiotics. Equally important, antibiograms should not be used to monitor the emergence of resistance during antimicrobial treatment, guide treatment later in admission, or after recent antimicrobial exposure [
2].
At the time of the study design, empiric antimicrobial therapy was source specific; however, most patients were empirically started on vancomycin and cefepime, despite the local, unit-specific antibiogram recommending empiric coverage with vancomycin and double coverage for potential Gram-negative pathogens. The anecdotal practice did not align with the automated RBT antibiogram (
Table 1). A quick pull of the source data for the
P. aeruginosa isolates utilized to generate the RBT antibiogram found several isolates drawn weeks after admission. It was hypothesized manual review of the data and creation of EA may have potentially significant implications for prescribing recommendations and future ASP practices. While the current antibiogram of study is already an EA (e.g., burn unit-specific), thought was given to common bedside prescribing practices when considering which additional rules to consider in the further stratified EA. The primary objective of this study was to compare the pathogens and susceptibilities of the current automated RBT antibiogram with EA manually collected through chart review with additional rules accounting for days since admission, risk factors for hospital-acquired infections, and initial courses of antibiotic therapy.
3. Discussion
This study has many implications for current and future antimicrobial susceptibility reporting, interpretation, and antimicrobial prescribing. Antibiograms aim to provide regularly updated data to guide antimicrobial selection for empiric treatment of initial infections. In this study, an EA specific to the institution’s burn center was further enhanced by additional manually-applied rules. Each rule selected was based on typical decision trees utilized in bedside differentiation in empiric antibiotic determination. Each rule provided a unique depiction of sensitivity alterations, especially compared to the current EA. While patient outcomes were not considered, this report is the first to analyse additional diverse strata applied to burn-specific EA. In a population at high risk for multi-drug resistant pathogens, any means to minimize exposure to unnecessarily broad spectra of antimicrobials has large downstream implications for the patient and the unit.
The “call” is clear for further research exploring the true impact of clinical decision support systems and antibiograms on ASP [
2,
28,
29,
30,
31,
32]. In their summary, Hindler and Stelling outlined the necessity for future researchers to look more critically at the produced antibiograms to optimize performance best and improve prudent prescribing [
2]. Treatment outcome was not directly measured in this study. However, the analysis provides clear evidence of how prescribing recommendations of empiric antibiotics are altered with additional EA considerations. Consider the results from the perspective of a case example, where a patient may be admitted with acute severe burns to 50% of their body. Using national averages, they will require more than ten acute surgical procedures and be in the hospital for around 70 days [
33,
34]. Sepsis remains the most common reason for mortality in patients with burn injuries surviving the initial 48 h, and wound infection is the most common source [
24,
35]. At some point in the stay, the patient will likely require systemic antibiotics, likely multiple courses. If, on hospital day six, cefepime, amikacin, and vancomycin are prescribed empirically for suspected sepsis, the subsequent infection (or perhaps during treatment) will likely be highly resistant. Traditionally, international burn-specific data is strongly correlated and suggests
P. aerugionsa,
A. baumannii,
S. maltophilia, or
carbapenem-resistant Enterobacteriaceae will soon follow [
36,
37,
38].
Table 6 and
Figure 4 demonstrate that multi-drug resistance is highly prevalent, even after a single course (or seven days of exposure) of antibiotics.
In some cases, empiric antibiotics are initiated without the attainment of cultures. Despite recommendations, some patients receive antibiotics without cultures to guide definitive treatment. This is obvious from the study results, as the current antibiogram (
Table 1) reflecting “first culture, per patient” would have more closely resembled the “initial” EA in
Table 5.
Recall during the analysis, empiric antimicrobial prescribing did not abide by the RBT EA and instead followed anecdotal evidence. There is potential for significant error with recall bias, and relying solely on anecdotes should be discouraged. The study hypothesis was created with this in mind. Notably, anecdotal evidence proved more reliable than the RBT utilized for past iterations of the EA antibiogram. CLSI recommends avoiding empiric monotherapy prescriptions for serious infection when susceptibility patterns indicate the chosen agent has less than 90% susceptibility for the likely pathogen(s) [
2,
3]. In the same recommendations, there is latitude given for susceptibilities down to 80% for certain infections and populations. The additional rules applied to the EA supported using a single antipseudomonal beta-lactam antibiotic plus an agent with activity against methicillin-resistant
S. aureus (MRSA) instead of two Gram-negative antibiotics.
Unfortunately, early empiric recommendations still indicate an antipseudomonal agent is necessary. Globally,
P. aeruginosa,
A. baumannii, and
Enterobacteriaceae remain common pathogens following burn injury [
36,
37,
38,
39]. Additionally, the prevalence of community-onset MRSA is growing (unpublished institutional data), which parallels statewide and national reports [
40]. While a single antipseudomonal beta-lactam antibiotic typically covers methicillin-sensitive
S. aureus, it is a poor choice for MRSA. Even in the best EA model scenario produced during the study, using only a single beta-lactam without an MRSA active agent would have resulted in 44% of patients being inadequately covered. Fortunately, a previously unpublished internal analysis noted a few isolates with a minimum inhibitory concentration in excess of 1 μg/mL, which improves the likelihood of treatment response.
Reflecting Dr. Berwick’s remarks, continuous investment in process improvement (PI) is imperative. Demonstrating an adequate PI program for burn center verification through the American Burn Association is necessary. Infection prevention and stewardship practices should be a cornerstone of PI, as the iatrogenic acquisition of multidrug-resistant bacteria carries with it proud morbidity and mortality. An easy method of preventing the creep of early multidrug-resistant pathogen prevalence is reducing iatrogenic spread. Attention must stretch beyond contact isolation and proper donning of personal protective equipment. An often-overlooked aspect of infection prevention is the various components of environmental cleanliness, especially for units caring for patients with burn injuries. Microbes are called such for a reason. Any small break in the infection prevention chain affords a massive opportunity for opportunistic pathogenesis. Units sufficiently monitoring culture data will see fluctuations and timing of “their unit-specific pathogens”, indicating when reinvestment in infection prevention audits may be indicated.
Knowing RBT or laboratory-based susceptibility reports may not present the clinically-relevant data is certainly not a novel concept [
2,
41,
42]. It is essential to understand not all bacteria are pathogens. For example, most infections in burn centers involve the wound. However, it is critical to understand a common misnomer, wounds do not have to be sterile to heal. Evidence is growing, especially as we can detect biobank species, and some bacteria promote wound healing [
43,
44]. Over-targeting bacteria or exposing patients to a broad spectrum of antimicrobials could be more detrimental than previous depictions. Therefore, it is imperative not to include surveillance data in antibiograms.
A significant limitation of this study is the reproducibility. While the hypothesis was supported, the number of man-hours required for the chart review and data collection could provide a sufficient workload to support an entire full-time equivalent (e.g., FTE), especially considering the other EA needed for the multiple hospital units. Each hospital unit typically houses patients from a single (or a small number) subspecialty and presents a unique environment/microbiota. Recall the demographic and clinical data for each admitted patient was reviewed in hopes of creating the additional EA and only including clinically relevant pathogens (e.g., reducing chances of reporting surveillance cultures). The evidence presented supports the need to invest in software development and integration. The point of RBT is to improve efficiency and accuracy substantially. We are not there yet. Due to wide confidence intervals and potential misrepresentation of the larger population (e.g., all patients admitted to the unit), samples (e.g., pathogens) should be either excluded or pooled with additional cohorts or in a multiyear fashion when analyzed at a drug-pathogen level. In this analysis, power was dramatically improved over individual drug-pathogen analysis by including (1) two years of laboratory and clinical data and (2) pooling all the pathogens. Antibiograms displaying sensitivities per individual pathogen have advantages when the source is known, and the likely pathogen can be narrowed. However, this is disadvantageous when the source is not known, and the pooled analysis was a better method to answer the hypothesis questioned in the study.
4. Materials and Methods
4.1. Study Design and Patient Population
The study was approved by both the University of Tennessee Health Science Center and Regional One Health Research Institute Institutional Review Boards (20-07615-XP). This dual IRB-approved study was an observational case series of patients admitted to a single verified burn center between 1 January 2018, and 31 December 2019. Patients were excluded for any of the following: (1) no positive bacterial cultures obtained, (2) less than 18 years of age, (3) incarcerated, (4) pregnant, (5) cultures collected after 30 days of admission, (6) culture results below quantitative thresholds, or (7) isolates not reported on the automated antibiogram (e.g., no comparison could be made). Patients were screened initially by reviewing burn center admission logs during the study period, and exclusion criteria were applied to generate a final sample of patients and cultures.
Computer-generated, rule-based antibiograms were compared to the manually-collected antibiograms over two years. The study period was chosen to ensure an adequate sample after applying inclusion and exclusion criteria and the ability to compare the last two annual antibiograms [
2,
45]. A priori estimates accounted for an estimated 750 admissions, with half being cultured for potential infection and a goal of at least 30 isolates for the most commonly reported pathogens (
S. aureus,
E. faecalis,
Enterobacter spp., and
P. aeruginosa).
The hypotheses driving this study attempted to capture bedside considerations when initiating new courses of antibiotics. The primary hypothesis of this study was including days since admission, as a rule, will significantly alter the antibiogram and associated sensitivities. The aim was to compare each pathogen from the autogenerated antibiogram to a manually collected version with an additional rule applied within seven days of admission. A second hypothesis was excluding patients with risk factors for hospital-acquired infections will significantly alter the ideal choice for empiric antimicrobial therapy. The second aim compared the automated version to a manually collected antibiogram with two additional rules applied: (1) within seven days of admission and (2) patients without risk factors for hospital-acquired infections. The third hypothesis was susceptibilities significantly decrease after a single course of antimicrobials. To test this hypothesis, susceptibilities were compared between patients with a prior history of antibiotic exposure.
4.2. Data Collection
Data were manually collected from the electronic medical record during individual chart reviews. Demographic data included: age, sex, race, comorbidities, date of arrival and risk factors for hospital-acquired infections (e.g., intravenous access, history of chemotherapy, positive urine drug screen or reported social history, resident in a nursing home or long-term acute care hospital, or admission to the hospital in the last 90 days). Burn injury characteristics included: etiology, presence of inhalation injury, % total body surface area burned, and % partial thickness and full thickness injury. Treatment data during the first 30 days of admission included: dressings utilized, topical and systemic antimicrobial agents and dates utilized, systemic antimicrobial indication, and systemic steroid use and dates. Based on the aims, outcome data included pathogens and sensitivities.
Every attempt was made to include only those considered pathogens (e.g., limit inclusion of surveillance cultures). Positive bacterial cultures were defined as pathogens meeting the positivity threshold for the source of culture or deemed to require therapeutic courses of antibiotics. Positivity thresholds were dependent on source: wounds (105 or semiquantitative tissue or exudate results treated with systemic therapy), bronchoalveolar lavage (105), blood (any growth that resulted in treatment with systemic therapy), urine (105), bone (any growth resulting in treatment with systemic therapy), other (typically semiquantitative results of drainage). Susceptible pathogens were defined as strains whose minimum inhibitory concentrations (MIC) were interpreted to be susceptible to a given antibiotic. Non-susceptible pathogens were defined as strains whose MICs were interpreted to be resistant or intermediate to a given antibiotic. During this study period, the institution’s microbiology laboratory utilized the bioMerieux Vitek 2 (Durham, NC, USA) automated system for identifying bacteria and bacterial susceptibility testing, along with secondary panels (Kirby Bauer and E test) for multi-resistant organisms. The Vitek 2 bacterial identification system is based on established biochemical methods and substrates measuring carbon source utilization, enzymatic activities, and resistance. Rules for antibiotic reporting and interpretation for MIC values from the Vitek 2 were based on FDA-cleared interpretations built within the automated Vitek 2 system. Quality control for Vitek identification and MICs followed the package insert for the Vitek products. Kirby Bauer and E test interpretation, reporting and quality control followed CLSI recommendations.
4.3. Sample and Statistical Analysis
4.3.1. Sample Size Determination
The primary objective was to compare resultant antibiograms between an automated, rule-based software and a manually derived antibiogram to include standard rules plus days since admission. In addition, given the rarity of some pathogens, the two years of admissions were reviewed to ensure at least 30 of the most commonly reported pathogens were included in the study sample.
4.3.2. Statistical Analysis
Patient demographic data and injury characteristics were reported using descriptive statistics. Nominal data were reported with n (%). The Shapiro-Wilk test was utilized to test for the normality of continuous data. Non-parametric data were reported as median (interquartile range). Normally distributed data were reported as mean ± standard deviation. Differences in sensitivities for pathogens and antibiotics were compared using Fisher’s exact test. Statistical analysis was performed in SigmaPlot version 11.2, Palo Alto, CA, USA.