Antibiotic Use and Presumptive Pathogens in the Veterans Affairs Healthcare System

Abstract

Background

Empirical antibiotic use is common in the hospital. Here, we characterize patterns of antibiotic use, infectious diagnoses, and microbiological laboratory results among hospitalized patients and aim to quantify the proportion of antibiotic use that is potentially attributable to specific bacterial pathogens.

Methods

We conducted an observational study using electronic health records from acute care facilities in the US Veterans Affairs Healthcare System. From October 2017 to September 2018, 482 381 hospitalizations for 332 657 unique patients that met all criteria were included. At least 1 antibiotic was administered at 202 037 (41.9%) of included hospital stays. We measured frequency of antibiotic use, microbiological specimen collection, and bacterial isolation by diagnosis category and antibiotic group. A tiered system based on specimen collection sites and diagnoses was used to attribute antibiotic use to presumptive causative organisms.

Results

Specimens were collected at 130 012 (64.4%) hospitalizations with any antibiotic use, and at least 1 bacterial organism was isolated at 35.1% of these stays. Frequency of bacterial isolation varied widely by diagnosis category and antibiotic group. Under increasingly lenient criteria, 10.2%–31.4% of 974 733 antibiotic days of therapy could be linked to a potential bacterial pathogen.

Conclusions

Overall, the vast majority of antibiotic use could be linked to either an infectious diagnosis or microbiological specimen. Nearly one-half of antibiotic use occurred when there was a specimen collected but no bacterial organism identified, underscoring the need for rapid and improved diagnostics to optimize antibiotic use.

More than two-thirds of antibiotic use among hospitalized patients may occur at stays without a bacterial isolate. Metrics linked to diagnosis and microbiology may be useful for antibiotic stewardship, and rapid and improved diagnostics are needed to optimize antibiotic use.

INTRODUCTION

Antibiotic-resistant infections continue to pose a serious threat to hospitalized patients [1, 2]. Resistant pathogens may readily emerge and transmit in the hospital because of the vulnerable patient population, shared environments and personnel, and transfers between healthcare facilities [2]. Additionally, high levels of antibiotic use, with approximately one-half of patients receiving at least 1 antibiotic during their stay [3], impose selective pressures for antibiotic resistance. Studies have estimated that 30%–40% of antibiotic use among hospitalized patients may be inappropriate [4–6].

A key challenge underlying inappropriate antibiotic use is identification of the causative pathogen of infection, a nontrivial task resulting from incomplete specimen collection, difficulties in obtaining high-quality samples [7–9], and the need to distinguish between colonizing and causative organisms, especially for infections at nonsterile sites [9]. At an individual level, identification of the causative organism can inform optimal treatment [10]. For instance, Pseudomonas aeruginosa should generally be treated with more intensive regimens because of intrinsic resistance mechanisms and the propensity to acquire others [11, 12]. Isolation of the causative pathogen also facilitates investigation of antibiotic susceptibilities and may guide a clinician to de-escalate (or escalate) a patient to narrower (or broader) spectrum therapy [13–15]. Prior studies found that microbiological results led to modifications in antibiotic treatment regimens for 30%–40% of cases who had a specimen taken [16, 17], though this number may be expected to vary based on local etiologies, antibiotic resistance patterns, and clinical practices. At the population level, comprehensive data on causative pathogens may be used to describe spatial and time-varying trends in etiology and to aid in the design of locally tailored antibiotic stewardship programs.

Here, we use electronic health records to comprehensively describe patterns of antibiotic use, infectious diagnoses, and microbiological specimens among hospitalized patients in the Veterans Affairs (VA) Healthcare System. To the extent possible, we then characterize the proportion of antibiotic use that can be attributed to presumptive bacterial pathogens and how these proportions vary by antibiotic class and category. Antibiotics were categorized based on the framework used for the Standardized Antimicrobial Administration Ratio (SAAR) reported by the National Healthcare Safety Network (NHSN) [18].

METHODS

The VA Healthcare System is the largest integrated healthcare system in the United States, serving approximately 9 million enrollees [19]. This analysis used electronic health records from 136 acute care facilities in the VA Healthcare System from October 2017 through September 2018. Hospitalizations were included if both admission and discharge occurred within the study period; stays without time in acute care wards (medical/surgical, intensive care, or medical observation) were excluded. To limit complex cases with multiple infection episodes, we excluded hospital stays that lasted longer than 10 days.

Antibiotic use in the VA is recorded using barcode medication administration software. Antibiotic use was measured in days-of-therapy (DOT), with any number of administrations of a specific antibiotic on a given calendar date counted as a single day of therapy. NHSN is a system coordinated by the Centers for Disease Control and Prevention to track healthcare-associated infections [20], and the SAAR metric was developed by NHSN to provide a standardized, risk-adjusted benchmark for antibiotic use [21]. In accordance with the SAAR protocol, we included antibiotics administered via intravenous, intramuscular, digestive, or respiratory routes; records with a different or missing route of administration were excluded (< 1% of all antibiotic administrations over the study period). Antibiotics were organized into the Adult SAAR Antimicrobial Agent Categories outlined by the NHSN (Table S1) [22], which were designed to capture key groupings for antibiotic stewardship programs. We focused on the 5 mutually exclusive SAAR antibiotic categories: broad spectrum antibacterial agents predominantly used for hospital-onset infections, broad spectrum antibacterial agents predominantly used for community-acquired infections, antibacterial agents predominantly used for resistant Gram-positive infections, narrow spectrum beta-lactam agents, and antibacterial agents predominantly used for extensively antibiotic resistant bacteria. We also included complementary agents, defined as antibiotics included in the “all antimicrobials” SAAR metric but not assigned to any of the categories mentioned previously.

Diagnoses were based on the International Classification of Diseases, 10th Revision, Clinical Modification scheme. All codes assigned during the hospital stay or up to 2 days before admission were included because diagnoses from an outpatient or emergency room visit within 2 days of admission were likely related to the subsequent hospitalization. International Classification of Diseases, 10th Revision, Clinical Modification scheme, diagnosis codes were grouped using a VA-developed classification scheme in conjunction with previously published work [23]. Codes were considered “infectious” if they were associated with a potentially bacterial infection and were further stratified into common indications, including urinary tract infections (UTI), lower respiratory tract infections (LRI), skin and soft tissue infections (SSTI), bacteremia, and sepsis (Table S2).

We included all specimens collected during the hospital stay or up to 2 calendar days before admission. Specimens with no growth, unspeciated growth, or questionable significance were considered negative results. For Clostridium difficile, positive specimens included identification of toxin and rapid test results (polymerase chain reaction or enzyme-linked immunosorbent assay) in addition to culture growth. The collection site of each specimen was categorized as “sterile,” “nonsterile,” or “other” (Table S3). Specimens classified as “other” were either donor specimens or likely collected for methicillin-resistant Staphylococcus aureus surveillance and were excluded from this analysis. For coagulase-negative staphylococci (CoNS), which is often a contaminant but increasingly found to be a pathogen [24, 25], we required 2 positive specimens from the same body site.

In general, organisms identified from sterile sites, excluding probable contaminants, are considered significant and clinically relevant, whereas organisms from nonsterile sites may represent colonizing flora depending on the species and specimen collection site. We categorized bacterial isolates from most to least likely to be a causative agent of antibiotic treatment according to the following tiers:

1. Bacterial organism identified from any sterile site.

2. Bacterial organism identified from a nonsterile site corresponding to a common infectious diagnosis (as defined in Table S4) from the same stay.

3. Bacterial organism identified from any other nonsterile site.

All antibiotic use at each stay was attributed to the isolated organism at the most likely tier. If multiple bacterial organisms of the same tier were identified at a single stay, antibiotic use was attributed to all organisms. For hospital stays at which antibiotics were administered but no bacterial organism was identified, we assessed whether any specimen was collected, any infectious condition was diagnosed, or any procedure warranting antibiotic prophylaxis was performed. Procedures were identified based on Current Procedural Terminology codes for which first- or second- generation cephalosporin prophylaxis is indicated in Centers for Medicare & Medicaid Services’ Quality Measures (Table S2) [26].

All analysis was conducted in R Version 3.5.3 (2019-03-11) “Great Truth” [27].

RESULTS

Over the study period, 482 381 hospital stays for 332 657 unique patients met the inclusion criteria (Figure 1). Patients were 93.7% male with a median age of 69 (interquartile ratio: 60–74), slightly more male and older than the overall enrollee population [19]. Across all included stays, 202 037 (41.9%) included any antibiotic use, 974 733 antibiotic DOT were administered, 544 477 specimens were collected (45 431 or 8.3% in the 2 days before admission), and 73 495 specimens resulted in isolation of at least 1 bacterial organism.

Figure 1.

Flowchart of study population.

Nearly 60% of hospitalizations (59.2%) resulted in at least 1 infectious diagnosis; about one-quarter (27.0%) included diagnoses for UTI, LRI, SSTI, bacteremia, and/or sepsis (Table 1). Specimens were collected at 56.4% of stays with any infectious diagnosis, and at least 1 bacterial organism was identified at 31.7% of stays with any specimen. These quantities were 17.3% and 9.1%, respectively, among stays without an infectious diagnosis. In most cases, both proportions increased substantially when the diagnosed condition was UTI, LRI, SSTI, bacteremia, sepsis, or some combination of these conditions. However, bacterial organisms were less frequently isolated for patients with LRI only (16.0%) and sepsis only (29.5%). In general, the likelihood of specimen collection and mean DOT per stay increased with an additional diagnosis of sepsis or multiple infectious conditions. Specimen collection and identification of a bacterial organism occurred more often when restricting to hospital stays with any antibiotic use (Table S5).

Table 1.

Overview of Hospital Stays, Antibiotic Use, and Specimen Collection by Diagnosis Category

Diagnosis	Hospital Stays, n	Mean Length of Stay, d	Any Antibiotic Administered, %	Mean DOT per Stay	Any Specimen Collected, %	Bacteria Identified From Specimen, %
UTI only	21 720	4.5	81.2	3.3	84.1	55.6
LRI only	45 958	4.6	68.9	3.8	71.4	16.0
SSTI only	21 973	4.6	88.4	5.5	77.4	34.5
Bacteremia only	2097	5.4	84.8	5.4	84.9	53.6
Sepsis only	5451	5.0	84.8	5.7	91.0	29.5
Sepsis + (UTI, LRI, SSTI, or bacteremia)	18 250	5.2	95.9	7.1	95.5	45.3
Other combinations of the above	14 621	5.8	94.1	7.5	93.3	61.6
Other infectious diagnoses	155 367	4.1	35.2	1.4	33.7	19.4
Any infectious diagnosis	285 437	4.4	56.4	3.1	55.4	31.7
No infectious diagnosis	196 944	3.4	20.8	0.5	17.3	9.1
Total	482 381	4.0	41.9	2.0	39.9	27.7

Columns were calculated across all hospital stays (n) with the exception of the final column, which was calculated as a percentage of hospital stays with any specimen collection. Diagnosis categories ending in “only” include hospital stays with only 1 out of the 5 infectious conditions explicitly listed here (UTI, LRI, SSTI, bacteremia, sepsis).

Abbreviations: DOT, days of therapy; LRI, lower respiratory tract infection; SSTI, skin and soft tissue infection; UTI, urinary tract infection.

The majority of antibiotic DOT (75.1%) occurred during stays with at least 1 specimen collected from any site; this percentage increased to 90.0% when considering stays with at least 1 specimen or any infectious diagnosis code (Figure 2B). Diagnoses varied predictably across SAAR categories; for example, broad-spectrum agents typically used to treat community-acquired infections (broad-spectrum antibacterial agents) were often administered at hospitalizations with UTI diagnoses, whereas SSTIs were more common among patients receiving antibiotics predominantly used to treat resistant Gram-positives (Figure 2A). For most antibiotic groups, the most common scenario was an infectious diagnosis and specimen collection without identification of a bacterial organism (43.7% overall, Figure 2B), reflecting the trends observed in Table 1. Overall, about one-third of DOT (31.4%) occurred during hospital stays with at least 1 bacterial isolate, ranging from 19.8% for narrow-spectrum beta-lactam agents (NSBL) to 48.1% for agents predominantly used for extensively antibiotic-resistant bacteria (XDR) (Figure 2B). The vast majority of bacteria isolated from sterile sites were found in blood samples (Figure S1A), whereas the most common nonsterile sites resulting in positive specimens were urine for UTI patients, sputum for LRI patients, and wound and foot sites for SSTI patients (Figure S1B). Not all antibiotic use could be categorized into one of the specified tiers. For example, a substantial fraction of DOT for NSBL (19.3%) and agents not included in SAAR (30.6%) occurred at hospital stays lacking any specimen, infectious diagnosis, or included procedure code (Figure 2B), potentially indicating misclassification, empirical use, or use for other types of prophylaxis or noninfectious conditions. Antibiotic use patterns tended to be similar across antibiotic classes within the same SAAR category, with several exceptions: carbapenems were linked to higher levels of bacterial isolation compared with drugs in the same category and cephalosporins were linked to more potential prophylaxis for procedures compared to other NSBL (Figure S2, Figure S3).

Figure 2.

Percentage of antibiotic days-of-therapy (DOT) by SAAR antibiotic category and (A) diagnosis group and (B) tier. The number above each bar is the total DOT for that antibiotic category in thousands. The rightmost bar represents DOT across all included antibiotics. SAAR antibiotic categories include: BSCA, broad-spectrum antibacterial agents predominantly used for community-acquired infections; BSHO, broad-spectrum antibacterial agents predominantly used for hospital-onset infections; complementary, agents included in “all antimicrobials” SAAR but not listed in any particular category; GramPos, antibacterial agents predominantly used for resistant Gram-positive infections; NSBL, narrow-spectrum beta-lactam agents; SAAR, Standardized Antimicrobial Administration Ratio; XDR, antibacterial agents predominantly used for extensively antibiotic resistant bacteria. See Table S1 for details.

Only a small fraction of DOT could be potentially attributed to any particular organism from either a sterile or nonsterile site (Figure 3). Under the most stringent criterion (isolation from a sterile site), 2.4% (range across antibiotic groups: 0.5%–5.9%), 1.9% (range: 0.6%–3.4%), and 1.0% (range: 0.5%–2.6%) of use could be presumptively attributed to S aureus, Escherichia coli, and Enterococcus spp., respectively. These values increased to 6.0%, 5.2%, and 3.1% over all antibiotics when also considering organisms isolated from a nonsterile site corresponding to a concurrent infectious diagnosis. There were distinct differences in isolated organisms between SAAR categories. For example, use of GramPos agents was most often tied to isolation of S aureus. Some evidence of species-specific treatment can also be observed within SAAR categories; for instance, carbapenems were used more often at hospital stays with isolation of E coli compared to other agents in the same category (Figure S4). For all antibiotic categories, a substantial proportion of use was potentially attributable to “other” organisms that were not in the 5 organism groups shown in Figure 3. For XDR agents, Acinetobacter species were the most commonly isolated “other” organisms from nonsterile sites; for most other SAAR categories, the most commonly isolated “other” organisms from both sterile and nonsterile sites were Streptococcus spp.

Figure 3.

Percentage of days-of-therapy (DOT) of each antibiotic group by isolated bacterial organism and tier. Organisms were abbreviated as follows: Entc, Enterococcus spp.; EC, Escherichia coli; Kleb, Klebsiella spp.; other, other bacterial organisms; PA, Pseudomonas aeruginosa; SA, Staphylococcus aureus.

DISCUSSION

We have conducted a comprehensive assessment of potential reasons for inpatient antibiotic use in the VA Healthcare System through analysis of diagnosis codes and microbiological test results. Our analysis, which used routinely collected, structured data from electronic health records, was designed to be readily replicable in other hospitals and health systems. We found that the majority of antibiotic use (75.1%, Figure 2B) occurred at hospital stays with at least 1 specimen collected, presumably signaling physician intent to identify the causative pathogen. However, identification of a bacterial organism varied widely across diagnoses, occurring less than one-half of the time in specimens collected from patients with LRI, SSTI, and sepsis. Bacterial isolation was more common among patients with conditions requiring positive cultures (55.6% for UTI and 53.6% for bacteremia, Table 1), but still lower than expected, potentially indicating low specificity of diagnostic coding. Overall, one-third (31.4%, Figure 2B) of all antibiotic DOT could be linked with any bacterial isolate, and nearly one-half (43.7%, Figure 2B) occurred when a specimen was collected but no bacteria were isolated, a pattern that was consistent across SAAR drug categories, even for drugs in the XDR group. Nontargeted use of antibiotics increases the risk of imposing selection pressure for resistance with minimal benefit to the patient.

There are many reasons why antibiotics may be prescribed despite negative microbiological tests. Negative cultures do not rule out bacterial infection. For example, blood cultures have low sensitivity for most types of infections. Contributory factors to negative cultures, even when a bacterial infection is present, include specimen collection more than 24 hours after initial antibiotic administration (15.9% of specimens in our data) or other challenges in specimen processing. We also required confirmation with 2 tests to conclude a positive identification of CoNS; approximately 10% of bacteremia hospitalizations included a single isolation of CoNS. Conversely, a recovered bacterial isolate cannot necessarily be construed as the intended target of antibiotic therapy nor is it unambiguous evidence of a true bacterial infection. Our tiered approach was designed to create a gradient of diminishing uncertainty about the causative organism for a given patient’s infection.

Comparisons to existing literature are challenging because of varying healthcare contexts, diagnostic criteria, and study populations. Prior estimates of the frequency of specimen collection among hospitalized patients who received antibiotics ranged from 49% across 22 hospitals in the Netherlands [28] to 85%–90% in other European hospitals [17, 29], compared with our estimate of 64.4% (Table S5). Prior work in the VA Healthcare System found that specimens were collected within 2 days of admission from blood or select respiratory sites in 84.5% of pneumonia hospitalizations [30] compared with our estimate of 71.4% (Table 1). Identification of a bacterial organism appeared to be less common for both SSTI (34.5% [Table 1] vs 54% [31]) and LRI (16.0% [Table 1] vs 30.6%–50% [32, 33]) in our analysis, albeit compared with several studies across different contexts. These discrepancies may reflect broader diagnostic criteria for infection or differences in specimen collection practices or etiology in our study population.

Our analysis is subject to limitations. First, we make several key assumptions to analyze this large, complex dataset. We use a simplified tiered approach to categorize isolated organisms from most to least likely to be the causative agent of treatment. Results should be cautiously interpreted, especially for isolates from nonsterile sites; in reality, the clinical team may also consult additional microbiological and laboratory data, patient history, and other variables to determine the most likely causative pathogen. For a similar reason, it is challenging to evaluate whether any antibiotic use or specimen collection practices were or were not “appropriate.” Actions may be clinically advised based on clinical signs and symptoms or other variables that were not included in this analysis. For instance, specimen collection is not warranted when antibiotics are used prophylactically or when clinical signs of infection are absent. In the latter case, microbiological results may lead to overdiagnosis and unnecessary antibiotic treatment, as observed with asymptomatic bacteriuria [34]. However, the ecological nature of our analysis allows us to highlight areas for further investigation by simultaneously conveying potential stewardship concerns and their frequency. For example, a substantial fraction of NSBL use could not be linked to a specimen, infectious diagnosis, or procedure warranting prophylaxis; more detailed follow-up work could identify potential reasons for this use. On the other hand, given the greater resistance concerns and more frequent use of broad-spectrum antibacterial agent therapies, follow-up work may be inclined to focus on the proportion of DOT in this category that were linked to an infectious diagnosis but lacked a positive specimen. We also do not account for chronological ordering of events within a stay and assume that patient data over the course of a short-duration (≤ 10 days) hospitalization is linked to the same infection episode.

Next, diagnostic coding is imperfect and may not represent clinical status for a number of reasons, including automated entry or coding for rule-out purposes. Studies investigating the accuracy of diagnosis codes for infections have reported highly variable sensitivity and specificity depending on syndrome and clinical context [35–37]. In our baseline analysis, we consider codes assigned at any point during a hospitalization, the most inclusive approach that may be prone to low specificity. This misclassification may contribute to the lower-than-expected frequency of bacterial isolation for infectious conditions, especially UTI and bacteremia. “Infectious” diagnosis codes were also defined broadly as any potentially bacterial condition. Some conditions, such as chronic obstructive pulmonary disease exacerbation, may warrant antibiotics only if confirmed with other clinical data. As sensitivity analyses, we conducted our analysis with discharge (Table S6) and principal (Table S7) diagnosis codes, which are likely more specific as they are assigned at the conclusion of a hospital stay [36]; in both cases, antibiotic use, specimen collection, and bacterial identification increased across all infectious diagnoses. As demonstrated here, diagnosis codes have the potential to streamline surveillance of antibiotic use, but context-specific validation studies are needed to quantify their accuracy.

Finally, the enrollee population of the VA Healthcare System is skewed toward older males, who tend to have poorer health status compared with the general patient population, limiting the generalizability of our findings [38].

CONCLUSION

We used data from electronic health records to describe patterns of antibiotic use, specimen collection, and bacterial isolation in the VA Healthcare System. We found that specimens were usually collected when antibiotics were administered in the hospital, but only about one-third of antibiotic use occurred at a hospitalization with a bacterial isolate. This analysis is easily generalizable to any facility recording clinical diagnoses and microbiology results, and future work could examine the degree to which the observed patterns vary across hospitals and health systems, including between the VA facilities that were assessed in aggregate in this work. Extensions could also incorporate antibiotics administered at discharge, which were not available for this analysis. Further research is needed to improve our ability to rapidly and accurately determine the causative pathogens driving treatment, and ultimately to optimize antibiotic use.

Notes

Acknowledgments. The authors thank Dr. Makoto M. Jones for his thoughtful comments on the study design and contributions to preparing the data for this analysis.

Financial support. This work was supported by the Centers for Disease Control and Prevention (grant number CK000538-01); the National Institute of Allergy and Infectious Diseases (grant number T32AI007535 to C. T.); and the IDEAS Center at the Department of Veterans Affairs, Veterans Health Administration (to M. H. S.). This content is solely the responsibility of the authors and does not necessarily reflect the official views of the Centers for Disease Control and Prevention, Department of Health and Human Services, National Institute of Allergy and Infectious Diseases, National Institutes of Health, or Department of Veterans Affairs.

Potential conflicts of interest. M. L. has received grants from Pfizer, National Institutes of Health (US), National Institute for Health Research (UK), Centers for Disease Control and Prevention (US), Open Philanthropy Project, and Wellcome Trust (not related to the topic of this manuscript) and personal fees from Merck, Bristol-Meyers Squibb, and Sanofi Pasteur. C. T., M. N., and M. H. S. report no conflicts.