Keywords
antibiogram - electronic health records - clinical decision support systems - analytics
- data visualization - pediatrics - quality improvement
Background and Significance
Background and Significance
Hospitals routinely perform antimicrobial susceptibility testing for bacterial pathogens
and then summarize the results in a table called an antibiogram. Clinicians refer
to antibiograms to guide optimal empiric antibiotic therapy and reduce inappropriate
antibiotic usage.[1] Antibiograms can also be used to track changes in antibiotic resistance over time,
perform surveillance for emergence of drug-resistant organisms, and identify areas
for intervention by antimicrobial stewardship programs.[2]
[3]
The Clinical and Laboratory Standards Institute (CLSI) developed consensus guidelines
in 2007 (revised in 2014) for the collection, storage, analysis, and presentation
of antimicrobial susceptibility data.[4]
[5] Hospitals' adherence to these guidelines has varied widely, and a 2013 survey determined
that only 39% of U.S. pediatric hospitals compiled an antibiogram on a yearly basis.[6] One possible reason for this limited adoption is that creating an antibiogram is
typically a labor-intensive, manual process. At our hospital, one physician was responsible
for the month-long task of compiling the annual institution-wide antibiogram manually.
This hospital-wide antibiogram reported aggregate data with limited granularity within
specific patient groups and patient care units.[7]
Objective
Thus, we embarked on a quality improvement project to automate this manual process
by using electronic health record (EHR) data to generate an electronic antibiogram
(“e-antibiogram”) that adheres to the CLSI guidelines and contains filters for patient
age, medical conditions, unit location, and other factors that provide the user with
access to more detailed, up-to-date data than a hospital-wide, annually produced antibiogram.
Methods
This study was performed at the Children's Hospital of Philadelphia (CHOP), an urban,
pediatric hospital with 535 beds. The CHOP Institutional Review Board deemed the study
protocol exempt from review. The hospital system has approximately 29,000 inpatient
admissions yearly and has an ambulatory care network consisting of more than 50 outpatient
primary and specialty care locations that receive approximately 1.2 million outpatient
visits each year. The hospital's infectious disease diagnostics laboratory (IDDL)
processes all microbiology cultures obtained during inpatient, emergency department,
and outpatient hospital encounters, and data are recorded in the hospital's EHR (Epic,
Verona, Wisconsin, United States) since November 2011. Prior to this study, the IDDL
compiled an annual institution-wide antibiogram that aggregated the susceptibility
data of isolates that were obtained from all specimen types across the entire hospital.
The IDDL uses the automated VITEK 2 (bioMérieux, Durham, North Carolina, United States)
system to perform organism identification and antimicrobial susceptibility testing
in conjunction with disk diffusion and gradient strip testing per CLSI guidelines.
D-testing for inducible clindamycin resistance was performed for all Staphylococcus aureus isolates. For S. aureus, oxacillin-susceptible isolates were identified as methicillin-susceptible (MSSA),
while oxacillin nonsusceptible isolates were identified as methicillin-resistant S. aureus (MRSA). We do not test routinely for extended-spectrum β-lactamases per CLSI recommendations.
Cumulative antimicrobial susceptibilities were calculated separately for MRSA and
MSSA isolates. Pathogen antimicrobial susceptibility estimates were based on clinical
factors entered by the EHR user and include the following: patient age, specimen type,
and anatomic source of culture; location of clinical encounter; and key early features
of the microbiologic result, including Gram stain, morphology, and biochemical testing
results.
All bacterial isolates originating from positive cultures collected during inpatient
and outpatient encounters during the study period were included. In the case of multiple
isolates from the same patient, the first isolate of each species from a given patient,
in each hospital unit, per year was included. Selective reporting of antibiogram data
was implemented per CLSI guidelines. Screening and surveillance cultures were excluded.
Demographic data in the microbiology records were retrieved and included patient age,
sex, and the location where the clinical specimen was obtained (inpatient, outpatient,
emergency department). Patient age was classified into one of six groups: newborn
(less than 1 month of age), young infant (1–3 months), infant (3 months–1 year), child
(1–5 years), school child (5–15 years), and adolescent (greater than 15 years). Outpatient
locations were grouped by clinical activity into primary care and subspecialty clinics.
Dashboard Development
As we have described previously, we followed a five-step process to create a secure,
EHR-linked, condition- and patient-specific visual analytics e-antibiogram that aggregates
and displays bacterial isolate data following the CLSI guidelines.[8]
[9]
[10] First, we retrieved the relevant data from the EHR relational reporting database
(Epic Clarity) and then merged thousands of conventional relational database tables
into a single dimensional model within visual analytics software (QlikView, Radnor,
Pennsylvania, United States).[11]
[Fig. 1] shows the e-antibiogram's dimensional model diagram based on a snowflake schema
(data relationships). The dimensional model presents the data in an intuitive framework
that contains information regarding the Procedure Order, Order Result, and Order Sensitivity
tables as the “fact” or main table (MAIN_FACT_TABLE). The model also contains the
original antibiotic, organism, procedure, and admission, discharge, and transfer tables
as well as the curated lookup tables that are described in a later section. This dimensional
model is highly denormalized and eliminates the complexity from entity relationship
diagrams while providing the same approach to the logical and physical data models.
The second step of dashboard development consisted of exploratory data analysis. Histograms
and other visual summarizations of the data were created to explore the data structure
and perform initial data validation. The third step of the dashboard creation process
included data modeling, transformation, and additional validation. The fourth step—information
visualization—involved constructing an interactive user interface that adhered to
seminal visualization principles to achieve graphical excellence and integrity.[12] This fourth step included selecting the appropriate colors, eliminating distractors,
determining the proper data density, and formulating an aesthetically pleasing “golden
rectangle” interface design layout.[13] Some examples of how these principles were implemented include careful selection
of font types and sizes to maximize readability, standardized spacing among and within
cells, neutral colors for the areas framing the green-red-or-yellow e-antibiogram
cells so as not to distract the clinician, and appropriate shortening of antibiotic
and organism names to maintain uniform cell sizes. During the fifth step, iterative
usability testing, end-users used the “live” e-antibiogram dashboard and then reported
feedback that refined the dashboard's final interface. The evaluators in the fifth
step included informatics-trained pediatricians and a usability expert. They provided
suggestions on features of the interface including the e-antibiogram's layout and
color scheme.
Fig. 1 The electronic antibiogram's dimensional model diagram is based on a snowflake schema.
The dimensional model contains information regarding the Procedure Order, Order Result,
and Order Sensitivity tables as the “fact” or main table (MAIN_FACT_TABLE). The model
also contains the original antibiotic, organism, procedure, and patient and order
location tables as well as lookup tables that are curated by infectious disease specialists.
Antimicrobial susceptibility testing results were collapsed into two categories, susceptible
and nonsusceptible, and were displayed as a percentage of susceptible isolates. An
alert was added when the denominator consisted of fewer than 30 isolates.[14] Routinely tested antibiotics were included. Infectious diseases specialists at our
hospital selected the organisms for inclusion in the antibiogram based on their clinical
relevance and the frequency with which they were isolated.
Infectious disease specialists curated the content to five lookup tables (Organisms,
Antibiotics, Specimen Source, Procedures, Specimen Taken Location) to map all possible
and valid combinations. For example, for the Organisms lookup table, we created groups
for Genus, Species, Gram-Negative/Positive, Morphology, and Special Morphology. These
external mappings have enabled us to adapt the e-antibiogram to varying content in
different laboratory information systems. While the initial content for our electronic
antibiogram came from Meditech (Meditech, Westwood, Massachusetts, United States),
the e-antibiogram was adapted easily to SoftLab (SCC Soft Computer, Clearwater, Florida,
United States), which has a different set of organism descriptions. The lookup tables
are updated monthly based on the frequency of laboratory orders and the dates when
they were ordered or processed.
Susceptibility data were grouped into nine clinical scenarios as follows: community-acquired
infections (isolates obtained from outpatients, emergency department patients, or
inpatients less than 72 hours after hospital admission) sourced from blood, urine,
or a wound; hospital-acquired infections (isolates obtained from inpatients more than
72 hours after hospital admission) sourced from blood or urine; and the hospital unit
(oncology, pediatric intensive care unit, neonatal intensive care unit, and cardiac
intensive care unit).
Dashboard Validation and Clinical Comparisons
Infectious diseases specialists validated the e-antibiogram using a set of diverse
clinical microbiology scenarios, simulating the sequence and timing of information
that clinicians receive during their typical clinical workflow. Clinical validation
was performed to assess the accuracy and validity of the e-antibiogram.[15] This validation consisted of the development team exploring a variety of clinically
relevant use cases, one of which is included as an example in the “Results” section.
Dashboard Usage
The e-antibiogram was made accessible to all users via an external link located in
a drop-down menu in the EHR. Because the e-antibiogram utilizes Web-accessible visual
analytics software, only the operating system login credentials (and not the EHR login
credentials) are recorded when the antibiogram is launched. Many EHR workstation terminals
at our hospital do not require operating system logins, and the users who access the
e-antibiogram on these workstations are recorded as “unknown.”
The EHR database was queried for the operating system login credentials of users who
launched the e-antibiogram Web link from within the EHR. The users were categorized
using their operating system login credentials into the following categories: physician
house staff (fellows and residents), attending physicians, pharmacists and nurses,
physician assistants and advance practice nurses, and medical and nursing students.
Statistical Analysis
Descriptive statistics including frequency distributions were performed using Excel
(Microsoft Inc., Redmond, Washington, United States). Antimicrobial susceptibilities
were compared between cumulative and stratified e-antibiograms using chi-square or
Fisher's exact test. Difference among proportions over time was tested using a chi-square
test for trend. Exact 95% confidence intervals for binomial variables were calculated.
A two-tailed p-value of < 0.05 was considered statistically significant. All data were analyzed
using STATA v.12 software (Stata Inc., College Station, Texas, United States).
Results
The e-antibiogram was developed, implemented in the EHR, and made available to users
in July 2015. More than 6,000 inpatient, 4,500 outpatient, and almost 4,000 emergency
department isolates from January 2012 to July 2015 were included in the e-antibiogram.
The e-antibiogram is updated monthly based on EHR data. The initial user interface
screen consists of the e-antibiogram with nine preselected clinical scenarios. A variety
of filtering options frame the central display region and are accessible via open
item selectors, drop-down menus, and a date range selector slider.
Once the user has selected a clinical scenario (such as Community-Acquired Infection—Blood),
a chart with the frequency distribution of isolated pathogens is displayed ([Fig. 2]) followed by a table summarizing the antimicrobial susceptibilities for the antibiotics
that are both routinely tested and potentially useful ([Fig. 3]). Users also have the option to filter results by either rolling 12-month period
or total aggregated data. The e-antibiogram also allows users to explore the pathogen
and susceptibility data in granular detail using various tabular and graphical formats
(e.g., time series, heat maps).
Fig. 2 A screenshot of the user interface for the Community-Acquired Infection (CAI) Blood
12-month case scenario within the electronic antibiogram is shown. The initial selection
shows a graph of the distribution of isolated pathogens, while the lower panel shows
the antibiotic susceptibility summary. Users can return to the main screen, apply
the date filter, and view the pathogens that are contained in the “Others” group.
Abbreviations: CoNS, coagulase-negative staphylococci; MRSA, methicillin-resistant
Staphylococcus aureus; MSSA, methicillin- susceptible Staphylococcus aureus.
Fig. 3 A screenshot is shown of the user interface for the Community-Acquired Infection
(CAI) Blood case scenario antibiotic susceptibility summary within the electronic
antibiogram. Users can use the icons in the top right to return to the main screen,
apply the date filter, and return to the summary graph. Abbreviations: Amp/Sulbac,
ampicillin-sulbactam; Cipro, ciprofloxacin; MRSA, methicillin-resistant Staphylococcus aureus; MSSA, methicillin-susceptible Staphylococcus aureus; Pip/Taz, piperacillin/tazobactam; TMP/SMX, trimethoprim/sulfamethoxazole.
E-antibiogram Use Case
To illustrate how an e-antibiogram can aid clinicians to better target empiric antimicrobial
therapy, we selected the combination MRSA and clindamycin in the e-antibiogram due
to their clinical relevance. MRSA is a leading cause of skin and soft-tissue infections
with few oral antibiotic treatment options for children. Significant differences were
found between the hospital-wide clindamycin susceptibility of all MRSA isolates and
the clindamycin susceptibility showed in unit-specific e-antibiograms. For example,
90% of wound MRSA isolates obtained in the emergency department were susceptible to
clindamycin compared with 79% in the hospital-wide e-antibiogram [chi-square (1) = 187.74,
p ≤ 0.001]. Stratifying results by type of specimen (wound) and type of infection (community-acquired)
showed significant differences in susceptibility results. During the study period,
the clindamycin susceptibility among MRSA isolates was at least consistently 10% higher
among isolates obtained in patients presenting to the emergency department who were
discharged home compared with the cumulative clindamycin susceptibility among all
MRSA isolates. The age group filters showed similar results. The demographic data
captured in the e-antibiogram showed slightly lower clindamycin susceptibility among
wound community-acquired obtained MRSA isolates in neonates (78%) compared with older
children (88%), although this was not a significant difference between age groups
(c.001]. St = 0.37, p = 0.71).
E-antibiogram Session Counts
The e-antibiogram sessions counts have increased steadily since the tool's release.
Monthly usage data during the study period is shown in [Fig. 4] and the distribution of usage across the various user categories is listed in [Table 1].
Table 1
Categories of electronic antibiogram session users based on electronic health record
login credentials from March 1, 2016 to March 31, 2017
|
User category
|
Number of sessions
|
Percentage of total sessions
|
|
Unknown–Workstation
|
3,284
|
75
|
|
Physician house staff (fellows, residents, interns)
|
415
|
9
|
|
Pharmacists, nurses
|
331
|
8
|
|
Attending physicians
|
244
|
6
|
|
Physician assistants and advanced practice nurses
|
57
|
1
|
|
Medical and nursing students
|
45
|
1
|
|
Total
|
4,376
|
100
|
Note: The “Unknown–Workstation” category refers to use of the electronic antibiogram
at computer workstations where users' credentials cannot be tracked.
Fig. 4 Monthly electronic antibiogram aggregate session counts from March 2016 to March
2017.
An average of 11 e-antibiogram sessions per day were recorded over a 12-month period
across both antibiotic prescribers (attending physicians, fellows, residents, and
interns) and nonprescribers (pharmacists, nurses). E-antibiogram session counts increased
from an average of 261 sessions per month during the first 3 months of the study to
345 sessions per month during the final 3 months. During the study period, on workstations
that required login credentials, physicians, pharmacists, and nurses accounted for
the majority of e-antibiogram session totals.
Discussion
We developed and implemented a novel, dynamic, e-antibiogram that has successfully
and permanently taken the place of a yearly manual compilation of a static antibiogram.
Our study has three main findings. First, it is feasible to implement a visual analytics
e-antibiogram that supplies susceptibility maps for all organisms and antibiotics
in a fully comprehensive report that is updated on a monthly basis, in contrast with
a manually produced annual antibiogram that lists only the most common organisms and
antibiotics. Second, an e-antibiogram can easily stratify laboratory data by source
(wound, blood, or urine), hospital unit, and hospital- or community-acquired status,
providing greater detail than a traditional antibiogram. Third, the e-antibiogram,
which is accessed by a link in the EHR and thus available to all clinicians at all
times, has logged an increasing number of sessions per month over a 12-month period.
Prompt, appropriate empirical antibiotic treatment improves patient outcomes, while
the selection of unnecessary broad-spectrum antibiotics can increase harm to patients,
antimicrobial resistance, and hospital costs.[16] The marked escalation in the prevalence of antimicrobial resistance has made the
selection of empirical antimicrobial therapy increasingly complex, and it is based
large part on the susceptibility rates compiled in an institution's antibiogram.[17] However, some manually compiled antibiograms might have outdated antibiotic susceptibility
rates that can affect a clinician's selection of optimal empiric therapy. Our e-antibiogram
may facilitate the selection of empiric therapy by delivering near real-time access
to detailed information including organism speciation, antibiotic susceptibility,
and the cost of antibiotic options.
We implemented filters for patient characteristics and unit locations as a novel solution
to two potential shortcomings with a hospital-wide antibiogram. First, aggregate hospital
data can mask differences in susceptibility data across specific patient characteristics—such
as patient age and disease—and patient care units. Second, it can overlook pockets
of resistance within a hospital if a highly resistant pathogen affects only one patient
care unit.[18]
[19] For example, pathogens that are isolated from patients with chronic diseases typically
have higher rates of antimicrobial resistance than the pathogens obtained from patients
without chronic diseases, and a similar pattern is often seen with pathogens from
patients in intensive care units compared with patients in outpatient settings. The
impact of this improved access to granular patient- and unit-specific data remains
the focus of future study.
The e-antibiogram session data are encouraging and the use by nonprescribers suggests
the tool's potentially broad utility. We anticipate that the rate of utilization will
increase as we are actively promulgating the e-antibiogram as well as incorporating
the e-antibiogram into clinical workflows such as embedding the e-antibiogram link
in order sets. Unfortunately, analysis of the session data was hindered by the lack
of identifiable user credentials at widely available “generic” clinical workstations.
Preliminary workflow analysis has determined that attending physicians logged into
the EHR more frequently from their hospital-owned computers—which require an operating
system login and thus contribute a greater proportion of identifiable antibiogram
sessions—than other health care providers who lack individually assigned hospital
computers. This may also explain students' low percentage of identifiable sessions,
as students are more likely to use the EHR at a generic clinical workstation that
does not require an operating system login. Thus, the access patterns across the health
care provider categories might differ significantly if the “unknown” logins were to
be distributed accurately, and it is likely that users who typically lack assigned,
hospital-owned computers (such as students, pharmacists, and nurses) constitute the
majority of generic workstation e-antibiogram sessions.
Other available e-antibiograms include a Web-based antibiogram at Stanford University
and a commercial offering called iAntibiogram (Teqqa, Jackson, Wyoming, United States).[20]
[21] The Stanford e-antibiogram allows users to generate sensitivity tables from static
yearly snapshots of data that can be filtered by organism and antimicrobial drug,
but filtering by community versus hospital infections or on a rolling 12-month basis
is unavailable, as our e-antibiogram offers. iAntibiogram delivers monthly updates
as well as functionality such as filtering by hospital, unit, and source that are
very similar to our e-antibiogram. Furthermore, iAntibiogram can be used on smartphones
and the application can be installed locally, both functions that our e-antibiogram
lacks.
It is important to note the limitations of both manually compiled and e-antibiograms
to guide treatment in general. In many infections, such as pneumonia, the identification
of the pathogen is the exception. Furthermore, sicker patients may be more likely
to have a pathogen identified, which will bias any antibiogram toward more resistant
isolates. Lastly, the best use of antibiogram data is not readily evident—for example,
what percentage of resistant isolates in a given pathogen precludes empiric use of
a certain agent? Unfortunately, fully addressing these points involves decision making
beyond the implementation of an e-antibiogram. At CHOP, the selection of empiric therapies
is done in multidisciplinary committees who collaborate to create clinical pathways.
Antibiograms help to make data-driven decisions, but certainly other factors such
as the severity of the patient's illness or immune status also play a role. Antibiograms
can only relay data on isolates submitted to the microbiology laboratory. This is
a limitation of both cumulative and specific antibiograms, yet they remain the only
way to account for local patterns of resistance. However, despite their inherent limitations,
antibiograms are excellent tools to help decision making once the organism is known.
Prior to that, the user also has to take into account the likely frequency of the
organisms that might grow in culture. Site-specific data such as that presented in
the e-antibiogram might help with this clinical challenge because the e-antibiogram
can give relative frequencies of the different organisms at each site.
This study had several additional limitations. First, the e-antibiogram was developed
for use with our hospital's EHR data. Our results may not be fully generalizable to
hospitals using other EHR systems. Second, while we have obtained the session data
for the e-antibiogram, we have not assessed clinicians' empiric antibiotic prescribing
patterns both before and after the implementation of the e-antibiogram. This will
be the focus of future research. Third, while the traditional cut-off period for inpatients
is less than 48 hours after hospital admission, we used a 72-hour cut-off period that
is customary at our institution. Fourth, we did not track time efforts and costs associated
with developing the e-antibiogram. Lastly, at this time the only user interaction
that is tracked is the launching of the e-antibiogram within the EHR. The user's subsequent
interactions with the e-antibiogram, such as selection of case scenarios and application
of filters, cannot be tracked currently.
The e-antibiogram is an integral part of an ongoing quality improvement initiative
at our hospital to optimize empiric antibiotic treatment based on our institution's
EHR data. As users increasingly use the e-antibiogram, topics to be explored include
antibiotic prescribing patterns, complications and errors associated with empiric
antibiotic administration, and optimizing the e-antibiogram interface and design based
on surveys, focus groups, and user feedback. Furthermore, there is a national effort
among pediatric hospitals to share antimicrobial data and implement antimicrobial
stewardship programs, and our e-antibiogram could be a useful tool to implement at
other pediatric hospitals.[22]
[23] While we have shown that the development and implementation of an e-antibiogram
can successfully replace the manual process of antibiogram compilation, our goal is
to determine how the technology can best be utilized to improve clinical practice
and ideally, patient outcomes.
Conclusion
An e-antibiogram that is generated and updated monthly from laboratory and pharmacy
data are a feasible replacement for a yearly, static, manually compiled antibiogram.
The e-antibiogram enables users to view stratified data in ways that can elucidate
differences in susceptibility patterns that a traditional hospital-wide antibiogram
cannot. The e-antibiogram has shown signs of increased usage over a 12-month period.
Future work will examine the impact of the e-antibiogram on antibiotic prescribing
patterns.
Clinical Relevance Statement
Clinical Relevance Statement
First, it is feasible to implement a visual analytics e-antibiogram that supplies
susceptibility maps for all organisms and antibiotics in a fully comprehensive report
that is updated on a monthly basis, in contrast with a manually produced annual antibiogram
that lists only the most common organisms and antibiotics. Second, an e-antibiogram
can easily stratify laboratory data by source (wound, blood, or urine), hospital unit,
and hospital- or community-acquired status, providing greater detail than a traditional
antibiogram. Third, stratifying data in an e-antibiogram can elucidate differences
in susceptibility patterns that a traditional hospital-wide antibiogram cannot.
Multiple Choice Question
-
When constructing an interactive user interface that adheres to seminal visualization
principles to achieve graphical excellence and integrity, it is crucial to do which
of the following?
-
Vary spacing among and within cells that display data
-
Use bright, playful colors for the areas framing data cells
-
Select font types and sizes that maximize readability
-
Use nonuniform cell sizes to maximize the use of space
Correct Answer: The correct answer is c. When constructing an interactive user interface that adheres
to seminal visualization principles to achieve graphical excellence and integrity,
it is crucial to eliminate distractors, determine the proper data density, and formulate
an aesthetically pleasing “golden rectangle” interface design layout (Tufte[12]). The goal of these principles is to optimize the human–computer interaction by
building an interface that is accessible, easy to use, and efficient.
Thus, font types and sizes in the interface should be selected carefully to maximize
readability. Spacing among and within cells should remain uniform to avoid a haphazard
arrangement of cells that might adversely affect users' processing of the tabular
information. Cells containing data should maintain uniformity (if possible) for the
same reason. Lastly, the areas framing data cells should be neutral so that the user's
attention is not drawn away from the relevant information contained within the cells.
