Keywords clinical decision support systems - workflow - emergency department - antimicrobial
stewardship
Background and Significance
Background and Significance
Clinical Decision Support in Emergency Departments and Clinical Workflow
Electronic health record (EHR)-based clinical decision support (CDS) systems have
the potential to change clinical behavior in the emergency department (ED) setting.
Decision support interventions tailored to context and end-user needs have been shown
to be more likely to result in desired behavior change.[1 ] Although CDS has promise to impact prescribing practices, the ED setting presents
unique challenges to CDS implementation and effectiveness.[2 ]
[3 ] Clinicians in the ED are frequently forced to make rapid decisions in a chaotic
setting. Decision making is often complicated by patient turnover, multiple interruptions,
the need for empiric therapy, and limitations such as a lack of perceived follow-up
and medicolegal fears.[4 ]
[5 ]
[6 ]
[7 ] To be successful, current literature and CDS experts strongly recommend that developers
assess and account for local ED processes and clinician perceptions during the creation
and refinement of CDS, but prior to implementation.[8 ]
[9 ]
[10 ]
[11 ]
Antimicrobial stewardship programs (ASPs) aim to optimize appropriate antibiotic prescribing
in the clinical setting.[12 ] CDS can deliver antibiotic prescribing recommendations at the point-of-care, and
overcome barriers unique to the ED setting.[13 ]
[14 ]
[15 ] Therefore, CDS is a potentially effective method to implement ED-based ASPs. However,
an understanding of ED workflow prior to development of CDS for antibiotic prescribing
is necessary to develop an effective and sustainable ASP intervention. Proper design
and implementation of CDS can be integrated into the usual workflow without disrupting
ongoing practice.[16 ]
[17 ] The purpose of this study was to characterize workflow for four clinician roles
in a pediatric ED, to inform the future development of EHR-based CDS, as the centerpiece
for an ED-based ASP.
Potential Role for CDS in Antimicrobial Stewardship Programs
Antimicrobial resistance is among the greatest threats to health care quality and
pediatric public health.[18 ] In outpatient settings such as the ED, it is estimated that 25 to 63% of patients
are prescribed inappropriate or unnecessary antibiotics, contributing to the development
of antimicrobial resistance.[19 ]
[20 ]
[21 ]
[22 ] ASPs can improve guideline adherence for antibiotic prescribing and minimize the
impact of antibiotic-resistant organisms.[23 ] However, ASPs are not sufficiently utilized in ED settings, owing to the erratic
workflow and variety of providers unique to this setting.[24 ] Innovative methods are required to implement ASPs into the ED setting.
Use of health information technology is one possible method to potentially optimize
ASPs in the ED.[13 ]
[14 ] Incorporation of health information technology and use of EHR is viewed as an essential
method for future CDS interventions to mitigate antibiotic prescribing.[14 ]
[15 ]
[25 ] Implementation of CDS using the EHR as a vehicle to deliver recommendations has
positively impacted provider antibiotic prescribing in various clinical settings.[15 ]
[25 ]
[26 ]
[27 ]
[28 ]
[29 ]
[30 ] Trials of CDS have demonstrated success in improving antibiotic prescribing in inpatient
and office settings, though ED-based studies are limited.[25 ]
[27 ]
[28 ]
[31 ]
[32 ]
[33 ]
[34 ]
[35 ]
[36 ] However, the success of these CDS interventions largely depends on proper integration
into the clinical workflow.[26 ]
[37 ]
[38 ]
Methods
Setting
Workflow was systematically examined in a single, tertiary-care academic children's
hospital ED that serves as regional referral center, receiving ∼73,000 visits annually.
The ED consists of 40 acute care beds and is staffed by a variety of providers, including
attending physicians (pediatric emergency medicine [PEM] specialists and general pediatricians),
nurse practitioners (NPs), physician assistants (PAs), and resident physician trainees.
At any given time, the ED is usually staffed by 2 to 3 PEM specialists, 5 advanced
practice providers (NPs or PAs), and between 2 and 4 residents. The ED is grouped
into three teams of providers: two teams are a combination of PEM specialists, residents,
and advanced practice providers. A third team is composed of two advanced practice
providers who evaluate patients of low acuity. In the system, all advanced practice
providers act independently, without direct supervision. Residents present to PEM
specialists only, and are fully supervised.
Sample
ED clinicians of four provider types were observed in the ED over a 5-month period
(March–July 2016). Data were collected from a total of 23 participants (11 attendings,
3 residents, 5 NPs, and 4 PAs). [Table 1 ] shows the participating providers and observation characteristics. Participants
volunteered from a pool of 40 attending physicians, 30 residents (on ED rotation),
and 70 advanced practice providers. We focused recruitment on the day shift, to ensure
homogeneity of the data.
Table 1
Study sample
Type of providers
Clinicians
(n )
Sessions
(n )
Time spent in the ED
(h)
Time captured with observation tool
(h)
Decision point questions
(n )
Attending
11
13
41
38
65
Resident
3
4
11
10
12
Nurse practitioner
5
8
24
22
32
Physician assistant
4
6
14
13
18
Total
23
31
90
83
127
Abbreviation: ED, emergency department.
Procedure
Data were collected by direct observations supported by probing questions. Workflow
activities and information regarding decision points were recorded on a tablet computer,
using a user-friendly software tool for collection of time and motion data, where
the tablet automatically records the time each recording of an activity was entered
([Fig. 1 ]). The software tool[39 ] was initially developed for and validated in an ambulatory care setting. We modified
the tool by adding the ED setting and observer names while keeping the task list and
other settings. Coding of workflow activities is outlined in [Appendix A ]. A total of 90 hours of observations were conducted.
Fig. 1 Data collection tool.
Appendix A
CATG_ID
CATG_DESC
A1
Computer–Communicating
B1
Dictating
C1
Computer–Entering
C2
Computer–Login
C3
Computer–Logout
C4
Computer–Processing
C5
Computer–Printing
C6
Computer–Reading
D1
Paper–Copying/Faxing
D2
Paper–Reading reference
D3
Paper–Reading/Writing
D4
Paper–Retrieving/Accepting
D5
Paper–Sorting/Filing/Delivering
D6
Paper–Writing
E1
Phone–Answering
E2
Phone–Calling
E3
Phone–Transferring
F1
Talking–Coworker
F2
Talking–Patient
F3
Talking–With others
G1
Walking–Alone
G2
Walking–With coworker
G3
Walking–With patient
H1
Meeting
I1
Performing–Exam room preparation/Cleaning
I2
Performing–Hand sanitization
I3
Performing–Handling
I4
Performing–Measuring
I5
Performing–Medical procedure
I6
Performing–Physical exam
I7
Performing–Other
J1
Personal–Cell phone
J2
Personal–Computer
J3
Personal–Eating/drinking + Socializing/chatting
J4
Personal–Restroom
J5
Personal–Other
K1
Cell phone/iPad–Reading/typing
To better understand the critical decision-making moment, clinicians were queried
in real time about the timing of diagnosis, and disposition decision points, for each
patient that had a prescription or discharge instruction entered in the EHR. Decision
points were based on the following question: “For [this specific patient] at what
point during her/his visit did you decide how to proceed with the treatment?” A total
of 127 decision points were identified for providers ([Tables 2 ] and [3 ]). We primarily targeted diagnosis point rather than disposition with this question.
However, we acknowledge that diagnosis and disposition can be very close temporally
for some health care encounters. We acknowledge that consideration of both decision
points is very important. Regardless, from the ASP CDS design perspective the ambiguity
does not matter because in both cases (diagnosis or disposition) the function of the
CDS would be the same.
Table 2
Summary of decision points
Decisions (total frequency of responses for each decision)
Provider type
Frequency
1. After/during examining or talking to patient or relative (55)
Attending
27
NP
17
PA
6
Resident
5
2. After talking to a specialist (13)
Attending
8
NP
3
PA
1
Resident
1
3. After test/image is resulted and discussion with patient/family (13)
Attending
8
NP
3
PA
2
4. During/after discussion with a resident (5)
Attending
5
5. After administration of medication and reexamining (4)
Attending
1
NP
2
PA
1
6. During/after discussing with an attending (2)
Resident
2
7. After reviewing patient history and conducting exam (2)
NP
2
8. Reading through the patient chart (2)
Attending
2
9. After seeing the test results and conducting a second physical exam (2)
PA
1
Attending
1
10. After making calculations on a Web site (1)
Attending
1
11. Talking to a respiratory therapist (1)
Attending
1
12. After going through head injury protocol; and observation protocol (1)
NP
1
13. After asthma assessments (1)
Resident
1
14. After gathering her history, conducting a physical exam, and talking to an attending
(1)
PA
1
15. After obtaining a second opinion (1)
NP
1
16. After obtaining medical history and final exam (1)
NP
1
17. After referring to TBI algorithm (1)
PA
1
18. After medical procedure (1)
Attending
1
19. When hearing other attending taking the transfer phone call (1)
Attending
1
20. Prior to the patient's arrival at the ED, as well as after speaking with the patient's
doctor who referred the patient (1)
Attending
1
21. After reading triage notes (1)
Resident
1
Total
110
Abbreviations: ED, emergency department; NP, nurse practitioner; PA, physician assistant;
TBI, traumatic brain injury.
Table 3
Summary of cases with no specific decision
Reasons for no decision
Provider type
Frequency
1. Complex patient so no decision during the observation period
Attending
5
NP
2
PA
1
Resident
2
2. Social worker consultation
Attending
1
3. Psychiatry patient
Attending
2
PA
3
4. Observer could not ask
Attending
1
Total
17
Abbreviations: NP, nurse practitioner; PA, physician assistant.
The observer recorded reflection statements after each of the 31 observation sessions.
These statements were used to clarify both qualitative and quantitative data. Reflection
statements are important to explain why there was a decision or if there is any additional
important information (e.g., location) that might be important for analysis.
Decision Analysis
We used a qualitative approach to analyze decision points and a quantitative approach
to analyze workflow data. Qualitative analysis of the 127 decision point responses
was accomplished by summarizing the decision points within the overall care delivery
process and highlighting the patterns by roles of the four types of providers. Responses
by the participants were categorized using constant comparison. Key events associated
with decision making were identified.
Sequential Pattern Analysis
The quantitative analysis of workflow data consisted of two parts. In the first part,
the analysis was accomplished by sequential pattern mining,[40 ]
[41 ] a subfield in data mining that focuses on identifying frequent patterns in sequence
data. Prior to applying the sequential pattern mining methods, we manipulated the
workflow activity data into an appropriate format for the algorithm input. Specifically,
tasks observed in the care delivery process were categorized ([Appendix A ]) and transformed into a string sequence for each observation. For example, if a
clinician walks alone (G1), talks to a patient (F2), and then performs a physical
exam during the observation (I6), a sequence of “G1-F2-I6” for this observation is
generated. To make the patterns more interpretable for clinical experts, we constrained
the sequential pattern mining method to identify only adjacent events with a focus
on start and end points. We ignored frequent task patterns that did not occur in a
consecutive manner. To accommodate this constraint, we added a pseudo start and end
point, “S1” and “S9” correspondingly, to each observation sequence.
The manipulated workflow activity data were then fed into a computerized algorithm
developed by the second author (D.T.W.) to identify frequent sequence patterns. This
identification was guided by the a priori algorithm, which has two parameters. The
algorithm constrained a minimum support of 20%, i.e., a pattern is considered frequent
when one-fifth of the observations contained this pattern, and the minimum confidence
of 80%, i.e., a task is added to the frequent pattern list if this task is observed
4 out of 5 times using conditional probability. Once the frequent task sequences were
identified, we visualized these patterns in a directed network graph ([Fig. 2 ]) for each clinical role, with the nodes being the frequent categories and the edges
being the connection in the patterns. Based on these parameters, the network graphs
were laid out in a circle to facilitate interpretation. It is worth noting that the
research team iteratively reviewed and interpreted the frequent patterns in the network
graphs to adjust the parameters to include a reasonable amount of meaningful patterns,
resulting in the threshold of minimum 20% support and the 80% confidence.
Fig. 2 Workflow patterns for the four providers.
The second part of the quantitative analysis focused on determining whether the frequency
of, and the time allocated in the tasks, were significantly different among the providers
based on their clinical role. In particular, the frequency (count) and the time allocation
(percentage) were aggregated based on the task categories per observation using the
same mapping definition in [Appendix A ]. Multitasking events were split into several single events and their elapsed time
were evenly distributed to the single events. For example, if event A lasts 10 seconds,
event B lasts 15 seconds, A and B are overlapped by 5 seconds, then the multitasking
part is split equally between A and B so that A is 7.5 seconds and B is 12.5 seconds
in a single tasking manner. The mean of the frequency of, and the time allocated to
the categories by clinical roles, was compared using the Kruskal–Wallis test, assuming
two independent samples with a nonnormal distribution. The significance level was
set at 0.05.
The data were manipulated in Python 2.7 and stored in a SQLite database (https://www.sqlite.org/ ), a relational file-based database format. The sequential pattern mining procedure
and the statistical test were also conducted using Python 2.7. Specifically, the string
matching procedure was implemented using the generic Python regular expression library;
the Kruskal–Wallis test was implemented using the Python Numpy and Scipy libraries.
The network graphs were produced using the igraph library in R-package in the R-Studio
environment, an open-source interactive code editor for R.
During the study, we collected qualitative and quantitative data simultaneously. Our
analysis of the qualitative data showed saturation before data collection was complete.
We argue that our sample is sufficiently representative because: (1) we did not pick
the participants purposefully, it was based on volunteering; (2) observers did not
note a significant event during the observations that may affect the usual workflow;
and (3) the sample size is consistent with similar studies.
Results
The 90-hour observations resulted in 6,060 records with 64 distinct activities. Among
these activities, Five were conducted at different frequency or time allocation across
the four clinical roles, including computer entering (C1) and reading (C6), talking
with coworkers (F1) and walking alone (G1), and personal activities (J3) as listed
in the second part of [Fig. 2 ]. We also identified the frequency of decision points and compared the sequential
patterns of the activities across the four roles.
Decision Points
During the observation period, of 127 decision point questions, in 17 questions (9
attending, 2 NPs, 4 PAs, and 2 residents), no specific decisions were made toward
the treatment (see [Table 3 ]). A common reason for the lack of a decision point was among those are complex patients;
still need further tests and consultations, and those coming to ED for mental health
reasons; these patients were being screened in the ED for medical clearance. The 110
remaining decision points within the care delivery process are presented in [Table 2 ].
The most frequent decision point for each of the four providers was “after/during
examining or talking to patient or relative” in 55 (50%) responses. Other frequent
decision points were: “after talking to a specialist” and “after test/image result
is resulted and discussion with patient/family” with 13 (12%) total responses for
each. The second and third most frequent decision points for attending physicians
and NPs were “after talking to a specialist” and “after test/image result is resulted
and discussion with patient/family.” “After test/image result is up (and talking to
patient/family)” was the second frequent decision point for PAs, while “during/after
discussing with an attending” was the second most frequent decision point for residents.
Out of 110 visits, antibiotics were prescribed in 20 cases. In 12 of those cases,
the decision was made “after/during examining or talking to patient or relative.”
The reflection notes highlight how various activities were conducted in various places
within the ED. For example, conversations among ED staff members took place at physician
stations, hallways, or in front of electronic board. Similarly (face-to-face or phone),
conversations between ED clinicians and specialists took place in various places within
the ED. Although the majority of physical exams took place in patient rooms, at least
in one case it took place in a triage room. Test results were accessed by clinicians
through desktop computers a majority of the time; however, attendings could be aware
of test results by residents in hallways.
Sequential Patterns
[Fig. 2 ] illustrates the most frequent sequential pattern in a network graph for each clinical
role. Attending physicians demonstrated the most clear-cut pattern (fewer frequent
patterns in total) compared with other providers. The attendings frequently started
with talking to coworkers (F1), walking alone (G1), and personal activities such as
socializing (J3). They were frequently involved in computer entering (C1) and reading
(C6) toward the end of the observations. Other providers, on the other hand, involved
more tasks that the attendings did not frequently do, such as answering phone calls
(E1), hand sanitization (I2), and performing physical exams (I6). This pattern reasonably
reflects the roles these clinicians play on a care team. It is worth noting that residents
and PAs were more likely to talk to patients (F2) in the beginning of the observations.
They also tended to talk to their coworkers at the end of the observations (F1–>S9),
which likely involved reporting to attending physicians and seeking guidance and approval.
Residents exhibited the most complicated and frequent task patterns compared with
other providers, meaning that residents had higher numbers of frequent patterns than
others, and the sequence of these patterns had more steps and more switches between
task categories.
Frequency and Time Allocation
Significant differences (p < 0.05) were seen in the frequency and task categories among providers ([Table 4 ]). Attending physicians demonstrated a very different pattern compared with the other
clinical roles. For example, they more frequently talked to coworkers (F1) and read
electronic documents (C6) than other providers. Attending physicians talked to coworkers
about 20 minutes longer than other providers. NPs, on the other hand, tended to allocate
more time than attendings to entering tasks on the computer (22.29 vs. 8.27 seconds),
but performed measures (1.0 vs. 2.0%) less frequently than attendings. Residents walked
alone longer then NPs (4.68% vs. 1.86%).
Table 4
Activities that were conducted in different amounts by four roles
Measure
Category
Roles compared (median)[a ]
Number of observations
Median difference
p -Value
Time allocation (s)
F1: Talking–Coworker
Attending (2926.87)
NP (1140.65)
13, 8
–1786.22
0.000
Attending (2926.87)
PA (893.22)
13, 6
–2033.65
0.002
Attending (2926.87)
Resident (1520.75)
13, 4
–1406.12
0.000
Time allocation (%)
C1: Computer–Entering
Attending (8.27)
NP (22.29)
13, 8
14.02
0.001
F1: Talking–Coworker
Attending (30.58)
NP (12.54)
13, 8
–18.04
0.000
Attending (30.58)
Resident (19.02)
13, 4
–11.56
0.001
G1: Walking–Alone
NP (1.86)
Resident (4.68)
8, 4
2.82
0.006
Frequency
C6: Computer–Reading
Attending (57.38)
NP (22.13)
13, 8
–35.25
0.000
Attending (57.38)
PA (26.67)
13, 6
–30.71
0.000
Attending (57.38)
Resident (30.25)
13, 4
–27.13
0.002
E1: Phone–Answering
Attending (9.92)
NP (1.8)
13, 5
–8.12
0.001
Attending (9.92)
PA (2.0)
13, 3
–7.92
0.001
Attending (9.92)
Resident (3.0)
13, 4
–6.92
0.006
E2: Phone–Calling
Attending (6.45)
PA (1.8)
11, 5
–4.65
0.009
F1: Talking–Coworker
Attending (50.92)
NP (26.0)
13, 8
–24.92
0.000
Attending (50.92)
PA (15.17)
13, 6
–35.75
0.001
Attending (50.92)
Resident (29.75)
13, 4
–21.17
0.003
I4: Performing–Measuring
Attending (2.0)
NP (1.0)
3, 2
–1
0.000
I5: Performing–Medical procedure
Attending (1.14)
NP (2.0)
7, 2
0.86
0.001
Time allocation (s)
C1: Computer–Entering
Attending (707.50)
NP (2186.75)
13, 8
1479.25
0.001
Attending (707.50)
PA (1914.00)
13, 6
1206.50
0.011
C6: Computer–Reading
Attending (1546.50)
NP (923.58)
13, 8
–622.92
0.010
E1: Phone–Answering
Attending (419.17)
NP (80.00)
13, 5
–339.17
0.006
F1: Talking–Coworker
Attending (2880.50)
NP (992.33)
13, 8
–1888.17
0.000
Attending (2880.50)
PA (677.42)
13, 6
–2203.08
0.002
Attending (2880.50)
Resident (1520.75)
13, 4
–1446.50
0.005
G1: Walking–Alone
NP (182.5)
Resident (303.0)
8, 4
120.50
0.009
Time allocation (%)
C1: Computer–Entering
Attending (7.72)
NP (22.89)
13, 8
15.17
0.001
Attending (7.72)
PA (23.64)
13, 6
15.92
0.005
Attending (7.72)
Resident (16.92)
13, 4
9.20
0.005
E1: Phone–Answering
Attending (4.43)
NP (0.86)
13, 5
–3.57
0.009
F1: Talking–Coworker
Attending (31.52)
NP (10.95)
13, 8
–20.57
0.000
Attending (31.52)
PA (7.43)
13, 6
–24.09
0.011
G1: Walking–Alone
Attending (2.80)
Resident (4.62)
13, 4
1.82
0.012
NP (1.86)
Resident (4.62)
8, 4
2.64
0.003
I6: Performing–Physical exam
Attending (3.84)
Resident (6.39)
13, 4
2.55
0.012
Frequency
C6: Computer–Reading
Attending (58)
NP (22)
13, 8
–36.00
0.000
Attending (58)
PA (24)
13, 6
–33.50
0.001
Attending (58)
Resident (29)
13, 4
–29.00
0.006
E1: Phone–Answering
Attending (9)
NP (1)
13, 5
–8.00
0.003
Attending (9)
PA (2)
13, 3
–7.00
0.011
E2: Phone–Calling
Attending (5)
PA (1)
11, 5
–4.00
0.008
F1: Talking–Coworker
Attending (48)
NP (22)
13, 8
–26.00
0.000
Attending (48)
PA (10)
13, 6
–38.00
0.001
Attending (48)
Resident (31)
13, 4
–17.00
0.003
Abbreviations: NP, nurse practitioner; PA, physician assistant.
a Nonnormal distribution, using Kruskal–Wallis test to compare the pre- and postmedian.
Discussion
We examined general workflow to provide context around a specific task, i.e., antibiotic
prescription to inform the development of ASPs, in EDs. Although the parent project
of this study was ASP implementation and we conducted the study within the ASP context,
the main focus of this study is to highlight workflow variability across clinical
roles. Because of higher relevancy, we identified decision points within the workflow.
The most frequent decision point for each of the four providers was after/during examining
or talking to patient or relative in 55 (50%) responses. Where an antibiotic is prescribed,
in 60% of the cases, the decision is made at this same point. However, as we demonstrated
in this study, decisions are made at different times and at different locations. The
sequential pattern analysis also demonstrated that clinicians with different roles
may have different and more frequent patterns in their activities. Our analysis showed
that residents have the most complicated frequent patterns involving more diverse
(e.g., talking and computer-related) and interconnected activities. On the other hand,
while attendings have more frequent and longer conversations with coworkers overall,
they tend not to end their processes with talking to coworkers but computer reading
and entering. This is consistent with their role in demonstrating participation in
decision making and “signing off” on the appropriateness of the plan of care. Therefore,
the design of CDS should consider clinical roles, the accessibility at different points
of care, and physical locations. CDS should provide best decision using up-to-date
evidence-based clinical guidelines and most available data about the patient. These
findings highlight the following three main principles regarding the design and implementation
of the CDS for antibiotic prescribing that would support development of an ED-based
ASP.
CDS should support a variety of workflows: CDS should activate at the time when clinical
decisions are made.[42 ] Our results indicate that providers with different backgrounds often take different
steps and make decisions at different points within care delivery. This difference
reflects their role and expertise in health care. CDS has the potential to understand
a user's practice and provide decision support individualized to clinicians' roles.
For ASP, the CDS could not only identify the choice of antibiotic and potential variability
in prescribing practices, but consistency of practice with guidelines. CDS can benefit
best when providing some standardization without harming flexibility that is due to
the nature of the work.
CDS should support clinicians within different physical locations in the ED: Clinicians
are often at different locations when they make decisions regarding treatment plans.
For example, talking to a patient or their families mostly occurs at the patient's
bedside while talking to a specialist often occurs in the hallways or on the phone
in front of the computer screen. CDS should ideally be activated the moment when the
decision regarding antibiotic prescription is made at these diverse locations.
CDS should support decisions at different points of care: Qualitative analysis ([Table 2 ]) shows that depending on the patient's symptoms and diagnosis, prescription decisions
are made at different temporal points within the care. Quantitative analysis showed
that overall workflow varies across roles. Decision support systems should be contextually
sensitive in terms of understanding the clinician's overall workflow for a specific
patient, including identifying the timing of care to ensure that decision support
is provided at the right time. If CDS is not activated in time, the recommendation
is more likely to be ignored.
If the CDS is embedded in the EHR, convenient access to EHR at all locations and situations
should be provided to clinicians. Having computer access (e.g., strategically placed
screens on walls informed by workflow) and providing easy logging in/out features
can be solutions to support these principles. Another solution would be providing
mobile access to clinicians. CDS that follows these principles will be more context-sensitive
and will potentially provide better cognitive support to clinicians who then can make
better prescription decisions. Although our focus was CDS for ASP, these principles
apply to CDS for other purposes as well. CDS vendors could improve CDS performance
by developing diverse interaction points that allow clinicians to receive needed decision
supports at different physical locations as well as stage in process. An important
bottleneck of CDSs is availability of needed data (e.g., laboratory and imaging results)
in real time. For sooner access to needed data, vendors can also improve performance
by focusing on better interoperability across systems that provide the needed data.
While writing an outpatient antibiotic prescription may be similar to other medication
orders, the decision to prescribe is based on more complex data than other medications.
With antibiotic prescriptions, data from patient exam, results of diagnostic testing,
prior antibiotic use, etc., combined with local factors such as bacterial resistance
for the area, all must be considered. This decision making is usually less straightforward
than selecting a bronchodilator or analgesic. However, from a workflow perspective,
antibiotic prescription can be considered a prescription task. This specific study
is the first step in designing an ASP initiative. We intended to understand workflow
so that the new ASP initiative is congruent with the current practice.
Two recent studies[42 ]
[43 ] highlighted the necessity of examining workflow for the success of ASP implementation
through EHR in EDs. As opposed to previous studies[44 ]
[45 ] that focused on time distribution of tasks across various roles, this study also
examines the differences in the sequence of activities by various roles in ED. Sequence
is an important building block of workflow in EDs.[3 ]
[46 ] Development of next-generation CDS for ASP which follows the three principles listed
above may involve mobile applications and systems, which may significantly change
existing workflow. Therefore, the new workflow should be consistent with the variability
of care delivery work and should not hurt needed flexibility of clinicians' workflow
to accomplish high-quality patient care and avoid unintended consequences.
In terms of methodological contribution, we recorded the clinical workflow activities
using a tablet-based application and analyzed the data using a set of computational
methods developed by the authors.[39 ] The successful completion of the data collection and analysis of this study demonstrates
the potential generalizability of the application and the analysis methods, which
were designed based on two previous publications about standardizing time motion data
collection[47 ] and advancing time motion data analysis to uncover hidden patterns.[48 ] Another methodological contribution of this study is the combination of statistical
and sequential pattern analysis on the workflow activity data. The former provides
an overview of the frequency and time allocation of tasks in each clinical roles;
the latter dives into the processes to provide a more granular view of activity sequences.
This combined method enables both macro and micro view of workflow data, facilitating
a more detailed understanding of the nuances of behavior patterns in clinical workflow.
There are limitations to our findings. This study was conducted in a single setting
and the majority of observations were conducted by a single observer. However, pilot
observations (observations before we started collecting data that was not included
in the analysis) were conducted by two observers before data collection, and demonstrated
excellent interobserver reliability. Another limitation is that the workflow activity
data collected in this study, using a time and motion design, may not be comprehensive
due to the medium sample size. However, we believe our samples are representative
and we focused only on important tasks in the data collection process. This specific
study is the first step in designing an ASP initiative. We intended to understand
workflow so that the new ASP initiative is congruent with the current practice. We
focus on day shift to ensure homogeneity of the data. However, we added this as a
limitation.
Future studies will include leveraging EHR data to complement observation data with
patient-specific and operational information such as diagnosis, demographics, and
workload demands of the ED. Future studies will also include applying the data collection
tool in other scenarios (such as hand-offs) to improve its ability to be customized
for the setting and the generalizability of tool. This study examined general workflow
and diagnostic decision in general. Our future efforts will also focus on cases in
which antibiotics were prescribed and more detailed steps specific to antibiotic prescription
and its unique characteristics.
Conclusion
The design and implementation of the CDS to implement an ED-based ASP should support
all four ED provider roles, who often have different workflows. Clinicians make their
decisions about treatment at different points of care delivery; a CDS for antibiotic
prescribing also should support decisions at these different points of care.
Clinical Relevance Statement
Clinical Relevance Statement
CDS is the preferred method for implementation of ASP;[24 ] however, timing of CDS activation and presentation to clinicians is challenging
in the ED setting. The ED is characterized by an erratic interrupted workflow, and
a variety of providers and provider types that make decisions regarding antibiotic
prescribing at different time points during the ED visit.[6 ] Integration into the clinical workflow is among the most important preferences of
ED providers with respect to ED-based ASP implementation.[42 ] The results of our study demonstrate the complexity of the ED workflow, and the
differences that exist among various ED providers. These workflow patterns will be
essential to inform and develop our CDS for antibiotic prescribing, which will serve
as the centerpiece for an ED-based ASP.
Multiple Choice Question
Which of the following statements is incorrect when designing and implementing clinical decision support systems in EDs?
CDS should support workflow for various type of clinical roles
Clinical decision support should support clinicians within different physical locations
in the ED
CDS should support decisions at different points of care
The functionality of CDS should depend time of the day
Correct Answer: The correct answer is option d. As data analysis revealed in this study, CDS should
support workflow for various clinical role. CDS also support clinicians within different
physical locations and CDS should support decisions at different points of care. Although
functionality of CDS can depend on various contextual factors such as workload in
ED or individual characteristics of the clinicians, there is no evidence that the
functionality of CDS should depend on time of the day.
