Keywords
smart infusion pumps - analytics - clinical workflows - alerts - alarms
Background and Significance
Background and Significance
Infusion pumps are medical devices that deliver medication, fluids, and nutrients
in a precise, timely, and controlled manner that is critical to patient care. They
are widely used across both inpatient and outpatient care settings to provide critical
and life-saving care for illnesses that span the range of dehydration to cancer. Smart
infusion pumps are those equipped with dose error reduction systems (DERS) that can
alert users of programming mistakes (e.g., drug dose error by applying preset drug
limits) and operation errors (e.g., occlusion in intravenous lines). These safety
features have facilitated the adoption of the smart infusion pump in U.S. hospitals
with a rate of nearly 90% in 2017.[1]
To understand the varying impact of smart infusion pumps on patient care, many studies
have explored the care provider and work system perspective, for example, technology
design, usability, and sociotechnical system integration.[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15] Although some investigators have shown that automated smart infusion pumps improve
perceived safety, workflows, and workloads compared with manual pumps, many studies
highlighted concerns of the technology's impact on use errors, suitability across
the wide range of use environments, and workflow efficiencies.[16] Moreover, even though smart infusion pumps have the technological capabilities to
reduce the incidence of adverse drug event and medication administration errors,[17] they could also be a potential source of patient harm.[4]
[6]
[7]
[11]
[18]
[19]
[20]
[21]
[22] These show that smart pumps might have limited effects on improving patient safety.[7]
[22]
The extent of smartness of DERS depends on the dosing limit set in the drug library.
Errors in these limits can lead to programming alerts that are disruptive to nursing
workflows. Drug limits with wide ranges can potentially decrease the frequency of
alert occurrence, whereas more stringent limits can ensure adequate patient safety.
The balance between such efficacy and efficiency of a smart pump is largely reliant
on the hospital system's risk management strategy. The Institute for Safe Medication
Practices (ISMP) recommends that hospital systems routinely review and revise their
dose limit settings in the drug library to reflect changes in available infusion supplies,
clinical practices, or patient populations.[23]
Smart infusion pumps have a significant impact on care environments and care providers'
workflow. By design, they issue alerts and alarms to users. However, these audio-
or visual-based alerts and alarms can worsen the information overload situation already
experienced by clinicians. Moreover, clinicians have to address each of these infusion
alerts and alarms, thus distracting and disrupting their workflow. One study found
64% of perioperative monitor alarms in patients undergoing anesthesia to be clinically
irrelevant.[24]
[25] These clinically irrelevant alarms disrupt clinical workflows, compete for the health
care provider's attention, and can be troubling to patients. In addition to disrupting
workflows, a bad alarm system implementation can lead to alarm fatigue and, in some
cases, contribute to patient deaths.[26]
[27]
Multiple factors that adversely affect nursing staff's experience with smart infusion
pumps, including technical performance issues, alarm fatigue, and usability, have
been identified.[6]
[25] Many efforts to study these issues have been qualitative,[28]
[29]
[30]
[31] and interviews performed by researchers revealed numerous local nursing workarounds
to address mismatches between infusion pump interface design and patient care requirements.[32]
[33]
[34] Such workarounds are significant concerns and may jeopardize patient safety.[32] In addition, some quantitative and mixed method approaches to studying these issues
have also been reported.[32]
[35]
[36]
[37] One study identified the volume of infusion alarms specific to critical short half-life
infusions to evaluate user response times to these alarms.[38] In another study, formal central infusion monitoring and environmental changes in
a neonatal intensive care unit (NICU) were found to reduce alarm fatigue, improve
reaction time, and improve preempting of avoidable alarms.[39] These efforts are critical in improving nursing workflows, patient safety, and overall
health care delivery.
Objectives
This study aims to leverage the wealth of infusion event data recorded by the smart
infusion pump to understand the impact of infusion alerts and alarms on the care provider's
workflow during live use in dynamic and complex clinical settings. Specifically, we
examine the following with our dataset:
-
Risk factors that influence the occurrence of infusion programming alerts.
-
Risk factors that influence the occurrence of infusion operational alarms.
-
Contributory factors to long alarm resolution times that disrupt clinical workflows.
Methods
Dataset
The dataset used for this study was from a member health system of the Regenstrief
National Center for Medical Device Informatics (REMEDI) community of practice over
a 1-year period, within the past 5 years. REMEDI consists of a collaborative community
of 400+ hospitals that contribute infusion data to CatalyzeCare.org, a big data management
hub maintained by the Regenstrief Center for Healthcare Engineering at Purdue University.
The dataset contained infusion events from one specialty hospital and three community
hospitals. The infusion pumps of this health system were not interoperable with the
electronic health record (EHR) system at the time of data collection. The infusion
data were fully de-identified and contained only infusion event data.
Continuous full-day time-stamped “all-infusion detail report” data were extracted
from the Alaris System (Alaris System, BD/CareFusion, San Diego, CA[40]). In this study, we defined an infusion to be a series of infusion events grouped
by a unique infusion identifier (ID) assigned by the pump system. Exclusion criteria
(including incomplete and unreasonable records, patient-controlled analgesia, and
keep-vein-open infusions) were applied to 808,445 infusions, resulting in 745,641
unique infusions used for the analysis ([Fig. 1]). This dataset had 40 variables describing the state of the infusion pump, the type
or nature of the infusion, and the state of each infusion event.
Fig. 1 Exclusion criteria and justifications for the all-infusion dataset.
On the infusion pump, an alert sounds and visually appears on the screen when the
programmed infusion parameters exceed the drug's preset limits, whereas an alarm occurs
when an ongoing infusion is physically or operationally interrupted (such as due to
air-in-line or occlusion). For the purpose of this study, we deemed alerts and alarms
as undesirable “error states” during an infusion process because they are deviations
from normality, require troubleshooting steps and clinicians' extra cognitive attention,
and interrupt clinical workflows.
Data Analysis
All the analyses in this study were done with R[41] in the RStudio integrated development environment.[42] We used descriptive statistics, generalized linear models (specifically logistic
regression), and nonparametric tests for the all-infusion dataset in the study. Nonparametric
post hoc tests were performed for multiple pairwise comparisons. We identified three
dependent variables for the models used in this study: (1) presence of alerts, (2)
presence of alarms, and (3) alarm resolution times.
Some key variables like drug name, infusion type, event start time, event reason,
alarm type, etc., were extracted from the data used in this study. From these key
variables, new features including shift and day of infusion as derived from infusion
timestamps, total infusion duration, etc., were derived ([Table 1]). We categorized all infusions that started from 6 a.m. and ended before 6 p.m.
as occurring during the dayshift, whereas those that started from 6 p.m. and ended
before 6 a.m. as occurring during the nightshift. An infusion that spanned across
both day and night shifts (starting in the dayshift and ending in the nightshift and
vice versa) was categorized as shift “overlap.” Infusions spanning through Monday
to Friday were categorized as occurring during the weekday, whereas those spanning
from Saturday to Sunday were categorized as occurring during the weekend. We curated
a list of 32 high-alert medications (HAMs) commonly used in hospital settings based
on the American Society of Health-Systems Pharmacists (ASHP) Standardize 4 Safety
(S4S) initiative list of high-risk adult continuous infusion drugs ([Supplementary Table S1] [available in the online version]).[43]
Table 1
Custom variables derived from the all-infusion dataset
|
Derived variables
|
Definitions
|
Frequency observed
|
|
Institution
|
Hospital 1
|
207,960
|
|
Hospital 2
|
68,757
|
|
Hospital 3
|
288,745
|
|
Hospital 4
|
180,179
|
|
Shift
|
Dayshift
|
213,522
|
|
Nightshift
|
114,880
|
|
Overlap
|
417,239
|
|
Day of the week
|
Weekday
|
576,777
|
|
Weekend
|
168,864
|
|
High-risk medication
|
Yes
|
173,297
|
|
No
|
572,344
|
|
Total infusion duration (h)
|
Hospital 1
|
398,084
|
|
Hospital 2
|
171,691
|
|
Hospital 3
|
575,951
|
|
Hospital 4
|
392,665
|
|
Date range (months)
|
Hospital 1
|
12
|
|
Hospital 2
|
12
|
|
Hospital 3
|
8
|
|
Hospital 4
|
11
|
All predictor variables used in the models were selected by experts (clinical and
health researchers). [Table 2] lists and describes each of these variables and their justification for inclusion
in the models. Multicollinearity across the predictor variables for the generalized
linear model (GLM) was checked using the generalized variance inflation factor (GVIF)
where each GVIF value was raised to the power of (1/2 df) for comparability across
dimensions[44] (df = predictor variable degree of freedom). Because this value is analogous to
the square root of the usual VIF,[45] we square it and apply the usual VIF rule of thumb.
Table 2
Predictor variables and justifications
|
Predictor variable
|
Justification
|
|
1
|
Institution
|
Modeling this predictor variable accounts for institutional effects that can potentially
affect our response (e.g., presence of alerts and alarms) and human factors and systems
engineering focused predictor variables, e.g., shift and days of the week
|
|
2
|
Profile
|
Identifying critical care profiles at the highest risk of infusion-related errors
can help determine areas to focus medication safety efforts.[20] This can lead to minimized alerts and alarms and seamless nursing workflows. We
model this variable to further corroborate the literature on the effect of care settings
on nursing workflows
|
|
3
|
Shift
|
Researchers have studied the effects of day and night shifts on clinicians' cognitive
functions, fatigue level, attention, and performance.[57]
[58]
[59] One study also found that with each successive hour that passed in a nurse's shift,
response time to monitor alarms in a pediatric hospital was slower.[60] We also understand that night shifts tend to be more understaffed than day shifts
in hospitals. This is mostly anecdotal but very widely accepted. Therefore, we model
this variable to address any variability of these shift situations and routines and
to avoid trend bias
|
|
4
|
Day of the week
|
A recent study reported higher average number of alarms on Saturdays and Sundays and
thought it warranted further investigation.[36] We also understand that the weekend shifts tend to be less staffed than weekday
shifts in hospital settings. We model this variable to further investigate its effect
on the presence of alerts and alarms as well as alarm resolution times
|
|
5
|
High-alert medication (HAM)
|
Identifying HAM drugs as recognized by the ASHP[43] as a potential predictor variable that can affect the occurrence of alerts corroborates
the literature while shedding more light on how they can disrupt workflows
|
|
6
|
Number of alerts
|
We model whether or not the presence of an alert before an infusion has any effect
on the presence of alarms during the infusion
|
|
7
|
Infusion duration
|
For medications with short half-life, there is a tradeoff between flow rate and time
to occlusion alarm. Flow rate is a function of volume-to-be-infused (VTBI) and infusion
duration. One study reported long times to occlusion alarm in peristaltic infusion
devices at low flow rates.[61] We model this variable to further investigate any relationships between infusion
duration and the presence of alarms
|
|
8
|
Infusion type
|
Different types of infusions (e.g., intermittent or continuous) might potentially
affect the presence of alarms during an infusion. We consider this variable to get
more insights into the effects of the various infusion types in the dataset on alarm
occurrences
|
|
9
|
Alarm type
|
Depending on the nature of alarm encountered during an infusion, alarm resolution
times might differ.[54] We model this variable to determine which operational alarm types might affect alarm
resolution times
|
The three study objectives were addressed as follows:
-
Risk factors for alerts: A logistic regression model was used to explore the relationship between the predictor
variables and the occurrence of alerts at the infusion programming stage. The predictor
variables in the model for alerts include hospital, care unit, shift, HAM, and day.
The binary response variable for this model was coded as 0 (absence of an alert during
programming) and 1 (presence of at least one alert during programming).
-
Risk factors for alarms: A logistic regression model was used to explore the relationship between the predictor
variables and the occurrence of alarms during infusion. The following were the independent
variables for this model: hospital, care unit, shift, infusion type, HAM, day, number
of alerts, and infusion duration. The binary response variable for this model was
coded as 0 (absence of an alarm during an infusion) and 1 (presence of at least one
alarm during an infusion).
-
Alarm resolution times: As a result of non-normal alarm resolution times across all alarmed infusions, a
nonparametric alternative to the analysis of variance (ANOVA) was used. Specifically,
Kruskal–Wallis tests were conducted to evaluate the differences among factor levels
of some predictor variables on alarm resolution times. We applied some exclusion criteria
to the alarmed infusion dataset before conducting these tests ([Table 3]). The tests were corrected for tied ranks and Dunn's nonparametric post hoc tests
were also performed for multiple pairwise comparisons across factor levels. The evaluated
independent variables include alarm type, shift, and day.
Table 3
Infusion alarmed dataset exclusion criteria
|
Criterion
|
Justification
|
|
1
|
Excluded 77,016 (17.9%) unresolved/cancelled infusions
|
These were excluded because alarm resolution times could not be calculated from these
infusions. This could be due to incomplete data capture by the pump system
|
|
2
|
Excluded 17,605 (4.1%) infusions alarmed with zero resolution time
|
These were excluded because zero resolution times are not physically feasible. This
may be due to data capture issues and require closer examination
|
|
3
|
Excluded 33, 563 (7.8%) infusions with alarm resolve time greater than the 90th percentile,
i.e., outliers
|
These were excluded because of very long resolution times that could skew our results
|
|
4
|
Excluded 16,151 (3.8%) infusions due to missing pump values (incomplete data point)
|
These were excluded because of incomplete data possibly caused by pump data capture
error
|
Results
A total of 3,231,300 infusion events corresponding to 745,641 unique infusion IDs
were observed for 1,538,391 total hours of infusion (an average of 2.06 hours per
infusion) across the four facilities. Overall, 28.7% of all unique infusions had at
least one operational alarm and 2.1% of all unique infusions had at least one programming
alert.
Alert and Alarm Prevalence by Infusion Type
About 30% of all unique infusions encountered at least one error state. Grouped by
infusion type, fluid infusions had the highest percentage of error-state occurrence
(31%; [Table 4]). Intermittent infusions were the most common infusion type in the dataset (33%
of all infusions) with ∼30% of them encountering an error state. Thirty-one percent
of basic infusions, which are administered without using DERS, had alarms. By definition,
basic infusions do not trigger programming alerts. Continuous/bolus infusions were
the least common (18% of all infusions) and had the lowest percentage of error-state
occurrence (27%).
Table 4
Prevalence of error-free and error-state infusions in all-infusion dataset by infusion
type
|
Infusion type
|
Error free
|
Error state
|
Total infusion by type (% of total)
|
|
No alert/no alarm
|
% of total
|
Alarm only
|
Alert only
|
Alert and alarm
|
Subtotal
|
% of total
|
|
Basic infusion
|
147,753
|
68.6
|
67,723
|
0
|
0
|
67,723
|
31.4
|
215,476 (28.9)
|
|
Intermittent infusion
|
174,619
|
69.8
|
66,528
|
6,698
|
2,466
|
75,692
|
30.2
|
250,311 (33.6)
|
|
Fluid infusion
|
101,901
|
68.9
|
44,421
|
914
|
555
|
45,890
|
31.1
|
147,791 (19.8)
|
|
Continuous/bolus infusion
|
96,354
|
73.0
|
30,665
|
3,532
|
1,512
|
35,709
|
27.0
|
132,063 (17.7)
|
|
Subtotal
|
520,627
|
69.8
|
209,337
|
11,144
|
4,533
|
225,014
|
30.2
|
745,641
|
Approximately 70% of infusions encountered no error states. However, ∼29% of the infusions
had at least one alarm event that required the caregiver's attention. Prevalence of
alerts was lower than that of alarms at 1.5% of all unique infusions ([Table 4]). Finally, a very small percentage of infusions had both an alarm and an alert (0.6%).
High-Risk Medications
Alert and alarm prevalence varied for the 32 HAMs, as well as maintenance fluids ([Supplementary Table S1] [available in the online version]). These 32 high-risk drugs represented 17% of
all infusions, and 24% of HAM infusions encountered at least an alert or an alarm.
Propofol, heparin, insulin regular, fentanyl, and norepinephrine combined accounted
for 58% of all HAM infusions. About 2.5% of all HAM infusions had programming alerts,
whereas less than 0.1% of maintenance fluids had such alerts. About 22% of all HAM
infusions had operational alarms, whereas ∼33% of maintenance fluids had such alarms.
Risk Factors Modeling and Analysis
Separate models were developed to address the objectives of this study. Results showed
that all the variables used in the logistic regression models had VIF < 6, which suggested
that multicollinearity was not a concern[46] [[Supplementary Materials Tables S2] and [S3], available in the online version].
Presence of Alerts
Care unit (profile), shift, day of the week, and HAM were significant predictors (p < 0.01) of alerts ([Table 5]). The institution was not a statistically significant predictor variable in the
model. Respectively, the odds of an alert were 1.79 and 2.13 higher for infusions
in the pediatrics and labor/delivery profile than those in the adult medical/surgical
profile. A medication infusion in the HAM list of this study was 1.64 times more likely
to have a programming alert when compared with all other infusions in the study dataset.
For each of these odds ratio comparisons considered, all other predictor variables
are kept constant ([Table 5]).
Table 5
Logit model coefficient estimates, odds ratios for alert occurrence (keeping other
predictors constant), and corresponding p-values for predictor variables (with presence of alert as dependent variable)
|
Coefficients
|
Odds ratio
|
p-value
|
|
Intercept
|
–3.65
|
0.03
|
0.00
|
|
Institution
|
|
Institution 2
|
Reference group
|
|
Institution 1
|
–0.22
|
0.80
|
0.00
|
|
Institution 3
|
–0.22
|
0.80
|
0.00
|
|
Institution 4
|
–0.05
|
0.95
|
0.11
|
|
Profile
|
|
Adult medical/surgical profile
|
Reference group
|
|
Adult intensive care unit (ICU)
|
–0.32
|
0.73
|
0.00
|
|
Labor and delivery
|
0.76
|
2.13
|
0.00
|
|
Nursery
|
0.10
|
1.11
|
0.00
|
|
Pediatrics
|
0.58
|
1.79
|
0.00
|
|
Shift
|
|
Day shift
|
Reference group
|
|
Night shift
|
–0.41
|
0.66
|
0.00
|
|
Overlap shift
|
–0.01
|
0.99
|
0.00
|
|
Day of the week
|
|
Weekday
|
Reference group
|
|
Weekend
|
–0.16
|
0.86
|
0.00
|
|
High-alert medication (HAM)
|
|
All drugs not on the HAM list in this study
|
Reference group
|
|
HAM
|
0.50
|
1.64
|
0.00
|
Presence of Alarms
Institution, care unit (profile), shift, infusion type, HAM, infusion duration, and
number of alerts were significant predictors (p < 0.05) of alarms ([Table 6]). The day of the infusion was not a statistically significant alarm predictor. The
odds of an alarm occurring are slightly reduced in the labor and delivery profile
than in the adult medical/surgery profile (odds ratio = 0.92). The odds of an alarm
occurring during infusions that span across shifts was 2.68 times more than those
during the dayshift. Primary intermittent infusions were 1.68 times more likely to
have an alarm than fluid infusions. For every 1-hour increase in the infusion duration,
there was a 7% increase in the odds of an alarm occurring. For each of these odds
ratio comparisons considered, all other predictor variables are kept constant ([Table 6]).
Table 6
Logit model coefficient estimates, odds ratios for alarm occurrence (keeping other
predictors constant), and corresponding p-values for predictor variables (with presence of alarm as dependent variable) logit
model
|
Coefficients
|
Odds ratio
|
p-value
|
|
Intercept
|
–1.58
|
0.20
|
0.00
|
|
Institution
|
|
Institution 2
|
Reference group
|
|
Institution 1
|
0.11
|
1.12
|
0.00
|
|
Institution 3
|
0.08
|
1.08
|
0.00
|
|
Institution 4
|
–0.03
|
0.97
|
0.01
|
|
Profile
|
|
Adult medical/surgical profile
|
Reference group
|
|
Adult intensive care unit (ICU)
|
0.00
|
1.00
|
0.90
|
|
Labor and delivery
|
–0.09
|
0.92
|
0.00
|
|
Nursery
|
–0.52
|
0.59
|
0.00
|
|
Pediatrics
|
–0.67
|
0.51
|
0.00
|
|
Shift
|
|
Day shift
|
Reference group
|
|
Night shift
|
–0.05
|
0.95
|
0.00
|
|
Overlap shift
|
0.99
|
2.68
|
0.00
|
|
Infusion type
|
|
Fluid infusion
|
Reference group
|
|
Basic primary infusion
|
–0.08
|
0.92
|
0.00
|
|
Basic secondary infusion
|
–0.59
|
0.56
|
0.00
|
|
Continuous/bolus infusion
|
–0.01
|
0.99
|
0.13
|
|
Drug Calculation Continuous/Bolus infusion
|
–0.60
|
0.55
|
0.00
|
|
Primary intermittent infusion
|
0.52
|
1.68
|
0.00
|
|
Secondary intermittent infusion
|
–0.57
|
0.56
|
0.00
|
|
Day of the week
|
|
Weekday
|
Reference group
|
|
Weekend
|
0.01
|
1.01
|
0.39
|
|
High-alert medication (HAM)
|
|
All drugs not on the HAM list in this study
|
Reference group
|
|
HAM
|
–0.53
|
0.59
|
0.00
|
|
Infusion duration
|
0.07
|
1.07
|
0.00
|
|
Number of alerts
|
–0.06
|
0.94
|
0.00
|
Alarm Resolution Time
Mean number of alarms observed per unique infusion was 2.01 for a total of 430,585
alarm events. After applying all exclusion criteria, the data for this model had 286,250
alarm events. We defined alarm resolution as the process of an infusion moving from
an alarm error state to an alarm error-free state. Therefore, this resolution time
includes a wait time between when an alarm sounds and when the nurse gets to the patient's
bedside to address it. This captures the disruptions caused by infusion alarms on
clinical workflows as both wait times and actual time spent resolving alarms translate
to an interruption in the nurse tasks at hand. In the study dataset, a total of 12,822.5 hours
elapsed between when alarms sounded and when they got resolved. This translates to
a total of 10 to 13 hours of nursing time for alarm resolution in a day for the community
hospitals. For the specialty hospital in the study, it means a total of 2 hours of
nursing time in a day was spent on alarm resolutions. These are concurrent and may
include the wait times for nurses to get to the bedside to address these alarms. Mean
resolution time for 74.5% of the alarms in the study dataset was ≤1 minute. However,
∼8% of alarms took more than 4 minutes to get resolved.
Kruskal–Wallis tests were conducted to evaluate differences among the types of operational
alarms encountered during infusion, nursing shifts, and days on median change in alarm
resolution times. The tests, corrected for tied ranks, were all significant for each
independent variable at p < 0.001:
χ
2(11, N = 286,250) = 33,858, χ
2(2, N = 286,250) = 6,071.6, and χ
2(1, N = 286,250) = 358.74
Follow-up Dunn's tests were performed to evaluate pairwise differences among the 12
types of alarms, 3 levels of shifts, and 2 groups of day variables. Type 1 errors
across tests were controlled by Bonferroni's correction. Not all pairs of alarm types
were statistically significant and median resolution times were longest for both cumulated
air-in-line and patient-side occlusion alarms and shortest for door-close alarm (p < 0.001). All pairs of shift levels were statistically significant (p < 0.001) and median alarm resolution times were the longest for overlap shifts compared
with both day and night shifts. Finally, median alarm resolution times were longer
for weekends than weekdays.
Discussion
Infusion pump data analytics and human factor contributions continue to be an important
area of research to reduce medication-related patient safety events. To minimize potential
harm in clinical settings, the Association for the Advancement of Medical Instrumentation
(AAMI), the U.S. Food and Drug Administration (FDA), and the ISMP have highlighted
several patient safety priorities for use of infusion pumps, including establishing
processes for analyzing infusion incidents, mitigating use errors, and understanding
use environments.[16]
[25] In this work, we leveraged big data available from smart infusion pumps and identified
risk factors for programming alerts and operational alarms. Specifically, we analyzed
infusion data from four hospitals and quantified infusion alert and alarm impact on
nursing workflows, examined their potential effect on patient safety, and determined
their associative factors related to patients and health care providers. This showcased
how infusion pump informatics can facilitate our understanding of user–pump interactions
in the highly dynamic and complex clinical setting.
Our analysis discovered a high prevalence of error-state occurrences of programming
alerts and operational alarms during infusion pump use. An infusion alarm signals
a physical issue that has stopped the ongoing infusion and requires the clinician's
time and effort to resolve it. Unlike other types of alarms that occur in clinical
settings, studies have shown that infusion pump alarms are unique because they require
the caregiver's presence and interaction with the pump to resolve the alarm.[35]
[47] In the study dataset, alarms averaged 2.01 times per infusion for a total of 745,641
infusions and 430,585 alarm events. Similar alarm analysis done with a different vendor
infusion pump reported an average of 1.74 alarms per infusion for a total of 568,164
infusions and 987,240 alarm events.[36] This shows that even with differences in vendor infusion pumps and number of infusions
delivered, infusion pump alarms are prevalent in clinical settings. Other studies
have also reported this high prevalence of infusion pump alarms across various clinical
settings.[48] Specifically, ∼8% of the alarms took more than 4 minutes to resolve. These lengthy
resolutions of infusion alarms can draw time away from other clinical tasks or interrupt
the nurse's task at hand. Long infusion interruptions are even more undesirable as
medications that have a short half-life for effectiveness require rapid alarm resolution,
imposing an additional time pressure on clinicians.[38]
The findings of this work provide additional insights into the impact of smart infusion
pump alerts and alarms on nursing workflows. Infusion alerts and alarms are mechanisms
designed to aid medication administration by drawing caregivers' attention to issues
that can potentially affect a patient's safety. However, there are tradeoffs between
protecting patients and maximizing nursing workflow efficiency since alerts and alarms
cause infusion interruptions. Because time and efficiency constraints are critical
in busy, stressful, and very high-risk clinical settings, incessant infusion interruptions
are undesirable. Smart infusion alarms add to an environment that is already saturated
with other medical technologies that generate alarms. One study reported medical staff
members in an intensive care unit (ICU) being repeatedly exposed to an average of
45.5 alarms per patient per hour, and infusion pumps contributed to almost 10% of
the total alarm burden.[37] Consequently, large volumes of infusion pump false alarms can contribute to the
alarm fatigue problem in these environments and lead to clinicians getting desensitized
to these warnings. These alarms also sound at high decibels that can potentially lead
to rest or sleep interruption and cause concern, annoyance, confusion, or burnout
to patients.[49]
[50] These issues can expand clinicians' roles and duties, erode their trust in medical
technologies, and lead to stress and dissatisfaction.[51] Therefore, hospital systems should create efficient risk management strategies that
prioritize safe infusion processes for patients and clinicians alike.
Although the rate of infusion alert was lower than that of alarms in the dataset,
alerts also had an impact on nursing workflows as they occurred at least once in every
48 unique infusions, for a total of 22,568 alert interruptions. These interruptions
occurred during infusion programming steps, some of which resulted in reprogramming
attempts. We recognized a portion of them may be “good catches” that averted potential
patient harm. However, an increased alert rate can also contribute to the alert/alarm
fatigue and desensitize clinicians' safety awareness. This may result in the clinician's
override of an alert without proper confirmation. It may also contribute to ∼30% of
non-DERS (i.e., “basic infusions”) use observed in the dataset as a way to avoid undesired
alerts deemed by clinicians as disturbances. This fell short of the ISMP safe infusion
guidelines of targeting 95% use of DERS,[23] and it indicates the need for more nursing education and trainings.
In this study, we found that more alerts occurred during the programming attempts
of the selected HAMs—1.64 times more likely to occur than other drugs; other factors
held constant. Since no standards have been established regarding an optimal alert
rate, individual hospitals often conduct their own infusion practice reviews. If a
drug's alert rate is deemed too high, the hospital needs to evaluate its limit settings.
In some cases, these settings might not be in line with clinical use, or some nurses
might not be well informed of the hospital's practice. We also observe higher frequencies
of alarms and alerts for infusions that span across shifts in comparison to those
during day shift. In particular, alarms were 2.68 times as likely to happen during
infusions that span across shifts than during dayshift; other factors held constant.
This may be attributed to issues around nursing handoffs (because of shift change)
or increased patient movement during specific hours that lead to obstruction of infusion
flow. As seen in one study, implementing a mandatory alarm parameter checklist during
nurse staffing handoffs might reduce the incidence of more alarms during infusions
that span across shifts.[35]
Researchers have investigated the incidence of infusion pump alarms and alerts in
different care units.[11]
[37]
[50] This study also performs similar analysis to investigate the effect of care unit
on the occurrence of alerts and alarms. Our analysis showed that the odds for infusion
alerts in labor/delivery profiles were 1.79 and 2.12 times higher than those in adult
medical/surgical profiles; other factors held constant. Alarms were also as likely
to occur in labor/delivery profiles as in adult medical/surgery profiles. These may
be good indicators of higher complexity of medication use and patient conditions in
the labor/delivery and similar units. One study reported a high volume of alerts clustered
around specific patients and mediations in a NICU.[52] This can potentially lead to a high alert burden and limit DERS safety benefit by
desensitizing nurses to these alerts. Another study found that due to line occlusions,
drug incompatibilities, and patient factors, alarms from infusion pumps were frequent
in the NICU/pediatric ICU.[53] This also implies that more training or special coordination is required of nursing
staff in those units with respect to infusion administration.
Our analysis also showed that “door-close alarm” had the least effect on alarm resolution
times. This is reasonable since “door-close alarm” simply signals that the pump door
was properly closed after it opened during an infusion. We also notice that “cumulated
air-in-line alarm” had one of the most significant effects on alarm resolution times.
This aligns with our understanding of the physics of such alarms. They are caused
by a large number of air bubbles in the infusion line which are detected by the pump
sensor. It requires significant effort and time to clear the infusion line (including
potentially priming) and resolve the alarm. An increased rate of occurrence of these
alarms may require further investigation of the root cause. Depending on its exact
nature, potential interventions include nursing education/training on spiking and
priming processes, allowing cold intravenous solutions to warm before handling, and
adding assisting devices such as an antisiphon valve (ASV). These have shown to be
effective in reducing air-in-line for some drugs.[54]
Although the independent variables used in our models are not exhaustive, our analysis
identified some system variables that might be attributing to the presence of alerts
and alarms and longer alarm resolve times. Since these are disruptive, they lead to
clinical workflow interruptions. In 2005, Brixey et al showed that workflow interruptions
in clinical settings are multidimensional across person–device and person–person interruptions,
the physical layout of workspaces, and work practices within departments.[28] Factors like how fatigued clinicians are (gotten through psychophysiological measures),
the pump maintenance culture in the facility, and the level of training given to the
nurses can also affect the occurrence of alerts and alarms and their resolution times.
More work is needed to explore how such factors directly influence the time needed
to resolve alarms.
The results from this study have several implications for hospital management. For
example, the proposed all-infusion analysis approach can be applied to prioritize
medication administration processes or quality improvement initiatives. Specifically,
medications with higher alert counts and rates may warrant a review of the drug limit
settings. These reviews can help check for inconsistencies in the ordering set and
the pump drug library. An unusually high frequency of some infusion alarms, such as
an air-in-line alarm, may indicate a training issue with the nurses' priming practice.
Furthermore, when the average infusion alarm resolution time is significantly longer
at night and overlap shifts than day shifts in a specific unit, it may indicate nightshift
understaffing in that unit. In summary, we envision these analyzed results can raise
red flags and help the leadership prioritize how to utilize their limited resources
(including technology, manpower, and time) on improving workflows and patient safety.
Limitations
Infusion data in this study were from one pump vendor. Thus, findings may not generalize
to hospitals using a different vendor. However, the workflow analysis framework is
applicable to all infusion pumps with similar available data. Infusion pump data in
this study were not linked to clinical incidents or patient outcomes, and linked datasets
are needed to understand direct infusion pump impacts on patients' safety. Studies
have also shown that smart pump–EHR interoperability can decrease alert firing rate.[23]
[55]
[56] Future intervention studies are needed to address the workflow disruptions caused
by infusion events on user's trust of the pump, as well as their overall job satisfaction.
Finally, the models in this study only explore the relationships between a nonexhaustive
list of predictors and the presence of alerts and alarms during infusion. Machine
learning algorithms for variable selection and predicting the occurrence of infusion
alerts and alarms might provide more insight into future works.
Conclusion
This study presents an analysis of a detailed infusion pump dataset that captured
over 700,000 unique infusions across four hospitals. Our results highlight several
implications of pump alerts and alarms that may impact patient care and nursing workflows.
This analytics-based approach can facilitate a greater understanding of infusion-related
clinical tasks, workflow coordination, and patient safety considerations across various
clinical settings through the use of infusion pump data in a manner not limited by
observation-based methods.
Clinical Relevance Statement
Clinical Relevance Statement
The findings from this paper adds value to the body of knowledge about smart infusion
pump use and practice, supporting how disruptive infusion pump alarms and alerts can
be to nursing workflows and patient safety in clinical settings. Moreover, risk factors
for these alarms and alerts have been identified, ensuring that clinicians are aware
of these contributory factors to workflow interruptions while using smart infusion
pumps. Consequently, hospital management can apply this approach to deploy efficient
risk strategies and quality improvement initiatives that prioritize safe medication
infusion processes.
Multiple Choice Questions
Multiple Choice Questions
-
Which of the following has been discovered as an adverse effect of the adoption of
smart infusion pumps in clinical settings?
Correct Answer: The correct answer is option d. In as much as smart infusion pumps have tremendous
advantages, their adoption comes with some adverse effects. These include disruptions
to nursing workflows (as alarms and alerts require the caregiver's attention), patient
harm, and alarm fatigue.
-
High levels of smart infusion pump programming alert rate can lead to:
-
High levels of alarm rates.
-
Non-DERS infusions (i.e., without using drug limit settings in the pump software).
-
Longer alarm resolution times.
-
Elimination of all programming errors.
Correct Answer: The correct answer is option b. High levels of alerts can contribute to alert fatigue
problem in the clinical environment and thus desensitize the caregiver's safety awareness
and lead to unsafe overriding of programming alerts.
