Context Usual drug-prescribing practices may not consider the effects of renal
insufficiency on the disposition of certain drugs. Decision aids may help
optimize prescribing behavior and reduce medical error.
Objective To determine if a system application for adjusting drug dose and frequency
in patients with renal insufficiency, when merged with a computerized order
entry system, improves drug prescribing and patient outcomes.
Design, Setting, and Patients Four consecutive 2-month intervals consisting of control (usual computerized
order entry) alternating with intervention (computerized order entry plus
decision support system), conducted in September 1997–April 1998 with
outcomes assessed among a consecutive sample of 17 828 adults admitted
to an urban tertiary care teaching hospital.
Intervention Real-time computerized decision support system for prescribing drugs
in patients with renal insufficiency. During intervention periods, the adjusted
dose list, default dose amount, and default frequency were displayed to the
order-entry user and a notation was provided that adjustments had been made
based on renal insufficiency. During control periods, these recommended adjustments
were not revealed to the order-entry user, and the unadjusted parameters were
displayed.
Main Outcome Measures Rates of appropriate prescription by dose and frequency, length of stay,
hospital and pharmacy costs, and changes in renal function, compared among
patients with renal insufficiency who were hospitalized during the intervention
vs control periods.
Results A total of 7490 patients were found to have some degree of renal insufficiency.
In this group, 97 151 orders were written on renally cleared or nephrotoxic
medications, of which 14 440 (15%) had at least 1 dosing parameter modified
by the computer based on renal function. The fraction of prescriptions deemed
appropriate during the intervention vs control periods by dose was 67% vs
54% (P<.001) and by frequency was 59% vs 35% (P<.001). Mean (SD) length of stay was 4.3 (4.5) days
vs 4.5 (4.8) days in the intervention vs control periods, respectively (P = .009). There were no significant differences in estimated
hospital and pharmacy costs or in the proportion of patients who experienced
a decline in renal function during hospitalization.
Conclusions Guided medication dosing for inpatients with renal insufficiency appears
to result in improved dose and frequency choices. This intervention demonstrates
a way in which computer-based decision support systems can improve care.
Renal insufficiency is relatively common among hospitalized patients,
and is associated with an increase in hospitalization-related morbidity and
mortality.1-4
Persons with acute and chronic renal insufficiency are hospitalized with increased
frequency compared with those nonaffected, due to renal disease per se, and
to the effects of renal insufficiency on other medical conditions, including
congestive heart failure and chronic liver disease.5,6
Practitioners caring for these patients are faced with the challenges of managing
the complex interplay between renal insufficiency and other organ system disease,
and of altering diagnostic studies (eg, angiography) and therapeutics to avoid
further renal injury. With regard to renal insufficiency and pharmacotherapeutics,
the majority of clinicians' attention has been directed at avoiding nephrotoxic
drugs in patients at risk for worsening renal failure; comparatively little
attention has been paid to the disposition of drugs, nephrotoxic and nonnephrotoxic,
in patients with renal insufficiency. In prescribing drugs for patients with
renal insufficiency, most practitioners rely on their clinical experience
or the advice of consultant physicians or pharmacists to guide dosing regimens.
Since few clinicians are expert in this area, and medication orders can rarely
be delayed until consultation is obtained, the capacity to provide information
on drug disposition in real time might be of great value to the practicing
clinician and the patient.
The problem of error in medicine has been found to be important and
costly.7 Adverse drug events (ADEs) are common
and often associated with errors.8 Even basic
computerization of physician ordering with relatively little decision support
was associated with a 55% decrease in serious medication errors, and an 84%
decrease in near misses or potential ADEs in 1 study by our group.9 However, only a 17% decrease was seen in preventable
ADEs. The study suggested that computerized advice regarding the dosing of
drugs in the setting of renal insufficiency might be among the most potent
additional preventive strategies.10
Thus, we hypothesized that the incorporation of guided dosing algorithms
for inpatients with renal insufficiency into an existing computer order entry
system would result in a larger proportion of appropriate dose and frequency
orders, and would be associated with shorter lengths of stay (LOS), lower
costs, and a lower frequency of worsening renal function.
The study was carried out at Brigham and Women's Hospital, a 720-bed
urban tertiary care academic medical center in Boston, Mass. The Brigham Integrated
Computing System (BICS) provides administrative and clinical computing services
at BWH. All inpatient orders are entered into BICS, including orders for medications,
laboratory and radiology studies, and for nursing interventions. The BICS
order entry application provides the physician with a range of possible dose
amounts for that medication (dose list) along with 1 dose that is highlighted
as the default or recommended dose amount (Figure 1A). The clinician is also offered a highlighted frequency
as the recommended dosing interval (Figure
1B). The clinician can also hit an additional key to see the data
used for calculation of creatinine clearance. Nearly all laboratory, radiology,
and pathology results, admission vital signs (including weight), and demographic
information can be accessed.
The BICS system had for some years contained an on-line, noninteractive
version that could be accessed separately from the order entry system. In
an attempt to enhance the impact of this application within the BICS, its
internal logic was integrated with the computerized laboratory results reporting
system, and was incorporated into the order entry system. Based on information
already in the reporting system, the new application first determined whether
a patient had renal insufficiency, defined as an estimated creatinine clearance
of less than 80 mL/min (1.34 mL/s), by the Cockroft-Gault equation.11 Next, based on the real-time calculation of the estimated
creatinine clearance and the drug being prescribed, the application would
modify the above-described dose list, default dose amount, and default frequency
(dosing interval) in the BICS (Figure 2).
After reviewing the relevant literature, an expert panel including a
nephrologist, a pharmacist, and a general internist convened to review all
medications in the hospital's drug formulary and selected those medications
that were renally cleared and/or nephrotoxic. New dosing suggestions were
generated in a subset of approximately 500 medications (approximately 2500
in total). To smooth dose recommendations, renal insufficiency was divided
into 3 categories: mild (estimated creatinine clearance, 50-80 mL/min [0.84-1.34
mL/s]), moderate (estimated creatinine clearance, 16-49 mL/min [0.27-0.82
mL/s]), and advanced (estimated creatinine clearance, ≤15 mL/min [≤0.25
mL/s]). The expert panel then determined optimal adjustments in dose list,
default dose amount, and default frequency for each of the medications in
the application in each of the renal insufficiency categories. The nonfixed
variables in the estimated creatinine clearance calculation (ie, weight, serum
creatinine) were the weight entered by the nurse or physician into the BICS
database on admission. The latest serum creatinine level was entered by the
laboratory and updated regularly during the hospital stay.
All persons admitted to the medical, surgical (including subspecialty
surgical services), neurology, and obstetrics and gynecology services between
September 1997 and April 1998, whose admission and discharge were within the
boundaries of 4 consecutive 2-month periods were included in the study. Admission
periods did not overlap.
Intervention and Evaluation
When renal insufficiency was detected and any medication was ordered,
the application potentially modified 1 or more of the dose list, default dose
amount, and default frequency. To test the effect of this application, an
intervention trial was designed. The study periods consisted of 4 alternating
8-week blocks of intervention and control subperiods. Throughout the intervention
and control periods, the application was active, determining whether the dose
list, default dose amount, and default frequency needed adjustments. During
the intervention periods, the adjusted dose list, default dose amount, and
default frequency were displayed to the order-entry user and a notation was
provided that adjustments had been made based on renal insufficiency. During
the control periods, these recommended adjustments were not revealed to the
order-entry user, and the unadjusted parameters were instead displayed.
A log was kept of all instances in which an application medication was
ordered and the application adjusted the dose list, default dose amount, and/or
default frequency. A log was also kept of the order finally made by the ordering
physician. A selection was considered appropriate if the dose amount or frequency
interval did not exceed the parameters set forth by the expert panel.
If use of a particular medication was considered potentially harmful,
the application would provide feedback to the ordering clinician, accompanied
by a recommendation for a suitable substitute when appropriate. For instance,
if meperidine hydrochloride were prescribed for a patient with an estimated
creatinine clearance of less than 15 mL/min (<0.25 mL/s), a warning regarding
its potential for promoting seizures would be issued, with a suggestion that
an alternative narcotic analgesic be prescribed. The clinician could then
either accept or override such a recommendation.
Patient outcomes were determined during discrete admissions. Length
of stay was recorded in days. Hospital and pharmacy costs were estimated from
billed charges and institution-specific charge-to-cost ratios.
Continuous data were presented as mean (SD) or median (interquartile
range), and compared with the t test or Wilcoxon
rank-sum test, as appropriate. Categorical data were presented as proportions
and compared using the χ2 test. Multivariable linear regression
analysis was used to compare LOS and costs (both log-transformed) in the intervention
and control periods. Age, sex, and diagnosis related group (DRG) weight12 were used as covariates in these analyses. In addition,
we evaluated (using multiplicative interaction terms) whether the effect of
the application intervention differed by age, sex, or DRG weight. To determine
whether the exclusion of patients whose admission extended across study periods
exerted any meaningful effects on the analyses of LOS, costs, and renal function,
we repeated the analyses without these exclusions. Each patient was assigned
to the group (intervention or control) based on the day of admission. All
reported P values were based on 2-tailed tests of
statistical significance. Analyses were conducted using SAS statistical software
(SAS Institute Inc, Cary, NC).
There were 19 982 admissions that either began or ended during
the 8-month study period; we focused on the 17 828 that were wholly contained
within a study subperiod. There were 7887 (39.5%) admissions wholly contained
in the 2 intervention periods and 9941 (49.7%) admissions wholly contained
in the 2 control periods (corresponding to 58 912 and 70 821 patient-days,
respectively). There were 2154 (10.8%) admissions that straddled a study-period
boundary and were excluded. In-hospital mortality rates were 1.8% and 1.9%
(intervention vs control, P = .61). Mean (SD) age
(52.5 [18.4] years vs 52.5 [18.3] years; P = .95)
and sex (61.4% vs 61.8% female; P = .78) were not
significantly different across periods. The mean DRG weight was higher during
the control periods (2.3 vs 2.1 in intervention periods; P = .004). The majority of patients (11 896 [60.1%]) had estimated
creatinine clearance values greater than 80 mL/min (>1.34 mL/s). One in 4
patients (4927 [24.9%]) had mild renal insufficiency. Fifteen percent had
moderate (2563 [12.9%]) or advanced (414 [2.1%]) renal insufficiency. The
mean estimated creatinine clearance at admission was higher during the intervention
periods (90.9 vs 84.7 mL/min [1.52 vs 1.41 mL/s] in control periods; P<.001).
There were a total of 2 278 723 orders during the study period,
773 113 of which were medication orders and 108 537 of which were
orders for nephrotoxic and/or renally cleared medications. We excluded 11 386
orders because of missing dose amount (3794 [33.3%]) or frequency interval
(5102 [44.8%]), or because of an uninterpretable estimated creatinine clearance
value (3696 [32.5%]), usually indicating an aberrant weight measurement, and
for a variety of other less common reasons (2588 [22.7%]). These exclusions
left 97 151 orders for analysis (orders could be excluded for >1 reason).
Of the 97 151 analyzable orders, the application generated a suggestion
for the clinician in 14 440 (15%). Table 1 shows a detailed array of these suggestions. Table 2 shows the proportion of orders deemed appropriate, stratified
by whether the the application's suggestion was dose-related, frequency-related,
or both. In the intervention vs control periods, the frequency of appropriate
orders was 51% vs 30% for all relevant orders, 67% vs 54% for orders involving
dose changes, and 59% vs 35% for orders involving frequency changes, respectively
(P<.001 for all comparisons).
LOS, Costs, and Renal Function
Table 3 shows unadjusted
LOS and costs (hospital and pharmacy) during the intervention and control
periods. The rightward half of the table shows the effect of including the
2154 hospitalizations that overlapped. Hospitalizations were categorized as
intervention or control based on conditions on the day of admission.
The adjusted mean LOS (adjusted for age, sex, and DRG weight) remained
significantly shorter during the intervention period, both when overlapping
admissions were included (P = .002) and when they
were excluded (P<.001). The effect of the application
on LOS was attenuated at higher DRG weights (P<.001).
In contrast, there were no significant differences in adjusted mean total
or pharmacy costs between intervention and control periods.
A 10-mL/min (0.17-mL/s) decrement in estimated creatinine clearance
from admission to discharge was considered to be of clinical significance.
The percentage of patients whose estimated creatinine clearance declined by
more than 10 mL/min (0.17 mL/s) was 11.8% and 11.5% (intervention vs control, P = .43). The mean (SD) changes in estimated creatinine
clearance were 1.9 (0.2) mL/min (0.03 [0.003] mL/s) and 2.3 (0.2) mL/min (0.04
[0.003] mL/s) during the corresponding periods (P
= .18).
We were successful in designing and implementing a computer order entry-based
application that provided real-time drug prescription decision support to
physicians. Compared with control periods during which information was readily
available on-line but not incorporated into the order-entry process, the application
intervention led to a statistically significant and clinically meaningful
increase in the proportion of prescriptions considered appropriate for inpatients
with renal insufficiency.
The large improvements in appropriateness of dosing and frequency were
probably realized in part because the application is largely transparent to
the clinician. Its key characteristics are that it remembers a huge amount
of data essentially impossible for clinicians to master (and keep updated),
and it makes it easy to do the right thing. Nonetheless, despite the overall
improved appropriateness of dosing, 49% of orders for the application's drugs
were still inappropriate in the intervention group. Some physicians may have
been reluctant to reduce drug dosages, particularly among more critically
ill patients. Others may have simply disregarded the advice in favor of their
own established practice patterns. Future studies with this application and
similar applications should investigate the reason(s) for accepting or rejecting
on-line advice regarding medication ordering, and it might be worthwhile to
consider stronger suggestions in specific situations.
A number of other studies have evaluated the impact of decision support
on dosing of medications for patients with renal insufficiency. For example,
Rind et al13 developed an application that
alerted physicians caring for inpatients when there was an increase in the
patient's serum creatinine concentration. An alert was triggered by a 0.5
mg/dL (44.2 µmol/L) increase in serum creatinine if the patient was
prescribed a potentially nephrotoxic medication (eg, aminoglycoside), and
a 50% increase in serum creatinine, to at least 2.0 mg/dL (176.8 µmol/L),
if prescribed a medication that was renally excreted (eg, digoxin). The alert
was delivered by e-mail to physicians who had accessed computer-based information
on the affected patient in the 3 days preceding and following the increase
in serum creatinine. The intervention resulted in a significant decrease in
the frequency of more severe renal dysfunction, although fewer than half of
the recipients (44%) found the alerts helpful and 28% found them "annoying."
It is also noteworthy that Rind et al excluded patients on all services other
than medicine, and all patients with preexisting moderate or severe renal
insufficiency (serum creatinine >3.0 mg/dL [265.2 µmol/L]).
In another important study, one in a series evaluating the influence
of computerized decision support, investigators at LDS Hospital in Salt Lake
City, Utah, incorporated renal function assessment into an application that
assisted physicians in prescribing antibiotics in an intensive care unit.14 These authors found that the use of their program
decreased the frequency of inappropriate antimicrobial prescriptions (ie,
orders for drugs to which patients had reported allergies, antibiotic susceptibility
mismatches, and excessive drug dosages), and ADEs. Among patients who received
recommended regimens, there was a significant decrease in LOS and drug and
total hospital costs. More recently, Nightingale et al15
implemented a program in the renal unit of a British teaching hospital. Clinicians
cancelled more than half of their orders when they were warned that the drug
dosage they had requested was excessive. In the Nightingale et al study, there
were no formal comparisons made between presystem and postsystem implementation
periods with regard to appropriateness of orders, costs, complications, or
hospital LOS.
The application used here differs from prior applications in that it
is generalized to all hospitalized patients, provides suggestions for a wide
range of drugs, and does so in real time. Feedback is most likely to be successful
if it is delivered in real time, and in close temporal proximity to the decisions
being made.9 As noted earlier, while we found
that computerized physician order entry reduced the frequency of serious medication
errors, it had a bigger impact on errors that did not actually cause injury
compared with those that did injure patients.9
We believe—although this needs to be validated—that part of the
reason for the larger impact on potential ADEs than actual ADEs was that the
systems evaluated did not include sophisticated decision support, such as
that provided by the application described here. With widespread application
of sophisticated decision support, major reductions in ADE frequency as well
as improvements in efficiency should be possible.
It is unclear why LOS was reduced by the new application's activity.
Typically, LOS is a downstream indicator of quality of care. Because of resource
constraints, we were unable to evaluate the more subtle effects of the application.
For example, avoidance of overdosing of selected drugs in elderly patients
may have led to fewer central nervous system or gastrointestinal tract adverse
effects or other complications. Alternatively, LOS may have been reduced by
other severity factors, which were not adjusted for by age, sex, and DRG weights.
The application had no effect on costs, but an effect may have been
present but obscured since all patients were included in the cost analyses.
In other words, restricting the analytic population to individuals prescribed
selected nephrotoxic or renally cleared medications might have allowed us
to show a difference. Regardless, the application itself is inexpensive to
implement within the context of an order-entry system, in contrast to other
prescription–quality-improvement programs, which generally have significant
labor costs and require ongoing expenditure or the effect wanes.
Our study has several limitations. First, the intervention and control
periods were not entirely analogous, since the number of admissions and the
hospital census were higher during the control periods. The higher census
may have prompted shorter LOS (in an effort to open beds), potentially decreasing
the relative effect of the application on LOS. Second, the calculation of
creatinine clearance by the Cockcroft-Gault formula may not accurately reflect
renal function under nonsteady-state conditions (ie, with increasing or decreasing
serum creatinine concentrations). In other words, the Cockcroft-Gault formula
may overestimate renal function when the serum creatinine is increasing, and
underestimate renal function when the serum creatinine is decreasing. However,
this misclassification should have affected individuals equally during the
intervention and control periods, and would tend to diminish the effect of
any intervention toward the null. Third, we did not consider the degree to
which individual orders differed from those considered optimal by the application's
definitions. In other words, we would have expected that dose-list modification
by the application would have led to a larger fraction of near-miss orders
during intervention periods, but due to the immense number of orders and resource
constraints, these were not calculated. Fourth, the program did not send notices
(pages or e-mails) to clinicians as soon as it had evidence of worsening renal
function, as did that of Rind et al,13 but
only alerted the clinician at the next occasion when the clinician was ordering
a medication. Finally, since the intervention was tested at a teaching hospital
where house officers write the majority of medication orders, the results
may not be generalizable to other, nonteaching hospital settings.
In summary, a computer order entry-based application to guide medication
dose and frequency choices for inpatients with renal insufficiency was tested
and resulted in a significant improvement in the appropriateness of drug prescription.
Provision of real-time advice in drug prescription may prove to be among the
most useful applications of medical informatics technology. Such applications
may provide clinicians "a better cockpit" and results in enhanced safety and
increased efficiency at minimal cost, with little intrusion into practice.
1.Hou SH, Bushinsky DA, Wish JB.
et al. Hospital-acquired renal insufficiency: a prospective study.
Am J Med.1983;74:243-248.Google Scholar 2.Levy EM, Viscoli CM, Horwitz RI. The effect of acute renal failure on mortality: a cohort analysis.
JAMA.1996;275:1489-1494.Google Scholar 3.Nolan CR, Anderson RJ. Hospital-acquired acute renal failure.
J Am Soc Nephrol.1998;9:710-718.Google Scholar 4.Obialo CI, Okonofua EC, Tayade AS, Riley LJ. Epidemiology of de novo acute renal failure in hospitalized African
Americans.
Arch Intern Med.2000;160:1309-1313.Google Scholar 5.Wang R, Mouliswar M, Denman S, Kleban M. Mortality of the institutionalized old-old hospitalized with congestive
heart failure.
Arch Intern Med.1998;158:2464-2468.Google Scholar 6.Dries DL, Exner DV, Domanski MJ.
et al. The prognostic implications of renal insufficiency in asymptomatic
and symptomatic patients with left ventricular systolic dysfunction.
J Am Coll Cardiol.2000;35:681-689.Google Scholar 7.Leape LL. Institute of Medicine medical error figures are not exaggerated.
JAMA.2000;284:95-97.Google Scholar 8.Bates DW, Cullen DJ, Laird N.
et al. Incidence of adverse drug events and potential adverse drug events.
JAMA.1995;274:29-34.Google Scholar 9.Bates DW, Leape LL, Cullen DJ.
et al. Effect of computerized physician order entry and a team intervention
on prevention of serious medication errors.
JAMA.1998;280:1311-1316.Google Scholar 10.Jha AK, Kuperman GJ, Teich JM.
et al. Identifying adverse drug events.
J Am Med Inform Assoc.1998;5:305-314.Google Scholar 11.Cockcroft DW, Gault MH. Prediction of creatinine clearance from serum creatinine.
Nephron.1976;16:31-41.Google Scholar 12.Edwards N, Honemann D, Burley D, Navarro M. Refinement of the Medicare diagnosis-related groups to incorporate
a measure of severity.
Health Care Financ Rev.1994;16:45-64.Google Scholar 13.Rind DM, Safran C, Phillips RS.
et al. Effect of computer-based alerts on the treatment and outcomes of hospitalized
patients.
Arch Intern Med.1994;154:1511-1517.Google Scholar 14.Evans RS, Pestotnik SL, Classen DC.
et al. A computer-assisted management program for antibiotics and other antiinfective
agents.
N Engl J Med.1998;338:232-238.Google Scholar 15.Nightingale PG, Adu D, Richards NT, Peters M. Implementation of rules based computerised bedside prescribing and
administration: intervention study.
BMJ.2000;320:750-753.Google Scholar