Keywords medical records' systems - computerized - medication reconciliation - patient safety
Background and Significance
Background and Significance
Transitions of care, defined as “movement of a patient from one setting of care to
another”,[1 ] are high-risk events. Important clinical information may be lost as the patient
transitions from one setting or provider to another.[2 ]
[3 ] Although adverse events may occur for many reasons, adverse drug events (ADEs) are
a common cause of morbidity and even mortality.[4 ] Unfortunately, medication discrepancies across care transitions are common.[5 ] To the extent that care transitions require that each medication must be accurately
entered, verified, represcribed, stopped, or changed, discrepancies are more likely
when there are multiple medications (polypharmacy) and when there are changes in medications.[6 ]
[7 ] Care transitions related to hospital admission and discharge are associated with
a high risk of inadvertent medication discontinuation—in both intensive care unit
(ICU) and non-ICU admissions.[8 ] Such errors of omission could increase risk of adverse events such as emergency
department visits, hospitalization, and death.[9 ]
Accurate and accessible medication records are required for safe clinical decision-making.
For example, Nirmatrelvir/Ritonavir (Paxlovid) for COVID-19 must be prescribed within
at most 5 days from the onset of symptoms but interacts with many commonly prescribed
drugs. Thus, to safely prescribe Paxlovid, the clinician must know the patient's current
medications. Medication reconciliation has been shown to reduce medication discrepancies
and ADEs,[10 ] and is required by various quality metrics and government programs.[11 ]
[12 ]
Ideally, medication lists that reflect medications that are active at a particular
point in time should match between institutions. Consider four medication lists: (1)
preadmission (outpatient), (2) at-admission (inpatient), (3) at-discharge (inpatient),
and (4) first posthospital visit (outpatient) list. Theoretically, the preadmission
and the at-admission medication lists should be similar, though discrepancies can
occur for a variety of reasons. For example, the patient could have seen a specialist
at a different institution between the last outpatient visit and admission. Similarly,
we expect the at-discharge list to be similar to the medication list at the first
follow-up visit; barring changes between these events. For example, the patient may
have completed a course of antibiotics prescribed at discharge. In contrast, we do
not expect the at-admission list to match the at-discharge list due to medication
changes that are likely to occur during hospitalization.
Previous studies have shown that discrepancies are highly prevalent among populations
of hypertensive patients.[13 ]
[14 ] We hypothesized that this kind of medication discrepancy is prevalent regardless
of disease type.
Objectives
Regardless of whether discrepancies are caused by errors, by health system fragmentation,
or other reasons, discrepancies between medication lists may increase ADE risk in
multiple ways. For example, errors of omission may occur if necessary medications
are inadvertently not continued and errors of commission may occur if unnecessary
medications are prescribed.[15 ] Therefore, we analyzed medication data across inpatient/outpatient care transitions
to quantify the discrepancies between medication records across care transitions and
across EHR systems.
Methods
We used data from two institutions in the southern United States that share patients
but are administratively separate with separate EHR systems. One institution, the
University of Texas Health Science Center at Houston (UTHSC-H) is an academic medical
center with an Epic EHR implementation that operates a network of outpatient clinics
but does not operate an inpatient facility (hospital). The second institution, Memorial
Hermann Hospital System is a Cerner site that is affiliated with the academic medical
center and operates both inpatient outpatient facilities. Patients often transition
between the two institutions. In this study, we analyzed inpatient admissions of patients
from the academic medical center to the hospital and their subsequent discharge back
to the outpatient clinics. This study has been approved by the Committee for the Protection
of Human Subjects (the UTHSC-H Institutional Review Board [IRB]) under protocol HSC-SBMI-13-0549.
We focused on the discrepancies in medication information between the two systems.
Specifically, we calculated the intersection (i.e., number of medications in common)
and the discrepancies between two lists. Discrepancies were calculated as follows:
Discrepancies = (no. of Medications on List A) + (no. of Medications on List B) − 2
× (no. of Medications on both List A and B)
For example, suppose list A contained aspirin, lisinopril, gabapentin and list B contained
lisinopril, gabapentin, tadalafil. The number of discrepancies would be: 3 + 3–2 × 2 = 2.
We used outpatient EHR data to identify adult patients (>18 years old) with an identified
primary care physician at the academic medical center who were seen in the outpatient
setting, admitted to hospital, and subsequently discharged home between January 1,
2022 and March 31, 2023. Patients with no primary care physician at the academic medical
center, who died, or who were transferred to other facilities, such as skilled nursing
facilities or acute rehabilitation, were excluded. [Fig. 1 ] provides an overview of our study.
Fig. 1 Overview. ToC, transitions of care.
We refer to discharges from the hospital as “discharge events” (DEs). DEs were the
unit of analysis because some patients were admitted more than once during the study
period. Active medications were assumed to be the medications that the patient was
taking at a particular point in time. Medication adherence (compliance) was outside
the scope of this study.
We used partially structured discharge summaries that were completed by the discharging
inpatient clinician (or designee) because structured medication data from the inpatient
facility were not available. The discharge summaries contained a partially structured
“discharge medications” section. The discharge medication section was further structured
into four categories of medications.
Unchanged : medications that were active both at-admission and at-discharge.
Changed : medications where some element was changed, such as dose or route of administration.
New : medications that were started during the hospitalization and were to remain active
postdischarge.
Discontinued : medications that were active before the hospitalization that the patient should
no longer take after discharge.
We compared the prescription start and end dates to the hospitalization dates to determine
the active medications at four specific times: preadmission, at-admission, at-discharge,
and at the first postdischarge encounter.
Preadmission : active medications before the admission date.
At-admission : all medications listed in the following discharge message categories—changed, unchanged,
and discontinued.
At-discharge : all medications listed in the following discharge message categories—new, changed,
and unchanged.
Postdischarge : encounter medications were defined as the active medications at the first encounter
after discharge. An encounter was defined as any recorded event that was not exclusively
financial or administrative.
One caveat to this grouping is that the “changed” category does not list what changed
in that medication order. The change could have been strength, frequency, or form
but that information was not provided in the discharge message.
We used two sections of the outpatient EHR (Epic) to define medication lists. The
first section was the active medication list and the second was the discharge summary.
The Active Medication list included the medications ordered and marked as active using
the metadata: start date, end date, cancel date. The discharge note in the outpatient
EHR (Epic) was received from the inpatient EHR (Cerner) and contained patient identifier,
discharging physician name, admission date, discharge date, and discharge note text.
To retrieve the discharge medications, the following steps were taken. The discharge
summaries were identified in the source inpatient database (Epic Clarity) system using
python scripts and inserted into a local database. We validated that the extracted
message matched what was seen in the outpatient EHR graphical user interface and in
the source inpatient database (Epic Clarity).
Discharge summaries were preprocessed by first identifying the location of the discharge
medication section and mediation categories: new, changed, unchanged, and discontinued.
These data were then processed using a Clinical Language Annotation, Modeling, and
Processing (CLAMP)[16 ] pipeline. CLAMP can parse clinical text and normalize the text using specific vocabularies.
For this project, we used a medication pipeline that identified medications and used
the RxNorm ontology for standardization. The output contained specific RxNorm concept
unique identifiers (CUIs) when dosage and route were specified and a generic RxNorm
CUI that was one level up in the RxNorm hierarchy.
The CLAMP output was processed and loaded into a relational database. After these
steps were taken, the active medication lists were imported for those patients who
had a DE. Using RxNav API,[17 ] the medications identified with RxNorm CUIs (we refer to this as RxNorm drug) were
further mapped to the ingredient (we refer to this as RxNorm ingredient). For multi-ingredient
formulations, all ingredients were compared as if they were separate medications.
For example, hydrochlorothiazide/lisinopril combination was treated as two medications:
hydrochlorothiazide and lisinopril. After ingredient classification, the four medication
lists were assembled using the admission and discharge times to identify medications
active before admission and after discharge. Dosages (e.g., 25 vs. 50 mg) and route
(e.g., oral vs. sublingual) were considered at the drug level, but not the more general
ingredient level. Frequency of administration was ignored. Notably, route and frequency
were often missing.
A random sample of 15 DEs were chosen for manual review and compared to the results
generated by the automated system. A clinical expert constructed the four lists independently
of the automated lists, extracting the medication information from the medication
and notes sections of the Epic user interface. The lists created by the clinical expert
were compared to the automatically constructed medication lists.
Statistical analysis was performed using Python 3.10 and R version 4.2.2. Descriptive
statistics such as mean, standard deviation, median, and interquartiles (Q1 = 25th
percentile, Q3 = 75th percentile) were reported. Wilcoxon's signed-rank test was used
to compare intersection mean percentage between two time points and p -values were reported. All statistical tests were two-sided and statistical significance
was defined at p < 0.05.
Results
Starting from 26,569 DEs and after applying the exclusion criteria shown in [Fig. 2 ], we analyzed 383 DEs. These 383 DEs were attributed to 353 patients (239 female
[68%], 114 male [32%]; 211 were aged between 18 and 64 [60%], 142 were aged >65 [40%]).
[Fig. 2 ] shows the breakdown of patients with single and multiple DEs. Notably two patients
had three DEs during the study period. [Fig. 3 ] shows the process of filtering and identifying DEs.
Fig. 2 Number of patients versus discharge events.
Fig. 3 CONSORT diagram of discharge events. CONSORT, Consolidated Standards of Reporting
Trials; DE, discharge event; PCP, primary care physician.
Term normalization to RxNorm drugs and subsequently to RxNorm ingredients are shown
in [Table 1 ]. As an example, albuterol sulfate is reconciled from four distinct RxNorm concepts
to one RxNorm Ingredient concept.
Table 1
Drug normalization example
Drug name
Strength
Form
RxNorm drug
RxNorm ingredient
Albuterol sulfate
108
HFA
142153
Albuterol
Albuterol sulfate
108
Puffs
745679
Albuterol
Albuterol sulfate
2.5 mg/3 mL
–
351136
Albuterol
Albuterol (Eqv-ProAir HFA)
90 μg/inh
Puffs
307782
Albuterol
Abbreviations: HFA, hydrofluoroalkane; inh, inhaler.
[Table 2 ] shows the number of medication entries in each of the four lists and in total. The
medication list comparison results are shown in [Table 3 ] using specific RxNorm drugs and more general RxNorm ingredients. Overall, the intersection
between lists from the same EHR system was substantially greater by all measures (mean,
median, percent of total medications on list) than the intersection between lists
from different EHR systems.
Table 2
Number of drugs and ingredients in each medication list
List
RxNorm Drug
N (mean, median, range, standard deviation)
RxNorm Ingredient
N (mean, median, range, standard deviation)
Preadmission
2,953 (7.71, 6, 0–33, SD = 6.78)
2,704 (7.07, 6, 0–28, SD = 6.11)
At-admission
2,736 (7.14, 5, 0–31, SD = 6.79)
2,254 (5.89, 4, 0–26, SD = 5.68)
At-discharge
2,978 (7.78, 7, 0–32, SD = 6.46)
2,430 (6.34, 5, 0–26, SD = 5.31)
Postdischarge
3,069 (8.01, 7, 0–33, SD = 7.12)
2,787 (7.27, 6, 0–31, SD = 6.36)
Total
11,736
10,175
Abbreviation: SD, standard deviation.
Table 3
Medication list comparison with respect to intersection (drugs in common)
RxNorm drug (range, SD)
RxNorm ingredient (range, SD)
Preadmission versus at-admission (across EHRs)
N = 383 DEs
Intersection mean
1.43 (0–10, 1.92)
2.96 (0–17, 3.52)
Intersection median
1
2
Intersection mean percentage
15.0% (0–100, 20.1)
33.9 (0–100, 34.0)
Intersection median percentage (Q1, Q3)
6.1% (0, 66.7)
29.4% (0.62)
Difference mean
11.99 (0–44, 9.22)
7.03 (0–28, 5.61)
Difference median
9
6
At-admission versus at-discharge
(same EHR)
N = 383 DEs
Intersection mean
6.33 (0–25, 6.07)
5.22, (0–23, 5.1)
Intersection median
5
4
Intersection percentage
71.3% (0–100, 39.8)
70.0% (0–100, 40.6)
Intersection median percentage (Q1, Q3)
94.1% (50, 100)
94.1% (44.6, 100)
Difference mean
2.26 (0–24, 2.87)
1.80 (0–14, 2.18)
Difference median
1
1
At-discharge versus postdischarge (across EHRs)
N = 376 DEs
Intersection mean
1.47, 0–9, 1.93
3.07, 0–17, 3.5
Intersection median
1
2
Intersection percentage
15.5% (0–100, 19.9)
37.2% (0–100, 34.9)
Intersection median percentage (Q1, Q3)
7.1% (0, 25.5)
35% (0, 66.7)
Difference Mean
13.14 (1–47, 9.02)
7.83, (1–30, 5.29)
Difference Median
11
7
Preadmission versus postdischarge (same EHR)
N = 376 DEs
Intersection mean
7.41, 0–33, 6.77
6.85, 0–28, 6.11
Intersection median
6
5
Intersection percentage
78.4% (0–100, 39.5)
78.8% (0–100, 39.5)
Intersection median percentage (Q1, Q3)
100% (87.1, 100)
100% (91.3, 100)
Difference mean
1.19, 0–16, 2.15
1.00, 0–28, 1.93
Difference median
0
0
Abbreviation: DE, discharge event.
Q1, 25th percentile; Q3, 75th percentile.
When using RxNorm drugs, the intersection percentages between two different time points
were examined using Wilcoxon's signed-rank test and p -values were reported in [Table 4 ]. The intersection percentage difference between the preadmission and at-admission
versus at-discharge and postdischarge were not statistically significant difference
(p = 0.79). However, all other list comparisons were significant.
Table 4
p -Values of pairwise comparison of intersection percentages between different time
points for pairs of medication lists
Time point
Pairwise comparison
RxNorm drug
RxNorm ingredient
Preadmission versus at-admission (across EHRs)
Ref
–
–
Ref
–
–
At-discharge versus postdischarge (across EHRs)
0.7903
Ref
–
0.012
Ref
–
At-admission versus at-discharge (same EHR)
<0.0001
<0.0001
Ref
<0.0001
<0.0001
Ref
Preadmission versus postdischarge (same EHR)
<0.0001
<0.0001
0.002
<0.0001
<0.0001
0.0005
Abbreviation: Ref, reference.
When using RxNorm ingredients, all comparisons were statistically significant.
Discharge message timeliness was assumed to be important so the time between discharge
and receiving the discharge message was also analyzed. We found a mean of 14.4 hours
and a maximum of 58 days.
To confirm the accuracy of the automated results, a manual review was done on 15 DEs
by an author with no prior knowledge regarding the system output for these 15 DEs.
These DEs contained a total of 455 medication records; one DE was not reviewed due
to restricted access to the patient record. There were a total of 42 discrepancies
(9.2%) between the manual and automated results. Twenty discrepancies (4.4%) were
related to nonmedication orderable supplies such as lancets, blood glucose monitoring
supplies, or diabetic test strips that had no RxNorm identifier. In 16 cases (3.5%),
CLAMP did not identify the text string correctly. Three discrepancies were caused
by reviewer error (0.7%). Finally, three (0.7%) had date-related artifacts that caused
an incorrect postdischarge encounter, this was associated with one of the fifteen
DEs.
Discussion
We found substantial discrepancies between medication lists associated with transitions
of care between outpatient and inpatient settings. These discrepancies were more associated
with transitions across EHR systems than clinical changes. For example, the difference
between outpatient (preadmission) and at-admission medication lists (different EHR
systems) was larger than the difference between at-admission and at-discharge lists
(same EHR system).
Strengths of our study included reliance on real-world data from two collaborating
institutions. We developed an automated system that allowed us to review over 10,000
individual ingredient entries, more than what would be practical to accomplish manually.
Manual validation of the automated system performed without prior knowledge of the
automated results increases confidence in the accuracy of the automated system.
At our institution, a group of community health workers and case managers at a transitions
of care “hub” review the four lists produced by our system to identify patients at
risk for ADEs. Patients deemed at risk such as those with 10 or more medications (polypharmacy),
high-risk medications (e.g., diphenhydramine in an elderly patient), or multiple new
medications are prioritized for human follow-up by a nurse case manager. Without the
medication extraction system, the case manager had to review the semistructured discharge
document, the current (postdischarge) medication list, manually reconstruct the preadmission
medication list, and then manually compare these lists. Anecdotally, this requires
approximately 10 to 20 minutes per case. An added practical benefit is that the system
described in this paper integrates information that previously required access to
multiple systems including the outpatient EHR and payer portals to identify recently
discharged patients.
Our study was limited to two collaborating institutions and reflects the clinical
workflow at these institutions. Clinical workflows, including medication reconciliation
practices, may be different in different settings and for different care transitions
(e.g., discharge to acute rehabilitation). Similarly, we restricted our analysis to
a general adult population. Results may be different for other populations such as
pediatric, geriatric, or obstetrical populations and in other geographic areas (e.g.,
rural).
Another limitation is that our data did not allow us to determine whether discrepancies
were errors or to attribute clinical consequences to discrepancies. As discussed above,
discrepancies could have occurred due to health system fragmentation (e.g., seeing
an outside specialist who changed medications between the last outpatient visit and
admission) or multiple other factors. Finally, our data did not allow us to assess
adherence (compliance). Some discrepancies could have been attributable to discrepancies
between prescribed medications and the medications that the patient was actually taking
at a particular point in time (e.g., at-admission) or what the patient reported to
the admitting clinician.
Previous studies have shown that it is possible to extract knowledge from unstructured
data.[16 ]
[18 ]
[19 ] Identification of medications from discharge summaries is not a new challenge and
we were able to leverage existing tools like CLAMP[16 ] and RxNorm resources maintained by the National Library of Medicine. A common classification
approach is to rely on the main ingredient of a combination.[18 ]
[19 ] A limitation of this approach is that ignoring some ingredients can lead to misclassification.
In our study, discrepancies between automated and manual medication identification
accounted for 9.2% of the manually reviewed data with only half of these attributable
to the automation; the rest were not related to medications (e.g., diabetic supplies),
date errors in the medical record or reviewer error. A substantial proportion of the
remaining errors were attributable to incorrect templates used in the notes system,
or a missing medication section.
Correct and current medication information is important for clinical performance measures[12 ] and patient safety[20 ]; particularly at care transitions. We found substantial discrepancies between medication
lists, particularly across EHR systems. This is not a new finding, nearly 20 years
ago, researchers noted that most medication lists were not accurate.[21 ] Our findings are similar to prior studies showing medication list intersections
between transitions of care from 30 to 80%.[22 ]
[23 ] Unfortunately, to the extent that discrepancies correlate with inaccuracies, regardless
of whether the discrepancy was due to an error or resulted from system-level properties
that inhibit routine, automated, and accurate information sharing, progress over the
past 20 years has been limited.
We found multiple instances of repeated medications that had different RxNorm drugs
in both structured data (outpatient) and discharge notes (inpatient). For example,
a list may contain two entries for Lisinopril “1: Lisinopril 10 mg po qd” and “Lisinopril,”
resulting in two RxNorm drugs. This accounts for some of the discrepancies in the
medication lists using RxNorm drugs. The comparison using RxNorm ingredient combines
entities with duplicate medications, brand names, and different formularies between
institutions (see [Table 3 ]), at the cost of abstracting away information about different formulations (e.g.,
liquid vs. pill form of the same drug). Another advantage of comparing ingredients
rather than RxNorm drugs is that supplies that are not likely to cause ADEs, such
as insulin syringes, were not identified as medications and thus not considered discrepancies.
We recognize that there could have been cases where a patient runs out of supplies
and is not able take a necessary medication (e.g., insulin).
Discrepancies in medication lists do not necessarily represent errors by either patient
or clinician. For example, a patient could have seen an outside clinician who changed
the medication regimen between the last preadmission visit and admission. Similarly,
administrative discrepancies such as inpatient formularies and outpatient insurance
coverage may require medication changes (e.g., therapeutic substitution[24 ]). Thus, we did not expect that medication lists between the two institutions would
be identical. However, quantifying these discrepancies can inform the design and implementation
of systems to support safe clinical decision-making.
Unfortunately, most clinics lack EHR functionality to support automated medication
reconciliation.[25 ] Examples of automated systems that may decrease the manual work required for medication
reconciliation include tools to (1) accurately extract medication lists from routinely
collected health data, (2) present these lists in an interface that makes comparison
simple[26 ]
[27 ] (e.g., displays both lists on the same screen), and (3) highlight discrepancies
between lists of medications.[28 ]
[29 ]
Medication lists within the same EHR (e.g., at-admission and at-discharge) tend to
be more similar than lists in different EHRs (e.g., preadmission and at-admission).
This finding highlights the importance of EHR interoperability for institutions that
share patients. Previous studies of interoperability found that implementations of
the same vendor tend to be somewhat more interoperable than implementations of products
from different vendors.[30 ] Standardizing representations, such as using the same vocabulary to represent medications
in both systems and sharing these as structured data rather than text notes may improve
concordance of medication lists across care transitions.
Conclusion
Active medication lists in the same EHR system are similar to each other, but there
are many discrepancies between active medication lists in different EHR systems. There
is opportunity to improve medication reconciliation. An automated system, such as
the one that we developed, implemented, and validated can identify medication list
discrepancies to be addressed within a transitions of care program.
Clinical Relevance Statement
Clinical Relevance Statement
Accurate automated identification of medication lists preadmission, at-admission,
at-discharge, and at first encounter postdischarge is possible. These lists can support
automated clinical decision support (e.g., drug interaction detection), support transitions
of care services (e.g., help community health workers and nurse case managers to identify
patients at risk of ADEs). Since transitions of care often involve transitions between
clinical systems (e.g., different institutional EHRs), interoperability between clinical
systems is important for patient safety across care transitions.
Multiple Choice Questions
Multiple Choice Questions
What medication lists were most concordant?
a) At-admission and at-discharge
b) Preadmission and at-admission
c) Preadmission and postdischarge
d) At-admission and postdischarge
Correct Answer: The correct answer is option c. The mean intersection was 7.41 (highest) and the
median difference was 0 meaning that these two lists were most similar.
What standard was used to compare medication lists in this study?
Correct Answer: The correct answer is option c. RxNorm ingredients were used to compared medication
lists in this paper. NDC codes were not used because they reflect packaging (e.g.,
30 tablet package vs. 90 tablet package) that are not necessarily clinically significant.
ICD10 is a controlled vocabulary of diseases/conditions and CPT is a controlled vocabulary
of procedures.
