Medical decision support: experience with implementing the Arden Syntax at the Columbia-Presbyterian Medical Center

Abstract

We began implementation of a medical decision support system (MDSS) at the Columbia-Presbyterian Medical Center (CPMC) using the Arden Syntax in 1992. The Clinical Event Monitor which executes the Medical Logic Modules (MLMs) runs on a mainframe computer. Data are stored in a relational database and accessed via PL/I programs known as Data Access Modules (DAMs). Currently we have 18 clinical, 12 research and 10 administrative MLMs. On average, the clinical MLMs generate 50357 simple interpretations of laboratory data and 1080 alerts each month. The number of alerts actually read varies by subject of the MLM from 32.4% to 73.5%. Most simple interpretations are not read at all. A significant problem of MLMs is maintenance, and changes in laboratory testing and message output can impair MLM execution significantly. We are now using relational database technology and coded MLM output to study the process outcome of our MDSS.

Full text

Medical
Decision
Support:
Experience
with
Implementing
the
Arden
Syntax
at
the
Columbia-Presbyterian
Medical
Center
Robert
A.
Jenders,
MD,
MS*t;
George
Hripcsak,
MD*;
Robert
V.
Sideli,
MD*;
William
DuMouchel,
PhD*T;
Haiying
Zhang,
MSI;
James
J.
Cimino,
MD*t;
Stephen
B.
Johnson,
PhD*;
Eric
H.
Sherman,
MD,
MS*t;
Paul
D.
Clayton,
phD*t+
Departments
of
Medical
Informatics*,
Medicinet,
Biostatistics$
and
Radiology+
Columbia
University,
New
York,
New
York
lfie
began
implementalion
of
a
mtiedical
decision
support
systein
(AIDSS,)
at
the
Columtbia-
Presbyterian
Aledical
Center
(
AIQIC)
using
the
lrden
Syntax
in
1992.
7he
('linical
Ivent
Monitor
which
executes
ihe
Aledical
Logic
Ak.odules
(MLAfs)
runs
on
a
miainframle
computer.
Data
are
stored
in
a
relational
database
and
accessed
Wia
PLi
programns
known
as
Dala
Access
Alodules
(DL,AAJIs).
Currently
we
have
18
clinical,
12
research
and
10
administrative
AILAIs.
On
average,
the
clinical
HAIMs
generate
50357
simple
interpretalions
of
laboratory
data
ancd
1080
alerts
each
inonth.
The
numiber
of
alerts
actually
read
'aries
by
subject
of
the
AILA1
frovn
32.4%
to
73.5%.
Alost
simnple
interpretations
are
not
read
at
all.
A
significant
problein
o?f
AAlMs
is
m)aintenance,
and
changes
in
laboralory
testing
and
mtiessage
output
can
imipair
AMLMA
execution
significanlvy.
IWe
are
now
using
relational
database
technologv
and
coded
MLM
output
to
sudvy
the
process
outcome
of
our
ADSS.
INTRODUCTION
The
integration
of
clinical
guidelines
into
medical
practice
has
becn
promoted
by
the
development
of
computcr-based
reminder
systems.'
Strong
evidence
suggcsts
that
some
mcdical
decision
support
systems
(MDSS)
can
improve
physician
performance.2
The
Ardcn
Syntax
for
Mcdical
Logic
Modules
(MLM),
developcd
in
part
at
Columbia-
Presbyterian
Mcdical
Ccnter
(CPMC).
has
been
promoted
as
an
opcn
standard
for
the
procedural
represcntation
and
sharing
of
mcdical
knowledge.3
The
CPMC
is
a
1477-bed
centcr.
In
1994
it
recorded
415256
outpaticnt
clinic
visits
and
53374
discharges.
As
part
of
our
ongoing
dcvclopmcnt
of
a
MDSS
at
CPMC.
wc
have
bcen
devcloping
MLMs
which
serve
the
needs
of
clinical
uscrs,
research
scicntists
and
hospital
administrativc
pcrsonnel.
We
begin
by
discussing
the
architecture
of
the
CPMC
MDSS
in
which
these
MLMs
residc
and
the
software
tools
used
to
develop
them.
We
then
discuss
the
MLMs
actually
implemented
at
CPMC
and
the
quantitative
operation
of
those
intended
for
clinical
use.
In
reviewing
our
experience
in
this
development
process,
wc
illustrate
the
benefits
and
problems
associated
with
the
implementation
of
our
MOSS.
Finally,
we
present
an
overview
of
our
future
work
in
studying
the
MDS
and
its
impact
on
the
process
and
outcome
of
medical
care.
MDSS
ARCHITECTURE
The
MDSS
at
CPMC
employs
a
compiler-interpreter
pair
to
crcate
executable
code
from
the
original
syntax
of
the
MLM.4
Although
a
MLM
can
be
created
in
this
system
using
any
text
editor,
we
primarily
employ
a
workstation
with
a
sophisticated
graphical
user
interface.5
In
particular,
one
of
us
has
written
an
editor
for
MLMs
employing
Asterx
running
under
X-Windows
on
a
RS/6000
machine
using
Unix.
MLMs
so
created
are
compiled
into
pseudocodes
using
a
software
program
built
with
lex
and
yacc.
Thcse
compiled
routines
then
are
uploaded
to
the
main
data
repository
of
CPMC,
a
3090
IBM
mainframe,
where
they
eventually
execute.
Laboratory,
demographic,
admission-discharge-
transfcr
(ADT)
and
inpatient
pharmacy
data
are
availablc
for
qucry
by
the
MLMs.
These
data
are
stored
in
a
relational
format
in
DB2.
Data
access
modulcs
(DAMs)
are
PL/I
programs
which
execute
on
the
mainframe
and
allow
the
MLMs
to
query
the
database
without
requiring
the
MLM
author
to
know
the
exact
relational
schema.
Parameters
passed
from
the
rcad
statement
in
a
MLM
to
the
DAM
specify
the
type
and
identity
of
the
data
elements
required
and
various
constraints
on
the
data
such
as
inpatient
status.
0195-4210/95/$5.00
C
1995
AMIA,
Inc.
169

Data
stored
in
DB2
arc
encoded
using
the
Medical
Entities
Dictionary
dcvclopcd
at
CPMC.6
Each
MLM
can
havc
onc
or
iiiorc
triggcrs
which
identify
when
the
Clinical
Event
Monitor,
which
serves
as
the
infercncc
cnginc
for
the
MDSS,
should
execute
the
MLM.
Thcsc
triggcrs
arc
coded
as
pairs
of
MED
codcs.
onc
of
which
idcntifics
the
typc
of
data
(such
as
a
pharmacy
ordcr)
and
the
other
which
signifies
the
identity
of
the
actual
data
clcmcnt
whosc
storagc
in
DB2
should
firc
the
applicablc
MLM.
The
type
of
data
elemcnt
is
spccificd
in
the
trigger
becausc
the
data
modcl
and
the
rclational
tablcs
arc
partitioned
in
this
manncr.
In
addition
to
storing
paticnt
data
using
MED
codes,
the
output
of
a
MLM
can
bc
written
using
MED
codes.
In
most
currently
implcmcntcd
MLMs,
howevcr,
the
interprctation,
alen
or
clectronic
mail
messagc
is
recorded
in
DB2
as
a
decision
event
whose
valuc
is
the
text
string
of
the
messagc.
The
application
program
which
displays
patient
data
for
clinical
users
thcn
qucrics
the
decision
cvent
tables
to
display
appropriatc
information
regarding
a
panicular
paticnt.
TYPES
OF
MLM
The
MLMs
currcntly
implcmcntcd
in
our
MDSS
can
be
divided
into
thrce
classcs
based
on
the
intended
audience
of
their
output.
The
largest
of
thcse
classes
arc
the
clinical
MLMs.
Thcsc
includc
alerts
which
identify
the
prescncc
of
abnormal
data
or
potentially
scrious
combinations
of
data
clcmcnts
in
the
paticnt
record.
In
addition,
somc
clinical
MLMs
provide
simplc
intcrprctations
of
laboratory
data,
such
as
calculating
an
anion
gap
or
a
calcium
lcvcl
corrected
for
albumin
lcvel
from
chcmistry
panel
results.
Thcse
intcrprctations
do
not
providc
any
advice
for
funhcr
cvaluation
or
trcatmcnt;
thcy
mercly
providc
simplc
calculations
derived
from
numbers
in
a
test
result.
The
second
most
numcrous
class
of
MLMs
is
the
rescarch
MLM.
Gencrally
thesc
provide
notice
to
clinical
investigators
via
clcctronic
mail
when
paticnts
meeting
ccrtain
characteristics
arc
admitted
to
the
hospital.
In
addition,
these
MLMs
can
notify
researchers
when
a
patient
already
cnrolled
in
a
study
has
been
admitted
to
the
hospital
so
that
appropriate
follow-up
can
occur.
The
third
principal
class
of
MLMs
are
those
intended
for
use
by
the
hospital
administrators.
sometimes
called
Administrative
Logic
Modules
(ALMs).
These
MLMs
generally
notify
administrators
via
electronic
mail
when
a
particular
condition
has
occurred
so
that
corrective
action
can
be
taken
in
order
to
reduce
hospital
expenses.
Examples
of
such
MLMs
include
notice
of
inappropriate
bed
classification
and
live
discharge
from
the
hospital
on
the
same
day
of
admission.
Also,
this
class
of
MLMs
provides
a
service
to
affiliated
community
physicians
by
providing
notice
via
fax
or
electronic
mail
of
admissions
or
discharges
of
their
patients
from
the
hospital.
MLMs
also
can
bc
classified
by
the
kind
of
output
they
produce.
Using
this
criterion,
current
MLMs
can
be
classified
as
alert,
interpretation,
or
electronic
mail
events.
Most
clinical
MLMs
are
alerts,
while
most
of
the
remaining
MLMs
produce
output
in
the
form
of
clectronic
mail.
We
are
just
now
creating
MLMs
which
produce
output
in
the
form
of
coded
messages
which
can
be
used
to
represent
intcrmcdiate
decision
states
in
a
complicated
clinical
guideline.
IMPLEMENTED
MLMS
We
currently
have
40
MLMs
in
production.
Eighteen
of
these
are
clinical
MLMs,
including
four
interpretations
and
fourteen
alerts
(Table
1).
The
alert
MLMs
produce
a
summary
average
of
1080
distinct
alerts
each
month,
while
the
interpretation
MLMs
yield
a
summary
average
of
50357
events
each
month.
Many
more
interpretations
than
alerts
are
produced
because
the
former
occur
primarily
on
all
valid
chcmistry
panels,
while
the
latter
occur
only
for
specific
abnormal
situations
typically
involving
combinations
of
data
elements.
Although
eighteen
distinct
clinical
MLMs
exist,
these
may
be
grouped
into
several
classes
which
act
upon
one
kind
of
abnormal
event.
This
occurs
because
distinct
MLMs
are
created
for
triggers
involving
distinct
kinds
of
events.
For
example,
three
MLMs
arc
used
to
alert
users
to
the
presence
of
hypokalemia
and
digoxin
use
which
might
lead
to
cardiac
dysrhythmia.
One
MLM
is
triggered
by
the
storage
of
a
pharmacy
order
for
digoxin;
another
MLM
is
triggered
by
the
storage
of
a
blood
potassium
result;
and
the
third
MLM
in
the
suite
is
triggered
by
the
storage
of
a
blood
digoxin
level.
Thus,
threc
different
triggers
with
three
different
queries
are
used
to
encode
this
example,
although
similar
logic
is
used
in
each
of
the
MLMs.
170

Tablc
1.
Operation
statistics
on
clinical
MLMs.
STANDARD
DEV
is
the
standard
deviation.
TB
is
infection
with
tuberculosis.
NSAIDS
are
nonsteroidal
anti-inflammatory
drugs.
HBV
is
Hcpatitis
B
Virus.
Chem
7
and
chem
20
are
7-
and
20-elcment
chemistry
panels,
respectively.
In
addition
to
thesc
clinical
MLMs,
we
currently
have
twelvc
rcscarch
MLMs
in
production.
Examples
of
these
includc
identifying
patients
with
abnormal
ccrvical
pathology,
paticnts
with
a
MB
fraction
of
crcatinc
kinasc
positivc
for
myocardial
infarction,
paticnts
with
a
discharge
diagnosis
of
lymphosarcoma
and
paticnts
with
hyperamylasemia
suggestivc
of
pancrcatitis.
Thcse
notify
the
appropriatc
rcscarchers
of
the
mcdical
record
number
and,
when
appropriatc
the
inpatient
location
of
a
paticnt
in
order
to
providc
the
rcscarchers
with
the
opportunity
to
cnroll
the
paticnt
in
a
study.
Other
MLMs
also
notify
rcscarchcrs
when
a
study
paticnt
is
admitted
to
the
hospital.
Finally
our
MDSS
includes
ten
administrative
MLMs.
Examples
of
these
are
notifying
community
physicians
when
their
patients
are
admitted
and
discharged,
notifying
administrators
of
live
discharge
from
the
intensive
care
unit
to
outpatient
status,
and
identification
of
patients
whose
payment
status
is
self-pay.
USER
INTERACTION
WITH
DECISION
EVENTS
Clinical
alerts
for
each
patient
may
be
displayed
on
the
hospital
information
system
at
the
option
of
the
individual
user
while
looking
at
a
particular
patient's
data.
This
option
is
available
as
a
choice
on
a
menu
171
SUBJECT
TYPE
NUMBER
NUMBER
TOTAL
AVERAGE
STANDARD
OF
MLMs
OF
EVENTS
EVENTS
DEV
MONTHS
PER
MONTH
positive
TB
alert
1
32
1098
34.3
11.9
culture
creatinine
interpretation
1
17
69689
4099.4
3925.9
clcarance
fractional
intcrprctation
1
25
3110
124.4
50.25
excretion
of
sodium
_ _ _ _ _
_ _ _ _
_ _ _ _
_ _ _ _
ncwborn
with
alcrt
2
36
450
12.5
7.6
mother
HBV
positivc
hypokalemia
alcrt
3
37
3224
87.1
23.9
and
digoxin
usc
crcatininc
rise
alert
1
30
7623
254.1
126.8
chem
7
interprctation
1
21
415724
19796.4
5378.6
chcm
20
intcrprctation
1
21
553084
26337.3
5084.8
hypokalemia
alcrt
2
3
198
66
16.6
and
diuretic
u
s
renal
failure
alert
2
3
168
56
20.7
and
aminoglycosi-
de
use
new
anemia
alert
1
3
1292
430.7
289.7
renal
failure
alert
2
3
417
139
96.0
and
NSAID
u
s
e _ _ _ _ _ _
_ _ _ _ _ _
_ _ _ _ _ _ _
_ _ _ _ _ _
_ _ _ _ _ _

Table
2.
User
interaction
with
clinical
MLMs.
These
data
concern
only
those
events
which
occurred
after
the
institution
of
logging
of
users
viewing
interpretations
and
alerts.
Average
Viewings
Per
Event
applies
only
to
those
events
(interpretations
or
alcerts)
which
were
viewed
by
at
least
one
user.
which
also
indicatcs
the
datc
of
the
most
rcccnt
alert.
If
at
Icast
onc
alert
for
a
given
paticnt
has
not
been
viewed
by
that
particular
user,
a
special
message
appears
informing
the
uscr
of
that
fact.
If
the
uscr
chooscs
the
mcnu
option
to
view
the
alerts,
he
is
then
prcscntcd
with
a
list
of
the
titles
of
the
alerts
in
rcvcrsc
chronological
order.
The
user
thcn
may
choosc
an
individual
alcrt
from
the
list
in
order
to
rcad
its
text.
Also,
interprctations
of
certain
laboratory
tests
may
be
vicwcd
by
choosing
an
option
while
reading
the
rcsult
of
a
tCst
for
which
an
intcrprctation
is
availablc.
Data
arc
availablc
on
which
user
rcad
the
text
of
which
alcrt
on
which
paticnt
since
January,
1994.
Thcse
data
are
summarized
in
Table
2.
The
total
number
of
events
are
thosc
alcrts
and
interpretations
which
occurred
after
the
initiation
of
the
logging
of
displayed
alerts.
The
number
of
alerts
or
intcrprctations
whose
text
actually
was
viewed
and
the
perccntagc
of
the
total
number
of
such
events
that
thcsc
vicwings
constitutc
also
are
indicated.
In
addition,
the
average
number
of
viewings
for
those
alerts
actually
viewed
is
delineated.
DISCUSSION
The
relatively
small
percentage
of
viewings
of
particular
MLM
subjects
may
be
explained
by
the
manner
in
which
alerts
are
displayed.
If
users
choose
the
option
to
display
alerts
while
viewing
a
patient's
results,
they
first
are
given
a
list
of
alerts
for
that
patient
consisting
only
of
the
title
of
the
alert.
The
title
alone
may
be
a
sufficient
alert
for
some
users,
who
then
do
not
read
the
text
of
the
alert.
The
data
in
this
paper
do
not
include
viewings
of
the
title
of
an
alert
without
reading
its
text.
Also,
maintenance
is
a
key
issue
in
providing
clinical
and
rescarch
system
users
with
reliable
service.
One
event
which
highlighted
the
importance
of
maintenance
was
the
change
in
laboratory
computer
systems
in
Summer,
1994.
This
caused
the
introduction
of
a
large
number
of
new
MED
codes
for
the
new
laboratory
tests
encoded
by
the
new
laboratory
software.
As
a
result,
the
MLMs
172
SUBJECT
NUMBER
NUMBER
NUMBER
OF
AVERAGE
PERCENT
OF
OF
EVENTS
OF
VIEWINGS
VIEWINGS
PER
EVENTS
EVENTS
EVENT
VIEWED
VIEWED
positivc
TB
culture
257
189
1416
7.5
73.5
crcatinine
clearancc
60889
1121
1436
1.3
1.8
fractional
cxcrction
1663
530
993
1.9
31.9
of
sodium
newborn
with
mother
209
115
219
1.9
55.0
HBV
sitive
__
hypokalcmia
and
1202
691
1425
2.1
57.5
digoxin
use
c
crcatininc
risc
1549
977
3048
3.1
63.1
chcm
7
233334
3017
3391
1.1
1.3
chem
20
361992
4847
5309
1.1
1.3
hypokalemia
and
198
95
154
1.6
48.0
diuretic
usc
renal
failure
and
168
70
109
1.6
41.7
aminoglycosidc
use
new
anemia
1292
419
864
2.1
32.4
renal
failurc
and
417
145
262
1.8
34.8
NSAID
usc
lI
I
lI_
_L

then
in
cxistencc
would
work
with
the
few
legacy
laboratory
tests
encoded
with
old
MED
codes
but
often
failcd
to
function
with
newly
generated
laboratory
data.
Clinical
uscrs
and
rcscarchers,
who
had
becomc
accustomcd
to
the
alcerts,
interprctations
and
notifications
providcd
by
the
MDSS
noted
their
absencc.
This
problem
helped
to
produce
the
rclatively
high
standard
deviations
seen
with
the
clinical
MLMs.
This
problem
was
solved
by
making
use
of
class
information
cncoded
in
the
MED.
In
the
MED
hicrarchy.
many
laboratory
tests
are
grouped
into
classcs
which
idcntify
rclatcd
kinds
of
tests.
Instead
of
using
the
code
of
spccific
tests
in
the
MLMs,
we
adaptcd
thcm
to
cmploy
the
code
of
the
class.
The
DAMs
wcrc
adaptcd
furthcr
to
allow
qucry
by
class.
This
allows
futurc
laboratory
tCsts
to
be
added
to
the
rclevant
classcs
without
disrupting
the
operation
of
the
MLMs.
Anothcr
problem
with
maintcnancc
occurrcd
when
the
laboratory
without
prior
notification
changed
the
format
of
certain
tcxtual
rcsults.
MLMs
which
looked
for
specific
strings
no
longer
functioned
properly.
A
statistical
monitoring
system
notifiecd
MLM
devclopcrs
that
ccrtain
MLMs
had
stopped
functioning,
and
a
rcview
of
the
new
laboratory
output
allowed
repair
of
the
rclcvant
MLMs.
FUTURE
WORK
All
of
the
MLMs
used
in
production
to
date
have
produced
output
in
the
form
of
decision
cvents
whose
value
arc
unconstrained
tcxt
strings
which
then
are
displayed
to
the
uscr
of
the
clinical
information
system
either
via
an
alen.
an
interpretation
or
clectronic
mail.
In
order
to
represent
the
intcrmcdiatc
decision
stcps
of
a
complex
clinical
guidclinc
wc
arc
devcloping
MLMs
which
write
decision
cvents
whosc
valucs
are
MED
codes.
This
cnablcs
MLMs
to
cxccutc
at
an
indetcrminate
future
timc
using
the
alrcady-calculated
result
of
other
MLMs.
Writing
coded
dccision
valucs
allows
as-yet
unspccificd
MLMs
to
cxecute
using
previously
produced
decisions
at
an
indeterminatc
future
time
point.
In
addition
wc
currcntly
arc
studying
the
impact
of
alces
on
Icngth
of
stay
and
various
clinical
outcomes.
SUMMARY
The
CPMC
MDSS
includes
40
MLMs
which
execute
on
a
mainframe
computer
and
access
patient
data
stored
in
a
relational
database
via
PL/1-encoded
DAMs.
Eighteen
clinical
MLMs
provide
a
large
number
of
alerts
and
data
interpretation,
while
research
and
administrative
M:LMs
provide
study
enrollment
screening
and
warnings
regarding
cost
to
these
different
classes
of
users.
Maintenance
is
a
key
issue
in
developing
the
MDSS,
and
using
the
MED
to
trigger
and
query
by
class
has
helped
ease
this
process.
We
now
are
using
coded
MLM
output
to
help
study
the
impact
of
our
MLMs.
References
1.
Jenders
RA,
Morgan
M,
Barnett
GO.
Use
of
open
standards
to
implement
health
maintenance
guidelines
in
a
clinical
workstation.
Comput
Biol
Med
1994;
24:385-390.
2.
Johnston
ME,
Langton
KB,
Haynes
RB,
Mathieu
A.
Effects
of
computer-based
clinical
decision
support
systems
on
clinician
performance
and
patient
outcome.
Ann
Intern
Med
1994;
120:135-
142.
3.
Hripcsak
G,
Clayton
PD,
Pryor
TA,
Haug
P,
Wigertz
OB,
Van
der
lei
J.
The
Arden
Syntax
for
medical
logic
modules.
Editor
Miller
RA.
Proceedings
of
the
Fourteenth
Annual
Symposium
on
Computer
Applications
in
Medical
Care.
New
York:
IEEE
Computer
Press,
1990;
200-204.
4.
Hripcsak
G,
Cimino
JJ,
Johnson
SB,
Clayton
PD.
The
Columbia-Presbyterian
Medical
Center
decision-support
system
as
a
model
for
implementing
the
Arden
Syntax.
Proceedings
of
the
Fifteenth
Annual
Symposium
on
Computer
Applications
in
Mcdical
Care.
Editor
Clayton
PD.
New
York:
McGraw-Hill,
1992;
248-252.
5.
Hripcsak
G,
Clayton
PD,
Cimino
JJ,
Johnson
SB,
Friedman
C.
Medical
decision
support
at
Columbia-
Presbyterian
Medical
Center.
In
Software
Engineering
in
Medical
Informatics.
Editors
Timmers
T,
Blum
BI.
Amsterdam:
Elsevier
Science
1991;
471-479.
6.
Cimino
JJ.
Clayton
PD.
Coping
with
changing
controlled
vocabularics.
Proceedings
of
the
Eightoenth
Annual
Symposium
on
Computer
Applications
in
Medical
Care.
Editor
Ozbolt
JG.
Philadclphia:
Hanley
&
Belfus,
1994;
135-139.
173