The APACHE III prognostic system.

Risk prediction of hospital mortality for critically ill hospitalized adults

WA Knaus, DP Wagner, EA Draper, JE Zimmerman, M Bergner, PG Bastos, CA Sirio, DJ Murphy, T Lotring and A Damiano Chest 1991;100;1619-1636 DOI 10.1378/chest.100.6.1619

The online version of this article, along with updated information and services can be found online on the World Wide Web at: http://chestjournal.org/cgi/content/abstract/100/6/1619

CHEST is the official journal of the American College of Chest Physicians. It has been published monthly since 1935. Copyright 2007 by the American College of Chest Physicians, 3300 Dundee Road, Northbrook IL 60062. All rights reserved. No part of this article or PDF may be reproduced or distributed without the prior written permission of the copyright holder (http://www.chestjournal.org/misc/reprints.shtml). ISSN: 0012-3692.

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

fiEl

.:i
-

clinical investigations in critical care

The APACHE PHPrognostic
Mortality
Ph.D.; M.D.; Carl A. Sirio, M.D.; MS.; and Anne Damiano, M.D.; MS.;

System*
for Critically

Risk Prediction of Hospital Ill Hospitalized Adults

William Elizabeth Marilyn A. Knaus, A. Draper, Bergner, M.D.; M.S.; Ph.D.; Douglas Jack Paulo P Wagner,

E. Zimmerman, G. Bastos,

Donaldj Murphy, M.D.; Ted Lotring, Frank E. HarrellJr., Ph.D.

The

objective

of

this

study

was

to

refine

the

APACHE

location

immediately

prior

to ICU

admission

to provide

risk

(Acute Physiology, Age, Chronic Health Evaluation) methodology in order to more accurately predict hospital mortality risk for critically ill hospitalized adults. We prospectively collected data on 17,440 unselected adult medical! surgical intensive care unit (ICU) admissions at 40 US hospitals (14 volunteer tertiary-care institutions and 26
hospitals services randomly nationwide). chosen to represent intensive care between We analyzed the relationship

estimates for hospital mortality for individual ICU patients. A five-point increase in APACHE ifi score (range, 0 to 299) is independently associated with a statistically significant increase in the relative risk of hospital death (odds ratio,
1. 10 to 1.78) within each of 78 major medical and surgical

the patients likelihood of surviving to hospital discharge and the following predictive variables: major medical and surgical disease categories, acute physiologic abnormalities, age, preexisting functional limitations, major comorbidities,

disease categories. The overall predictive accuracy of the first-day APACHE ifi equation was such that, within 24 h of ICU admission, 95 percent of ICU admissions could be given a risk estimate for hospital death that was within 3 percent of that actually observed (r 0.41; receiver operating characteristic 0.90). Recording changes in the
APACHE ifi score

on each subsequent

day of ICU

therapy

and
The

treatment APACHE

location

immediately

prior

to ICU

admission. options:

ifi
for

prognostic

system

consists

of two
patients

(1) an APACHE
stratification

ifi
defined

score,
patient

which
ill

can

provide
and
(2) an

initial

risk

severely

hospitalized

within

independently

groups;

APACHE

provided daily updates in these risk estimates. When applied across the individual ICUs, the first-day APACHE ifi equation accounted for the majority of variation in observed death rates (r0.90, p<O.000l). (Chest 1991; 100:1619-36)
HOC
= receiver

predictive equation, which uses APACHE reference data on major disease categories

and and treatment

III score

operating

characteristic

T
risk
been
*From

he ability mortality

to objectively or other for clinical

for

estimate

patient outcomes

risk

for

monitoring

resource

utilization,

and at prediction,

in

improving

important research.
important

is a new based have

in

quality

assessment.56 much
risk

Attempts less successful

however, individof daily

undertaking
assessments

Empirically clinical events

new therapies,

have been ual patient

clinical
in

in forecasting
the uncertainty

or in reducing making.7

extremely
the ICU

useful
Research

in evaluating
Unit, Department

decision

of Anesthesiology,

Objective risk the high-cost,

of

estimates emotional,

are

particularly important and technologically decare units (ICUs).

George Washington University Medical Center, Washington, DC ( Drs Knaus, Wagner, Zimmerman, Bastos, and Murphy; Mr Lotring); APACHE Medical Systems, Inc (Ms Draper); Health Services Research and Development Center, School of Hygiene and Public Health, Johns Hopkins University, Baltimore (Dr Bergner, Ms Damiano); Department of Critical Care Medicine, National Institutes of Health, Bethesda, Md (Dr Sirio); and

manding Because
assurance

environments

ofintensive

essential.6
the basis
in

high costs of ICUs, and utilization management Knowledge ofthe risk faced
ICU admission assurance could and time

the

precise quality strategies are by a patient on

an empiric activities.

Division of Biometry, NC (Dr Harrell).

Duke

University

Medical

Center,

Durham,

day for

of

provide utilization

quality

Supported by the Agency (grant No. HS05787); The 87267); the Department

University
Dr Bastos

Scientific Manuscript Reprint 20037

for Health Care Policy and Research John A. Hartfird Foundation (grant No. of Anesthesiology George Washington Medical Center; and APACHE Medical Systems, Inc. was supported by a grant from the National Council of and Technology Development (CNPq), Brazil.

Estimates
investigating deciding

during
how long

the course
the optimal to continue

oftherapy
for therapy.

could
discharge The

be useful
or demand in

for

intensive

treatment
constrained, One

is growing, half and many of all the

I 100

but

resources
that

are
now

received
requests:

May 15; revision

Dr Knaus, 2300

accepted
K Street,

August

13.
DC

N%% Washington,

increasingly overcrowded.9

ICUs
deaths

are already

CHEST

I 6 I DECEMBER,

1991

1619

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

occur
been

in ICUs made

take

place

only therapy

after

a decision be

has that with

Hospital and ICU Selection

We randomly
1,691 acute-care hospitals beds.

that

further

would

futile.#{176}

selected
in the

26 hospitals
United

to

be States

representative

of all and eligible 85 for based West),

Physicians, decision

patients, and society want to ensure making is accurate and compatible

continental beds in the Hospitals

with
were

200 or more

Approximately ICU study.

50 percent United listed

of all hospitals States on the

current

therapeutic

capabilities.

In this article,

we

percent
participation Hospital size(200

ofall

adult in the

1985 American
16 strata and

present the background, development, measurement, and validation for the system, which is aimed at addressing issues and challenges.

details of APACHE III these various

Association location to 350 beds

data

tape

were

grouped
Southeast,

into

on geographic

(Northeast,

Midwest,

The development
association iologic between balance and

ofAPACHE
acute short-term changes

III was based

in a patients risk of death.4

on the
physDe-

and >350 beds), and teaching status, as defined of resident house staff or the existence of an accredited graduate medical training program. A computer.based random number table was used to assign a number to each hospital. We invited the first hospital listed within each of the 16 strata to participate and the second hospital in ten strata. When a hospital by the presence
declined to participate, a randomly selected alternate was chosen.

tails of the two previous recently. Briefly, in developing

improve APACHE upon the risk
II by reevaluating

versions

have

been

reviewed

Among

APACHE
prediction

III,

we sought
and weight-

to

available

with

the selection

ing

of physiologic
in
to

variables
patient selection

while
for

examining
and timing

how
of

differences
admission

hospitals. between by risk

ICUs related We also sought

using of the mortality (2) expand

to outcome

variations

across

initial 26 randomly chosen hospitals, 23 agreed to participate. The three reasons for nonparticipation were the sale of a hospital; a severe nursing shortage, making data collection assistance unlikely; and a poor fiscal condition, making bankruptcy and closure imminent. In all three cases randomly selected replacement hospitals were chosen from the same computerized listing. The 26 randomly selected hospitals were combined with 14 volunteer institutions to complete the 40-hospital data base. The
14 volunteer hospitals were primarily tertiary-care referral with more instituthan data

the

to (1) clarify the APACHE scoring system within the size independently

distinction to stratify defined

risks

patient
ofmortality;

groups

and using it to estimate

individual

tions with an interest in this project. In hospitals one ICU, data collection took place in the unit annual admission rate. In two of the volunteer collection took place in two separate ICUs.
Ihtient The ICU SeleCtiOn inception patients cohort consecutively

with

the highest

institutions,

and representativeness

of our reference data regarding the selection led to this

base; and (3) examine issues of patients and the timing of

at each

ICU

consisted

ofapproximately

400

scoring.#{176}8 These and additional considerations which investigation are detailed in a comprehensive study
PATIENTS
AND

collection. consecutive

scheme
in the

When patient admissions, (eg, every second

for a minimum

admitted following initiation of data volume precluded data collection on we used an alternating data collection or third patient). A patient had to remain
of 4 h to be included in the aged study. We

ICU

did not include

METHODS III were to establish of candidate predictive (1) 16 years, out

patients

or individuals

with burn with chest

less than pain who were admitted to rule

injuries, patients however, were were

The two major analytic steps in developing APACHE the collection ofan appropriate database and (2)analysis a final system design. First, we assembled a list
variables constructs bidities, %firlable

myocardial infarction. Other cardiac diagnoses, included. Data for coronary artery bypass graft
thatwill

patients

and
(major

questions disease

for

each

of

the

five

major

collected in an independentdatafile No patients were excluded because

be reported

categories,

and origin
Selection

and timing

acute physiology, ofpatient selection).

age, comor-

of missing values. datacollectionwere reviewed andapproved by the data coordinating center, which continuously monitored data quality. Data collection began in May 1988 and was completed November 1989. All patients were followed up for survival
hospital obtained discharge. for all Data Medicare on survival after hospitalization

separately. All breaks in

by
at were

To determine the primary reason for ICU admission, we developed a comprehensive list of 212 disease categories. Each category classified the patient according to medical or surgical status, major organ system involved, and, when possible, specific etiology. Within 24 h of ICU admission, a patient was assigned to one of these major disease categories by using data available in the medical record that indicated the disease process most directly responsible for the patients ICU admission. On the basis ofpast experience and clinical
judgment, ofdisease. sheets ICU we Data selected 20 physiologic variables to measure severity

patients

and for a 15 percent

random

sample
Data

ofall

other patients.

Coliection

collectors obtained physiologic data from ICU and recorded the exact values at various times during

flow the

study.

Data on the patients prior age and the presence of limitations or comorbidities. during the initial 24 h ofICU
(emergency, recovery, hospital, patients

health status consisted of chronologic one or more preexisting functional Chart abstractors collected these data admission and also noted the location
oroperatingroom; ICU readmission;

After agreeing to participate, each hospital selected data collectors and sent them to Washington, DC, for a three-day training course that included detailed instruction in recording all aspects of disease, physiologic, and chronic health data. As an initial quality assurance test, we reviewed data collected on the first 20 patients. if this review demonstrated that the data collectors were able to complete dataformswithaccuracyandcompleteness, datacollection began; otherwise, additional training occurred. All patient data were entered into on-site microcomputers by using specially designed software with extensive internal logic and edit checks to increase the accuracy ofdata collection.

or transfer
time between

from another
the

ICU or hospital)
arrival in the

and,

when

applicable,

the
ICU

Reliability

ofData

Collection
were selected for a formal reliability study anecdotal datacollection problems. Because System (Knaus

emergency

room

and

admission. 1620

Eleven hospitals because they reported

APACHE

III Prognostic

eta!)

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

this

reliability III, Score

study scores

was

completed compared.

in 1989,

before

APACHE
APACHE Physiology (interclass

III data

base

was completed,
were

APACHE
Assignment II was highly

analysis of the II, rather than of the Acute reproducible

and

assigning
A.

zero

weight

to missing

physiologic

values

are

in

Appendix

We estimated

weights ranges using

componentofAPACHE

that

met our magnitude

for the comorbid chronic health variables and statistical criteria and for age divided the entire data file (both estimation and

coefficient=0.90). Frcent agMement values age, sex, race, ICU admission date, origin of admission, time from emergency room to ICU, and primary system failure were 91.8, 99.5, 95.4, 97.4, 98.0, 85.7, and 93.9, respectively. In 77.4 percent of cases with one or more chronic health items, both abstractors agreed that the patient had one of the 34 potential chronic health conditions. A reanalysis using the APACHE III weighting approach has demonstrated similar results. correlation

into fiveyear

for patient

validation halves). Wights assigned to the physiologic variables, however, remained those established by the estimation sample. We then converted the weights for physiology age, and chronic health conditions into points for construction of the final APACHE III score.
The APACHE To produce first location day

III Equation
an equation treatment, for predicting we combined hospital the mortality disease and after

Statistical Analysis
Our specific analytic objectives were to maximize the explanatory power of APACHE III, to improve discrimination at both low and high risk of death, and to maintain maximum measurement reliability. This involved examining weighting ofthe components of the APACHE III scale as well as investigating their relationship with an APACHE III equation that predicted patient outcome by using the
APACHE practices. III score, disease classification, and patient selection

the

of ICU

patient

ofdata collection, we first examined differences the 14 volunteerhospitals and the 26 randomly selected hospitals, using the APACHE II equation. No significant differences in risk-adjusted mortality rates between these two groups were found (p=0.20). Therefore, we combined data from all 40 hospitals into a common file for all subsequent analyses. Ninety percent ofthis data file was then randomly divided into estimation and validation halves, with approximately 50 percent of each hospitals patients included in each half. Analyses used in developing weights for the physiologic components of APACHE III were completed on the estimation half of the data file alone. The 40 diseasecategoriesfrom the APACHE II equalionw servedas diseasespecific controls. The chronic health, age, and previously described patient location variables were also controlled statistically in this
in mortalitybetween analysis.

Upon completion

coefficients with the relative weights assigned to the firstday values of the three components of the APACHE III score (physiolog#{231} age, chronic health). Because of the large number of disease categories in the APACHE III data base, coefficients for each disease classification with a sufficient number of patients and/ or deaths were obtained using the entire data file. The coefficients for specific diagnostic categories were examined to assess their stability with regard to the weights derived and used with APACHE II and the clinical experience of the investigators. The clinical homogeneity, the cell size, and the impact ofthe disease on shortterm outcome were the criteria used for assessing stability. Because of previous evidence suggesting that the patients treatment location immediately prior to ICU admission held important prognostic significance, we next assessed the effect on overall explanatory power of a variable describing selection for intensive care. We did this by assessing the effect of the patients

Table 1-Demographic
Total patients
26 random 14 volunteer hospitals hospitals

Characteristics

of Patients

10,941 6,499 17,440

Nonoperative Component Variables and

admissions room 6,199

Weights for the

Emergency Floor

physiologic variables, we used multivariable logistic regression analyses to determine the relationship between death rate and each of the 20 candidate physiologic variables, controlling for 19 other physiologic variables, age, chronic health conditions, operative status, and major disease categories by usingboth categoric and continuous weightingapproaches2l (Appendix A). We also explored interactions between physiologic variables by evaluating individual and combined variable weighting. For the
variables bicarbonate), reflecting acid-base disturbances (serum pH, Pco,, and

To estimate

weights

2,860 from other hospital

from other ICU 423

Transfer
Transfer

581
10,063 (5)*

1stoperafive Elective
Emergency

admissions surgery
surgery

5,811

1,566 7,377 (4)*

(299-449)

we found discrepancies that were not compatible with established physiologic principles. The computer-derived weights for a serum Pco2 above 50 mm Hg were consistently estimated as having little or no significant relationship to risk of death. We hypothesized that this was because the appropriate weighting for Pco2 is also dependent on the associated serum pH (is, whether there is a primary or secondary respiratory disorder). Therefore, we developed a combined variable, which included serum pH and Pco1, to establish weights for common acid-base disorders. The subsequent estimation of weights for this combined variable proceededasdidallothers. Inaddition, we found important interactions between urine output and serum creatinine, and also between respiratory rate, PaO, and ventilator use. In each case, we developed combined variables and compared them with individual variables for the clinical validity of their respective weights. All physiologic weights estimated on half of the data base were subsequently validated in the independent half. Details regarding choice of timing, assigning weights, disease-specific weighting,

number ofpatients Age, yr (mean, 59 yr)

Average <45 45-54
61

in each

unit

(range)

425 23.2t

10.8 18.0 25.5 17.2

65-74 75-84 85

5.3
44.8 55.2

Sext Male Female

Racet

White
Black Hispanic Asian

80.3
14.1 4.1 1.2

American *Values
tValues

Indian are percentages oftotal

CHEST oftotal number

0.3 of patients.

in parentheses are percentages

number
I 100 1

of patients. 6
1 DECEMBER.

1991

1621

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

APACHE

Ill PHYSIOLOGIC

SCORING

FOR

VITAt.

SIGNS

AND LABORATORY

TESTS

Ficuiu

1. APACHE

III scoring

for vital signs

and

laboratory

abnormalities.

location

immediately

prior

to

ICU

admission

(emergency

room,

outcome signed

resulted to the

when

additional of physiologic

weight

was

as-

floor, otherlCU, otherhospital)and the time spent in the emergency room on the entire model. We also investigated the influence of whether surgery was performed on an emergency basis on total explanatory power. In order to provide an initial basis for predictions over time that require incorporating changes in the patients physiologic status, we investigated the use of daily APACHE III scores as an additional independent variable within the above equation. RESULTh

extremes a zero

measurements

and
were

when

narrower

ranges
or

of physiologic
normal weight. the

measures
For some

assigned

variables, such with extremely risk associated ings. Overall also increased interactions:

as blood pressure, high recordings with equally

risk associated

is different
but for low the

from

the

extreme
power

record-

explanatory

for patient

outcome
following

The majority nonprofit, and

medical was school.

(89 percent) 54 percent

The average

of the 40 hospitals were were affiliated with a

number of hospital beds

when we accounted serum pH with PCO2,

with urine
use.

output,

and respiratory
of this analysis, the

serum creatinine rate with ventilator

weighting of each

As a result

359, the average number of ICU beds was 21, and the average number of ICU beds was 13. These characteristics reflect national statistics on the 1,691

Us hospitals
studied,

with 200 or more beds. Of the 42 ICUs 71 percent were mixed medical-surgical, 16

physiologic variable used in APACHE II was refined for inclusion in APACHE III, and five new variables (blood urea nitrogen, urine output, serum albumin, bilirubin, and glucose) were included
(Fig 1 and 2). Examination ofthe relationship death rate and each of the 20 candidate variables bicarbonate inclusion Glasgow revealed that serum potassium between physiologic and serum

remaining

were surgical, 10 percent were medical, and represented other specialties. Patient demographics, including location prior to ICU admission, appear in Table 1. A total of9,195 (53 percent) ofthe patients had positive answers to one or more of the 34 chronic health and functional status questions. This proportion decreased to 33 percent for patients admitted following elective surgery
3 percent Physiologic Variables and Their Weights

percent

did
criteria.

not

meet suggested

our

minimal

statistical the
scores

Reliability
Coma

results

that reformatting
similar

variables

to eliminate

with different clinical presentations We accomplished this by eliminating between incomprehensible words
sounds, flexion withdrawal and

would
the

be useful.
distinctions

and

inappropriate
rigidity,

decorticate

Compared
APACHE

to the II, increased

weighting explanatory

system power

used in for patient

and

decerebrate
the

rigidity
evaluation

and
of eye
APACHE

no responses,
opening (Fig

and
3).

by

simplifying

1622

III Prognostic

System(Knaus

eta!,

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

APACHE

ifi

PHYSIOLOGIC

SCORING

FOR

ACID

BASE

ABNORMALITIES

pCO2 pH

<25

25-<30

30-<35

35-<40

40-<45

45-<50

5O-<5S

55-<60

I
<7.2

I
-

7.15-

7*I
7.25-<7.30
,

9 <7.40 0

:::::

7.35-

7.45-

<7.50

*ci
7.55<7.60

!
3

=
2

-12

!ii
I
disturbances.

7.65

I
FIGURE

;:-2. APACHE III scoring for acid-base

On the basis
worst most component value appropriate of

ofcomparative
the scoring APACHE initial 24 III.

results,
h of for The approach

we retained
ICU the final care results physiologic

the
as the for

from

0 to 252; A missing

for

an

individual

variable, value

the

range a zero

is

over

0 to 48.

physiologic

is assigned

weight
the

(Appendix
of these and

A). In Figure
power

4, we have compared
of the
in the

explanatory

weightings
1 , 2, and APACHEIH
Ey op

of all physiologic 3. The range

variables

are in Figures
weight is

weights

and discrimination 17 physiologic variables

final
esti-

for total
SCORING

physiologic
FOR

mation

validation

halves

of the

data

base,

using

PHYSIOLOGIC
ar to W/v:bs1
orleited

NEUROLOGIC

ABNORMALiTIES

aiie
verbal

MimUlatiOsi

vaam
mothr
obeysveb.J

msthmnd couIvOslatioul

#{252}iappropriate wore and mOsmmwU 10

no rpoe

15

and Iocali 3 8 13 15 24

decorticate
-ty

rigity
25

ro

EyOs do not op
vbal

saneoisiy

or to painfUl/verbal
oriented mnyOs.sOs

stimulation conh&sed inappropriate work and no rmpome

conversation

mor . obesvaiI
command
.

incombIe mun#{232}
.

. .
,

16

Iocali flezioa

witb#{227}awal/ decorticate rigidity deOsrebrala rigidity! is,rupou

. . ., ..

:
.

: . -Th

16 33

- -,
. 1. 34

.-

. . .

-,

:.:
.

#{149}

4*
#{149} #{149} .

.,-

.
and unlikely
we had

. .

. .. , ,

:
FIGURE

The shaded
IMI SOs

15 dime

arms witbosdacores reprment calls. For the shaded areis

in any of thme cells should

manual
with scorm

PI_

a eIt

he done

after

dinical combinations. There were dita that pennit m to extrapolate valum. careful continnation of dinical findings.

3. APACHE
according

malities opening.

III scoring to presence

for neurologic or absence

abnorof eye

CHEST

I 100

I 6 I DECEMBER.

1991

1623

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

APACHE
N (Thousands) 5_

UI RANDOMIZED
VALIDATION

SPLIT
OBSERVED

HALVES
DEATH RATES 100% -80% -60%

4
3 2 1 0 0

-40%
20%

0%

10%
N ESTIMATION

PREDICTED
:

RISK
RISK RISK

RANGES
ESTIMATION LIDATION
Patient distribution and observed

L
FIGURE 4.

N VALIDATION
APACHE III physiology randomized split

halves

validation.

death
Disease

rate
and

for

each

10 percent

increment

of predicted

death

risk

are

estimation

half and 7,840 patients chronic health weights

in the validation half, using were from APACHE II.

APACHE

shown for 7,840 patients in the III physiologic weights in both.

the
for ating

40 APACHE
disease. The

II diagnostic
APACHE III 0.88

categories
physiologic

as controls
weights

23) and APACHE

age ill

(range, Score

0 to 24) variables

(Table

2).

forecast idation).

well
characteristic

to the

validation
[ROC]
=

sample

(receiver
0.87

operval-

estimation;

The score
groups
health)

that results
variables 0 to 252;

from
(physiology

the addition
age,
with a range

ofthe
and

of

three chronic
0 to III

is a cardinal

number

of 0 to 299

Comorbidity
Estimation

and

Age

Weights
34 candidate chronic health items

(physiology 23;
score.
was

chronic
III

health
score

evaluation, APACHE
population in this

of the

age,
50.

0 to 24).
The mean

It is referred
APACHE

to as the

yielded
syndrome tumor ously 2).

seven variables (acquired [AIDS], hepatic failure,

with metastasis, leukemia/multiple

immunodeficiency lymphoma, that

occur explanatory

solid previ(Table

myeloma,

immunocompromise,

and statistical
these variables improve

cirrhosis) criteria
did not

met

Table 2-APACHE
Chronic

ill Health

Pointsfor Evaluation

Age

and

established
Because

for inclusion admissions

frequently

Points

with
cause

elective
they did

postoperative
not

ICU
overall

and

be-

power

Age, yr 44 45-59 60-64

65-69

within these elective surgical conditions are not required admissions. when the

categories, for elective

the comorbid postoperative ofAPACHE surgery III case,

5
11

They are included as part patient is an emergency

to

13
16 17 24 condition* 23 failure
16 13

70-74
75-84

defined
ening Using estimation

as surgery
condition.

treat

an weights

immediate derived we then

life-threatfrom reestimated

85
Comorbid

the
half

physiologic of the data

the base,
from

AIDS Hepatic Lymphoma Metastatic

base,

the remainder of the equation this time using 78 diagnostic

on the entire data categories adapted

cancer Leukemia/multiple
Immunosuppression Cirrhosis *Exclud for elective surgery

11 myeloma 10

the original 212 based on frequency and death rate (Table 3). Using this equation, we also derived the final
weights for

10
4

the

chronic

health

evaluation

(range,

0 to

patients.

1624

APACHE

HI Prognostic

System

(Knaus

et a!)

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

Table 3-Major

Disease

Categories

in APACHE

III Prognostk

System Odds
Ratio for Increase Death III Rate Score Increase 95% Confidence Interval

Unadjusted Total Hospital Death Rate, %

in Hospital Risk in APACHE

for 5-Point

Disease
Nonoperative

Category

Patients

Cardiovascular/vascular
Cardiogenic Cardiac Aortic Peripheral Rhythm Acute Other Congestive arrest aneurysm heart vascular failure disease infarction diseases shock

41 414
53

65.9

1.20
1.24 1.11 1.30 1.56 1.33 1.38

59.9
26.4 21.0 13.7

1.07-1,34 1. 19-1.29
1.00-1.23 1.24-1.35

891
95 340 603 124

1.26-1.93
1.22-1.44 1.28-1.48 1. 13-1.52 1.22-1.40
1.03-1.18

disturbance
myocardial cardiovascular pneumonia pneumonia neoplasm (including (non-cardiogenic) pulmonary obstruction disease larynx, trachea)

Hypertension

272
59

10.6 10.1 8.1 21.0

64.4

1.31
1.30 1.10

Respiratory
Parasitic Aspiration Respiratory

102
54 131

50.0
51.9

1.18 1.12
1.17

1.09-1.28 1.04-1.21 1.09-1.25

Respiratory
Pulmonary Bacterial/viral Chronic

arrest
edema obstructive embolism airway pneumonia

107
454 362 165 68 140

38.2 37.4 32.8

23.8 19.4

1.21 1.21 1.28 1.24 1.30 1.40 1.22

1. 11-1.33 1. 16-1.26
1.20-1.36

Pulmonary

Mechanical
Asthma

Other respiratory diseases Gastrointestinal (CI)

Hepatic failure due due to varices

319
30

17.6 7.9 32.3

50.0

1. 15-1.34 1. 13-1.49
1.18-1.66

1. 16-1.28 1.02-1.23
1.16-1.53

1.12
1.34 1.21 1.25 1.28 1.44

CI CI CI CI

perforation/obstruction bleeding

87 210 (ulcerative colitis/crohns/pancreatitis

168

26.4 23.3
22.6

inflammatory
bleeding

disease

1. 13-1.28 1. 15-1.36
1.22-1.34

to ulcer/laceration

644
103 92 194
161 326

CI bleeding
Other
Neurologic

due to diverticulosis

CI diseases

14.4 6.8 28.3

54.1
36.6

1. 14-1.81
1. 14-1.41

1.27
1.37 1.39

Intracerebral
Subarachnoid

hemorrhage
hemorrhage

Stroke

30.4 27.5 25.0 15.6

15.2

1.25

1.27-1.49 1.26-1.52 1. 19-1.32

1.03-1.25

Neurologic
Neurologic Neuromuscular Seizure

infection
neoplasm disease diseases than urinary tract origin tract)

51 40 45 309

1.14
1.30

1.08-1.55
1.07-1.63
1.22- 1.42

1.32 1.32

Other
Sepsis Sepsis Sepsis Trauma

neurologic (other ofurinary

115
415

11.3
52.0

1.32
1.18 1.15 1.30 1.44

1. 15-1.51

1. 14-1.23
1.07-1.23

104 multiple
head

29.8 13.4 2.0

35.3

Head trauma
Multiple Metabolic trauma

(with/without
(excluding

trauma)

trauma)

477 399
68

1.23-1.37 1.23-1.67 1. 15-1.49 1. 13-1.34

1.22-1.65 1.20-1.50
1.

Metabolic
Diabetic Drug

coma
ketoacidosis

1.31
1.23

overdose

Other metabolic diseases Hematologic Coagulopathy/neutropenia/thrombocytopenia

Other Other hematologic medical diseases diseases

274 646 143 37

64 92 244

4.4 0.9 20.3

45.9

1.42
1.34
1.37 1.19

13-1.65

4.7
23.9 10.2

1.01-1.40
1.06-1.31

Renal diseases
Postoperative Vascular/cardiovascular

1.18 1.46

1.29-1.64

Dissecting/ruptured Peripheral vascular

Valvular Elective Peripheral heart artery surgery

aorta disease aneurysm

bypass

104

41.3 14.4

1.20 1.28 1.31

1.12-1.30
1.17-1.40
1.15-1.49

(no bypass repair

graft)

215 211 525

abdominal

graft

535

8.1 6.5 4.7

CHEST

1.27 1.51
I 100 I 6 I DECEMBER,

1.18-1.37 1.32-1.73 (Continued)

1991

1625

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

Thble

3 continued
Carotid Other endarterectomy cardiovascular infection diseases 429 225 2.1 9.8 1.78 1.29 1.64 1.40 1.32 1.47 1.31 1.28 1.26 1.32 1.32 1.30 1.23 1.64 1.17 1.35 1.34 1.56 1.36 1.52 1.26 1.41-2.25 1.17-1.44 1.13-2.36 1.25-1.57 1.08-1.61 1.27-1.70 1.22-1.40

Respiratory
Respiratory Lungneoplasm Respiratory/neoplasm

57
411

8.8
5.8 3.4 6.4 26.2 20.9 17.2 14.7 12.5 10.4 7.6 11.1 43.1 26.1 9.7 5.6

Other

respiratory

(mouth, diseases

sinus,

larynx,

trachea)

119 218

Gastrointestinal
CI perforation/rupture CI inflammatory disease
Globstruction

260 244
308 109

1.19-1.37

1.17-1.36
1.16-1.49 1.03-1.68 1.20-1.40 1.09-1.40 1.36-1.97

Glbleeding Livertransplant
Glneoplasm

40
500 170 180 hemorrhage 51

CI cholecystitis/cholangitis
OtherGldiseases Neurologic Intracerebral Subdural/epidural

hematoma
hemorrhage spinal cord surgery

88
93 214 437 126 multiple head trauma) trauma) 210

1.07-1.29 1.19-1.52
1.13-1.57 1.29-1.93 1.23-1.52 1.23-1.88 1.18-1.34

Subarachnoid
Craniotomy

Laminectomy/other
Otherneurologicdiseases

for neoplasm

5.5
7.1 22.9

Trauma
Head Renalt trauma(with/without

Multiple
Renal

trauma

(excluding

381 225
173 65

3.7
4.9 4.6 7.7 12.2

1.39
1.34 1.45 1.28 1.19

1.24-1.56
1.12-1.59 1.18-1.79 1.06-1.54 1.05-1.36

neoplasm Other renal diseases Gynecologic

Hysterectomy Orthopedic

Hip or extremity
*De tOdds

fracture represent

139

categories
ratio

the single

most specific

reason

for ICU

admission.

for 46 renal

transplants

not calculated

because

ofsmall

number

ofdeaths

(2.2%).

The
obtained

dfrect
during

relationship
the first

of the
day

APACHE
treatment

III score
within

plotted

of ICU

two homogeneous diagnostic groups, congestive heart failure and trauma, is illustrated in Figure 5. Within specific disease categories, the relationship of increases in the APACHE III score to the risk of death is also reported in Table 3, where the odds ratio of increased risk of death relative to APACHE III score is provided for each of 78 major disease categories included
odds regression ratios

against APACHE III score. When the III score is either low (20) or high (140), the relative importance ofdisease is small. Within the
APACHE

middle
lions
III

range
in risk

of scores,
are

classffication score.

however, variations associated importantly

in disease

predictions

for the same level

with variaof APACHE

The
lated

overall explanatory power of estimates calcufrom the data obtained during the first day of
and ROC

in the APACHE is calculated

analysis

III data from a specific

base. Each a separate disease

of these logistic category

ICU stay is evidenced by the total r of0.41 of 0.90 (Fig 7). Overall correct classification
first
7).

on the
(Fig

within

day at a 0.50

predicted

risk was

88.2

percent

The
another. Risk An and

equations

are

statistically

independent

of one

These

values

are significantly

better

than the over-

all Estimate equation

Equationsfor

Hospital the

Mortality power
category

explanatory power available from APACHE II = 0.85 and a correct classification at a 0.50 risk level of 85.5 percent).6 When the initial or first-day
(ROC

combining

explanatory

of of

APACHE

III

equation
for

is applied
most of mortality rates

across
the large (6

the individual
variation percent to in 42

the APACHE
prior patient hospital under

III score
location

with

major

disease
the

ICUs,
observed percent)

it accounts hospital

permits to those

calculation

death

risk estimates
similar

for ICU patients

in this

admitted
study.

circumstances

The relationship between disease classffication and risk prediction is further illustrated in Figure 6, where
predicted
1626

from these units (r = 0.90 p<O.000l). When we investigated the use ofAPACHE on the first and subsequent days to estimate risk over time for individual patients, we

III scores mortality discovered

risk

categories

for

four

major

disease

is

that the use of initial

and latest-day
APACHE III Prognostic

scores
Syem

achieved
(Knaus

eta!)

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

APACHE III AND 891 CONGESTIVE

N OF CASES

HOSPITAL HEART

DEATH FAILURE
OBSERVED

RATE FOR PATIENTS

DEATH RATE

2OO-

----

- -

100%
80%

150

I
I

60%

100

40%

50

0
0 10 20

30 40 50 60 FIRST DAY APACHE

UL

I
90 100
110+

70 80 III SCORE

N OF CASES

DEATH

RATE

N OF CASES

APACHE III AND HOSPITAL FOR 1,467 ICU TRAUMA

DEATH RATE ADMISSIONS

OBSERVED DEATH RATE

400

----------------- -------------

100%

300

80%

60%

200

40% 100
L

20%

0
0 10 20 30 40 50 60 FIRST DAY APACHE 70 80 III SCORE 90 100 110+

0%

maximum

DEATH

RATE

-*--

N OF

CASEj

FIGURE

5. Relationship

between

first-day

APACHE

III score and risk ofhospital death for patients with congestive heart failure (top) and trauma (bottom). relationship more greater complex between than blood pressure

explanatory the daily

power. risk of hospital

In

Figure mortality

8, we over

have

and

outcome than

was
with

indicated
hospitals.

for two septic

Each

shock
daily

patients
risk

drawn
estimate

from
(after

time one of the 40

the initial

originally

hypothesized

risk associated

with

hypotension

hyper-

day) is derived from the initial days and most recent days APACHE III score, along with major disease category and patient location variables.
DIscussIoN

tension, a relationship previously reported by others. We assign zero weight to missing physiologic data
because inferred weights would other artificially techniques inflate suggested risk

estimates and because that normal imputed

values

were power points

most enough
APACHE

appropriate.

Because
abnormalities of APACHE acIlls improve
the

a continuous
overall
of

weighting
application,

approach

did
III

not
uses

As anticipated, counted for the total new ologic

acute largest

physiologic proportion

explanatory
its

to warrant direct A of 1627

difficulty

explanatory power (Appendix A). In developing weights, we discovered that the impact of physiabnormalities in on hospital II. mortality For had example, been the APACHE

specific

physiologic

cut

that

permit

calculation

of the APACHE

III score

(see Appendix
contribution
1991

underestimated

for details). When we evaluated

the incremental
CHEST I 100

I 6 I DECEMBER,

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

APACHE III AND RISK OF DEATH: THE IMPORTANCE OF DISEASE

HOSPITAL
100% 80%
60%

RISK

OF DEATH

40%
20%
,

/
20 40 60 80 100 120 140 160

( FIGURE

and with

6. Relationship between predicted risk of hospital

APACHE

III score

death

for patients
-a--

postoperative subdural hematomas (S SDH), sepsis (other than urinary tract), bacterial pneumonia(BACT PNEUM), and postoperative gastrointestinal perforation (S GI PERF).

FIRST
SDH -4--

DAY
SEPSIS

APACHE
-aBACT

III SCORE
PNEUM -s-01 PERF

comorbidities

and

functional

status the patients

limitations

to

Clinical

Research

Applications
challenge for clinical

short-term
bid conditions status were requirements

outcome,

we discovered

that those

comor-

A fundamental

trials

involving

that influence the only ones for inclusion.

immunologic

acutely
groups

that met our statistical This is not surprising,

ill patients is that the treatment should be at an equivalent baseline

when
mation despite provides
is

patients
may

are randomly
detect risk

assigned,
differences

and control risk. Even prior risk estithat occurred

since

infection

is commonly

associated

with

both

hospital and ICU mortality. tions are also influential power for patients classffied surgery. Finally, we III equation, intensive capturing care

These comorbid condiin increasing explanatory as undergoing variable of the this to the patient emergency APACHE prior to at

randomization. a continuous

The use pretreatment

of APACHE III also risk measure. This

an improvement

over categorical
number statistical of patients significance

risk categories

and

added a new the location treatment;

can reduce the needed to attain

or observations between arms

variable

is aimed

the

prediction.
who are ICU

impact Among
readmissions,

of selection nonoperative
transfers

bias on outcome patients, those

from other units,

of a randomized trial.v Within each of 78 diseases listed in Table 3, an increased APACHE III score was statistically signfficantly associated with an increased risk of death.
the

Therefore,
statistical

in all
power

78
and

of these
precision

disease
of

categories,

trials

and admissions

from

the hospital

wards

have

marginadmitroom

ally increased risks ofdeath relative ted directly to the ICU from the (Appendix A).

to patients emergency

of experimental the end point measure such

therapies where could be improved as APACHE III.m

short-term by using

death is a severity

The

APACHE

III score

can also be direcily

calcu-

The APACHE III system consists oftwo options: (1) an APACHE III Score and (2) a series of predictive
equations
APACHE

lated and used dently defined

within disease categories or indepenpatient groups to provide severity

disease

linked to a reference patient data base. The III score may be used alone within a single category or any other independently defined

stratification
application

either
is similar

before

or after randomization.
use ofthe

This

to the recent

APACHE

patient
illustrated
APACHE

group

use of disease

relative risk stratification, as in Figure 5. Predictive equations link the III score to our reference data base by the separate variables for patient location and classification and can produce risk estimates

to perform

II score to assess hospitalized patients in such studies may

risk in studies of severely ill with infection. Since patients be selected on the basis of criteria

different from those ofthe APACHE III study, the risk levels associated with specffic APACHE III scores in this study may patients meeting not calibrate precisely with different selection criteria.8 those for

ofhospital

death

for individual

ICU patients

at various

times during applications,

their ICU stay (Fig 7 and 8). In both the use of the score with precise meas-

The same For example,

for trauma
subcategories

principle applies to disease classification. the risks displayed in Figure 5, bottom,

patients
listed

urement
amount (see
1628

of other
ofunexplained

patient

characteristics
in hospital

reduces
death

the
rates

are

those

for
3 (ic,

the

four

trauma

variation

in Table

due to previously
Appendix

unmeasured
B for examples

patient

characteristics

trauma,

both

nonoperative

head and multiple and postoperative). Each

III Prognostic System (Knaus

of application).
APACHE

eta!)

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

RISK STRATIFICATION BY FIRST DAY APACHE III EQUATION FOR 17,440 UNSELECTED ICU ADMISSIONS
OBSERVED HOSPITAL
MORTALITY RISK

0%

20%

40%

60%

80%

100%

PREDICTED

HOSPITAL OBSERVED

MORTALITY D R
45

RISK GROUPS
DEGREE

.-*--

CLASSifICATION

TABLES

A.

AT .11 PISDICITD

UIt

HOSPTTAL OIAIII

AT

PREDICITD

REOF

HOSPAL

DEATh

*u
TRUE

oce u
2I

TOTAL

AIAE TRUE AUVE DEAD TOTAL 742% 71.1 313

%.R

DEAD 573

TOTAL

.t
AD TOTAL

34

1U 15 lUG

I4414

15,

5$7

174

2IN

17.40
0.1%
04 NJ

%cOSRRCTaA$IIncAlKn4 w4smvnT atcincnv

PREOICflVEVALUE POSflWE

%coEncraAsslncATIoN

wnvnT swmcnv

PUEBIcUEYEvALUE

A171t

risnicmvs

PREDICtiVE VALUE POSRUVE

VALUE NEGA11VE

72.7 0.2

C.

AT .0 PREDICIID AliVE TRUE ALIVE lOAD TOTAL 14,04 17.414

RISE

HOSPITAL

DEATh

DEAD TOTAL 32 30 C4 14,414 324 17,444 147% 13J

NJ

FIGURE

7. Top,

Risk

stratification

by first-day

wOmvnY stcmai uzoicrivs

pREDIcm,E

% coiszcr

OASSIJ1CATION VALUE POSVE vws pim.tTnt

,z.s
$4.5

APACHE accuracy Bottom,

0.90

and assessment of predictive unselected ICU admissions. Classification accuracy at 0.10, 0.50, and predicted risk ofhospital death. III equation for 17,440

of these between variations justed

subcategories

has risk

a different are reflected

calibration in the

level

the APACHE
in baseline

III score

and outcome.

These
unad-

substantial (r2 = 0.90). studies observed

proportion of the variation This confirms the findings

hospital indicate that

death rates in J.Ible 3. These variations it is essential to specify precisely the

suggesting that the majority death rates across ICUs is related characteristics.#{176} units this could compare their base

in death rates in previous of variation in

to varia-

tions in patient Intensive

experience care with

patient selection and disease classffication, defined as the primary reason for ICU admission, with the results ofAPACHE III scoring.
Applications in Evaluating ICU Outcome

mortality

To evaluate

ICU patients,
combination and patient 42 individual
hospital

outcome for a multidiagnostic group the APACHE III score must be used
with
location

of in

an

APACHE
weighting. in this study,

III

ICUs

disease classffication When applied to the in which the observed

percent,

death rates varied from the APACHE III equation

6 percent accounted

to 42 for a

by using a patient-by-patient measurement of risk in order to compare the predicted mortality rate with the actual mortality rate. The difference between predicted and actual death rates is one measure of quality of care. This technique has proved useful in a variety of studies comparing the mortality experience of ICUs and investigating the incremental impact of specific treatment and of structural, process, or organizational changes on patient outcome.31 Because bed availaCHEST
I 100 1 6 1 DECEMBER, 1991
1629

reference

data

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

APACHE
PREDICTED RISK OF

III DAILY
HOSP. MORTALITY

RISK

ESTIMATES

0.9

0.8
0.7

0.6 0.5
0.4

0.3
0.2

0.1

0
FIGURE

2
TWO
-*---

3
DAYS IN ICU

4
WITH SEPTIC SHOCK

8.

Daily

risk

estimates

for

two

septic

shock

PATIENTS

patients

from the ICU with the highest average APACHE III score for all admissions. LOS =length of stay, expressed in days; HLOS hospital LOS.

PATIENT
#{149} 33

-e-

PATIENT
#{149} 21 -

ICULOS-8 HLOS OUTCOME

#{149} ALIVE

ICULOS-lO HLOS OUTCOME

DEAD

bility and screening tially across different in different hospitals variations tices. Potential Applications efforts with 100 in patient

for admission may vary substanICUs, comparisons among units may have to account for these selection and

explanatory from 0.84 judgments

ROC

power to 0.89. from

measured A recent a variety

in terms analysis

of ROC areas that combined recorded an

of clinicians comparison II with

discharge

prac-

of 0.85. inatory ability

clinicians

In a direct of APACHE

of the discrima combination of

correctly placed a

estimates,

physicians

in Clinical will percent never

Decision be able

Making to All predict

clinical

slightly
percent more ofdeath Objective

greater
risk level. discriminating on ICU

number
The admission at least

of patients
APACHE was three

above

the

90
were risk

Prognostic outcome

II predictions patients 10 percent potential whose or less.

specificity.

in identifying estimates

decision making, however, uses guide future decisions. Prognostic reliable ual
brated

past experience systems that and

range

to are calican

prognostic

derived

from

(eg, provide independent

accurate (eg,

identical
throughout

estimates of observer)
the

for an individwell
of risk)

APACHE

III have

advantages

patient that

ensure

the

experiences

of the

past

are taken APACHE (r

=

into III 0.41

compared with clinical judgment. First, they should be more reliable than individual estimates because they are based on reproducible data. Second, the data base than supporting the risk estimate any one clinicians experience, credibility are based not the or other
of the

consideration The overall system on the

in an unbiased manner.26 explanatory power of the initial day of ICU treatment well to that to that ofother classification

is substantially larger thereby providing

and ROC = 0.90) compares versions ofAPACHE15,m and systems. Overall correct decision rule 79. 1 percent

of previous prognostic using a 0.50

additional estimates treatment, for care prognostic patients estimates

Evaluations

to the prediction. Third, the risk solely on the patients response to order in which he or she presents used most heuristic
usefulness

commonly
potential

variables.
of objective

on the initial day with the Mortality 88.1 health, using in predictive application. percent we and our entire Because

of ICU treatment was Prediction Model, with APACHE weights components data base, may occur degradation III (Fig for the of some with will

estimates response over time

are

likely

to begin

with

compared
7, age, bottom). chronic

with

established disease patient

to therapy in the form of risk (Fig 8). In many cases, these daily

APACHE

III

degradation prospective

performance We anticipate

risk assessments will confirm uncertainty regarding the patients ultimate ability to benefit from treatment. The increased reliability and confidence inherent in objective estimates might reduce the potential for error.6 In a few cases, risk estimates could support clinical judgment that continuing current therapeutic efforts would be futile.8 A recent pilot study by Knaus et al#{176} suggested those available decision intensive making. care that probability estimates similar to with APACHE III might assist in such If our expands technical ability to provide capabilities
System (Knaus

be minimal, however, because most ofthe explanatory power resides in the derived physiologic weights, which forecast well to independent data (Fig 4).
APACHE

compare Recent estimates

1630

III outcome predictions (ROC = 0.90) favorably with those of physician judgment. reports analyzing the accuracy of physicians of hospital death have recorded overall

while
APACHE

our financial
III Prognostic

et a!)

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

become competing

more

restricted,

the

capability or their

of evaluating abilities more with to imthe

patients

requirements

analyses the potential use ogy to evaluate the quality outcomes patients anticipated constructs other requirement than for

ofthe APACHE methodolof ICU care and to predict death, ICU such therapies, as the or unique

benefit from intensive care portant.4 We emphasize,

exception of

could become however, that

hospital

evaluation been risk estimates, limitations

the study by of the usefulness The therefore,

Knaus et al,#{176} no formal of such estimates has yet value unknown. of objective

length of ICU and hospital stay. Since the that make up the APACHE III system

completed.

incremental remains

(disease, severity, age, and comorbidities) are also closely related to a patients therapeutic requirements,
the APACHE III methodology and other potential should care need be useful in as ICU method to theralso needs forecasting nursing well as the patients treatment. The most for evaluating patients apy by collecting requirements, for unique

Investigators and clinicians using APACHE III must consider the current limitations of the system and its proper application. There were at least three misunderstandings score without
III

efficient and accurate physiologic responses data over time

ofAPACHE consideration alone

II. First, was the use of the for disease. The APACHE only

physiologic

score

can be used

within

homogeneous

further investigation. The use of the initial and latest measurement is only one of a variety of analytic approaches
Comparisons

disease categories not risk prediction. the use ofa predictive for ICU treatment

and then for severity stratification, The second misunderstanding was equation calibrated on a patient sample by selection selected by that the firston all patients selecventilause of a
18

that are
role

could needed
for

be used.24
risk estimates to clinical

of objective

judgment
appropriate

both
these

to
estimates

determine

the

most deci-

in clinical

different criteria. 16.17 We now emphasize day APACHE III equation is calibrated selected for ICU admission, not tion criteria, tion.6 The first-day To address response, tients
calculate

sions and Physicians

to provide a guide now infrequently

to future improvements. use formal probabilistic data to guide decision makof the incremental impact,

any additional

such as the need for mechanical third misunderstanding was the at other times during the need for predictions we developed equations and updated risk individual estimates

reasoning or quantitative ing. Formal evaluations

acceptance, and value

of objective

risk

estimates

are

equation

the ICU

needed.

Especially

important

to explore the

is whether would of their

based on treatment that use the paIII over time scores (Fig to 8).

feedback of empiric probabilities result in an improvement in subjective probability ing. Fortunately, one way. Finally, prognostic framework, the need estimates one that to place into explicitly

to clinicians accuracy

initial

APACHE

estimates or their decision maksuch large-scale study is under these empirically derived a larger decision-malcing acknowledges and the values funin

APACHE III equations after the initial 24-h period use updated physiologic data and assign somewhat different weights to disease, chronic health, and age components as their implications change over time. We urge the potential user, however, to appreciate or with unusual risk may not be analytic one total classffication of selection Predictive techshould number is and estithat for patients with rare conditions presentations of common conditions, accurately estimated niques. In determining have in any risk estimate, by this the or other confidence ofthe disease scrutiny biases.

damental roles clinical decision

of patients preferences making, is essential.8

CONCLUSION

Every clinical

day, research

clinicians make

and complex

physicians decisions

engaged regarding

in

a review

of patients within a specific needed, along with careful other mates potential must confounding

the scope and intensity value of new therapies enhanced of patient questions

of treatment that might

or the potential be supported or

always be placed in an appropriate clinical context: a risk estimate for a critically ill patient whose clinical status is rapidly changing is likely to change; and the availability ofnew, untried therapeutic options should Future be taken Studies into account.

by an accurate and objective measurement risk. Indeed, many of the most important concerning the quality and appropriateness

of advanced medical care cannot be fully addressed until patient risk is accurately assessed and reliably recorded. The completion of the APACHE III prognostic system is an attempt to provide objective probability tients treated estimates in ICUs.
FINANCIAL All the authors DISCLOSURE

for critically

ill hospitalized

pa-

Empirically derived risk estimates have not been a traditional component of medical research.4 Discovering how best to use these risk estimates to improve the quality of patient care, the precision of clinical research, and patient outcome evaluation will require further investigation. We will be reporting in future

organization

certify that affiliations with or involvement in any or entity with a direct financial interest in the subject matter or materials discussed in this article are disclosed as follows: Drs Knaus, Wagner, and Zimmerman and Ms Draper are each
CHEST
I 100 I 6 I DECEMBER,

1991

1631

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

founders Systems,

and Inc

minority (AMS),

equity a for-profit

shareholders Delaware-based

of APACHE corporation study. based Dr

Medical that AMS upon Knaus, financial

24 h of ICU
When

funded, markets
some

in part, the research for the APACHE III a software-based clinical information system
of the concepts nor described in this article. Neither employees Dr Zimmerman, as foil-time

Dr Wagner,

of George

Washington University, is permitted to receive any direct payment from AMS and cannot participate in any business
decisions. Ms and has Draper has any is a full-time Dr received formal Sirio per affiliation die,n with employee of AMS consultative and

or policy
sits on

data found in which one abstractor scored higher or lower than another. For each of the three readings (admission, worst over 23 h, worst over 24 h), the absolute difference between the mean scores was not statistically significantly different from zero
mean

stay up to seven days. overall APACHE

compared,

II scores

between

abstractors

were

no consistent

pattern

was

(n

the Board for AMS

authors

ofDirectors.

has performed

services of the other

financial or

196). Overall,

data

availability

(the proportion

ofmissing

values)

payments.
or receives

None
any

favored
explanatory

the worst value over the initial

power.

24 h, as did maximum

other consideration

from AMS. We would like in the APACHE to acknowledge III data collection: the St

The second
weight
physiologic

ACKNOWLEDGMENTS: institutions that participated

provided based

consideration to each
were

was how to determine

physiologic divided into measurement. clinically

the exact
The appropriate

variables

Mary Hospital, East St Louis; White Memorial Hospital, Los Angeles; United Hospital, Clarksboro, WV; St Lukes-Roosevelt, New York; Cooper Medical Center, Camden, NJ; Monongahela Valley Hospital, Monongahela, Pa; Easton Hospital, Easton, Md;

ranges

partly

judgment.
a series

They
of separate

on cell size and were then incorporated

predictor variables

partly on clinical into the analysis as for each

were

range.

Burlington
NJ; St Medical

Medical
Lukes

Center,

Burlington,
Newburgh,

Ia; Union
NY;

Hospital,

Union,
Hospital!

Hospital,

Wyandotte

The
with

initial
basic

results
clinical

from
and

these

analyses

compared
Where

Center, Wyandotte, Mich; Daniel Freeman Medical Center, Inglewood, Calif; Terre Haute Regional Medical Center, Terre Haute, md; Flower Memorial Hospital, Sylvania, Ohio; Richland Memorial Hospital, Columbia, SC; Mount Auburn Hospital, Cambridge, Mass; University Hospital, Denver, Cob; Northeast Georgia Medical Center, Gainesville, Ga; layfront Medical Center, St Petersburg, Fla; Craven Regional Medical Center, New Bern, NC; Southwest Florida Regional Medical Center, Fort Meyers, Fla Mercy San Juan Hospital, Carmichael, Ca1i1 Kaiser Foundation Hospital, Los Angeles; Mesa Lutheran Hospital, Mesa, Ariz; St Lukes Hospital, Kansas City Mo; Presbyterian Intercommumty Hospital, Whittier, Ca1i1 Stanford University Medical Center, Stanford, Calif Winchester Medical Center, Winchester, Va; HenWilliam Beaumont Medical Center, Plattsburgh, NY; Mayo Clinic, Rochester, MInn; Cleveland Clinic, Cleveland; North Carolina Baptist Hospital, Winston-Salem, NC; Barnes Hospital, St Louis; Catherine McCauley Health Center, Ann Arbor, Mich; Lexington Medical Center, West Columbia, SC; Fafrfax Hospital, Falls Church, Va George Washington University nepin County Medical Center, Minneapolis;

physiologic

relationships.

discrepancies

existed

(eg, a mean

Hg, assigned a lower coefficient, death, than a mean pressure of ranges. Most of these variations
sizes in the originally designated

blood pressure of 60 mm indicating a lower risk of

Hg), we adjusted the

70 mm

were

due to small
In a few

sample
cases

ranges.

analyses remained incompatible with established physiologic patterns, we adjusted the estimated weights by using clinical judgment. Patterns of weights were also checked using restricted cubic splines
fUflCt1Ofl.

where

the

results

of the

Medical

Center,

Royal

Oak,

Mich;

C.V.PH.

Cubic
that

splines

analysis

is a statistical

smoothing
varying

allows

assignment

of a continuous

technique weight to a

physiologic variable. Because there are no discrete values for all physiologic measurements, the use

threshold
of contin-

Medical

Center,

Washington,
A:

DC.
TECHNICAL ISSUES

APPENDIX

Thning

ofMeasurement

and

Weighting

of approach to Se-

Physiologic

Variables ofinvestigating the optimal

coefficients for APACHE

In the process

leering and weighting physiologic III, there were a number ofspecific

ered.
value

technical issues considThe first was whether the patients initial or admission was preferred over our previous practice of using the
physiologic value over 24 h. For this reason, we

worst

weighting is attractive compared to using specific cut although it would have eliminated the ability to handscore APACHE III. In this data base, however, the use of cubic splines did not substantially increase total explanatory power. Our final method of selecting weights for the physiologic variables was still mainly empiric (le, deriving weights from a random half of the data base and validating it in the validation hail). As emphasized above, however, when these empiric weights conflicted with known physiologic relationships, they were adjusted.
points,

uous

collected

data on both. The first value obtained was the admission value (le, the first value in the initial hour of ICU treatment). if no first-hour measurement was available, the
data
obtained collector

Disease-Specific
Next, we

Weighting
explored

ofPhysiologic

Variables
that the derived weights

the possibility

looked

for

values
it as

obtained
missing.

1 h prior
The next value

to

for physiologic

admission

initial 23 h value (le, the most abnormal reading during the remaining initial 23 h of ICU treatment). We designed strict rules to define the most deranged value for the rare instances when there were abnormalities on both sides of a defined normal range. if no data were available for a particular physiologic variable either at admission or during the initial 23-h time period, we recorded the value as missing. Following data collection, we derived a worse over 24 h value for each physiologic measure using either the admission or worst over initial 23 h determinations (whichever was more abnormal). Finally, worst over 24 h values were obtained for each subsequent
1632

before recording was the worst over

different if examined within a specific disease category, as has been suggested by others.220 We examined the weights assigned to specific target physiologic variables in congestive heart failure, the one well-defined disease category with a large number of patients (891) and a substantial death rate (21 percent). We first estimated weights for the target variable (blood pressure), using only patients within the specific disease category, and then compared these weights to ones obtained by using all patients in the estimation data file except those in the specific disease categories. In this analysis, overall explanatory power was not improved by providing disease-specific weighting for individual physiologic variables. It is possible, however, that for some
APACHE III Prognostic System (Knaus eta

variables

would

be substantially

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

diseases,

use

ofspecific

combinations

ofphysiologic

variables

uniquely

captured

by the acute
for the following the remainder

physiology
portions (100

score.
of unique

The
the

other
explansum of

may enhance predictive ability. It is important to emphasize that when the outcome is dichotomous (eg, alie or dead), one must have a very large number of patients within
homogeneous disease categories to detect and statistically

factors
atory

account
power,

with

percent

validate severity
Missing

significant weighting.
Physiologic

advantages

in various

approaches

to

the parts) allocated jointly to all of the explanatory together: chronic health, 2 percent; age, 3 percent; 6 percent; patient origin, 1 percent. When an equation that uses the acute physiology
alone is estimated,

factors
disease,

score
that the

the
is 86 percent

x2is

5,501.

This
total

indicates of acute

Values

upper

limit

on the

relative

importance
ofthe

of assigning a weight of zero to did this by examining the pattern of missing values and by using dummy variables to estimate the most appropriate weight to impute to a missing physiologic value. Variations in laboratory-test ordering practices across patients and the various hospitals meant that data availability varied for the physiologic variables.
missing physiologic variables.

We examined

our practice

abnormalities

x2#{149}
trends (Fig

physiologic

Timing
sional
patient

ofICU

Admission of individual between patient initial 8), occaand

In our review
discrepancies

risk estimates

final

Ninety-nine on all four

pressure,

percent vital signs

of patients
(heart

had

complete

information

outcome were often due to ICU admission and initial APACHE assessment late in the course or treatment of an illness. The use ofa patient location variable (eg, emergency
room impact however, be coronary versus ofsuch ICU readmission) There
this

ICU.
values

rate, respiratory rate, blood and urine output) during their initial 24 h in the Serum sodium, serum potassium, and hematocrit
were

in APACHE

III reduces

the

differences. even with days For unit

will be occasional

patients,

for whom care

control

for lead-time
is admitted

bias may
to and a

available indicated

for 85 percent

and

arterial

blood

gas

inadequate.

example, an acute

a patient myocardial

measurements
24 h. Analysis

for 65 percent related

data.

of patients

during
ofmissing

the initial
values

infarction

that the proportion

acute
muscle.

mitral
After

value
three
by

insufficiency failure,
admission

was directly
by signs vital sign

to physiologic
Patients with

had

the largest

proportion

stability as determined normal or near normal vital of missing laboratory tests,

physiologic values with

congestive
is followed

heart
III

due to a ruptured papillary therapy for shock and an emergency valve replacement
of medical
to a surgical ICU. Using the

as determined
variables

by the mean per patient. Based

number of missing on estimation fill-in

APACHE

system,

this

patient

is classified not reflect

infarction

as an emer-

gency
death
by the coronary

postoperative
in the surgical
prior acute

valve
ICU
myocardial

replacement.
might

The first-day

risk of

dummy sumed Weighting

variables
magnitude

variables, missing normal and assigned

ofComorbid

physiologic variables zero weights.

were

as-

the risk implied

and shock in the of such patients is

care unit.
small, it is not

Because
possible

the number
to estimate

COnditiOns

very

the incremental

risk

We also based
measuring

the initial
physiologic

weighting reserve

of the and

34 candidate comorbidity data on

associated

therapy.
further bases.

with these variations in the onset of intensive Improved estimation ofthese risks will have to await
research
The time

a regression
and disease chronic

analysis
direction variable

on

the
of the

estimation
influence

file.

The

and
delay

of each (a coefficient

comorbid of 0.05) (T

the collection of larger between emergency

reference

data

on mortality

and the
ratio

overall
were in

statistical
considered APACHE

significance
in determining III. We also

of the
explored

influence

admission did not increase overall not included in the final system. Differences are at not the accounted
limitation important

room and ICU explanatory power and is

for

>2) include

which

variables potential

in patient for by
in

selection APACHE level.

ICU
III, also

therapy,
represent

which
an

to

interactions among chronic disease variables and ing ofacute physiology measurements by chronic (eg, a high serum creatinine level and low urine chronic conditions (eg, a high serum creatinine low urine output by chronic hemodialysis or
dialysis).

confoundconditions output by level and peritoneal

explaining
ICU

patient
Future

outcome

variations for these

hospital

and

interinstitutional characteristics

comparisons, variations

therefore,
by including variables

may
selected

have

to account

institutional

Relative

Importance

ofComponents

ofAPACHE

Ill

Score

or by limiting comparisons to hospitals pressures. Further details on these and other specific analytic issues are available from the authors and will be the subject of subsequent publications.
as control

with

similar

triage

and Equations To illustrate the relative importance of the different components of APACHE III and additional information in explaining interpatient differences in risk of death, we reestimated multivariate logistic regression equations, leaving out different groups of variables that measure global concepts. In the aggregate, the complete equation has a global of 6,426 (p<0.00001). The portion of the global
APPENDIX B: GUIDE
RISK

TO

APACHE

III

PREDICTION

of adult

The procedure for estimating the hospital mortality risk met ical and surgical patients admitted to an ICU
initial day of ICU treatment entails the following

on their

x1

x2
a

that measure

is

uniquely

associated

with

specific

variables

is

Step

ofthe unique importance ofthat factor in explaining risk of death. When the acute physiology score is deleted from the equation, the total drops to 3,396, indicating that 47.2 percent of total interpatient explanatory power is

x2

Choose the single most important reason for ICU treatment from the listing of nonoperative and operative major disease categories (Table 3). The disease category may not be identical to the primary
CHEST

hospital

diagnosis

(eg,
1991

for 1633

I 100 I 6 I DECEMBER,

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

patient with acute leukemia due o aspiration pneumonia, APACHE

leukemia nancy] in the

admitted for respiratory failure the major disease category for pneumonia;
[hematologic

from the authors. Example 1: Nonoperative

woman

III in the
most

scoring
as

is aspiration
a comorbidity

the the

acute
malig-

Admission

qualifies

APACHE
specific

III score).
diagnostic

Always
category

place
possible

patient
(eg, a

postoperative
who is admitted

patient

with a gastrointestinal
following

immediately

(GI) malignancy an abdominal exploshould Ifthe be classified patient does

with acute leukemia is admitted to the ICU from her hospital room following an episode of aspiration pneumonia. She had not been treated in the ICU during this hospitalization.
A 56-year-old

ration that revealed as CI obstruction,

an obstructed colon rather than CI cancer).

Major
five)
Treatment

Disease

Category:

Aspiration

pneumonia

(nonopera-

not qualify for any of the specific categories, use the most appropriate general category for the primary organ system affected. All patients admitted to the ICU from the recovery room or operating room are considered postoperative, and you must choose from one of the postoperative diagnoses. All other patients are placed into one of the nonoperative diagnostic categories.
Step

Location:

Hospital

room

APACHE
Age
Chronic

Ill Scoring:
=

(56 years)
health (leukemia)

5points 10

points
within

Acute
initial

physiologic
24 h): rate (125

abnormalities
beats/min)

(most

abnormal
=

2
patient
floor,

Pulse If the patient

the

7points

is a nonoperative was being treated

admission,
immediately

indicate
prior

where
to ICU

Mean blood Temperature Respiratory PaOJP(A-a)02

pressure
Rate

(75 mm

Hg)

= = =

6points
Opoints

(39.8#{176}C) (361mm) mm Hg; (Pa02=68

admission.
hospital

Treatment
other

locations
hopsital, other

include
ICU,

emergency
and readmission

room,
to

9points

the same ICU. ifthe patient was in a regular hospital room, the correct variable is floor; if they were in a specialized ICU, it is other icu: ifthe patient was previously treated in the same ICU during this hospitalization, the designation is ICU readmission. If the patient was admitted to the ICU immediately following surgery, was the surgery performed on an emergent basis? Emergency surgery is defined as surgery required immediately to correct a life-threatening condition. Step 3
Calculate summingpoints the

F1o2=0.70, mechanically ventilated; therefore, calculate P(A-a)O2: PCO1 = 26 mm Hg; P(Aa)O2 =433)
(24%)

= =

Hematocrit White

ilpoints 3points
5points

blood cell count

(2.2 mg/dl)

(1,2OWcu

mm)

= =

Creatinine
Urine

7 points 5points 12 points 0 points 6 points 6 points 3 points 9 points

output
urea (136

Blood
Sodium Albumin

(1,200 nitrogen
mmol/L)

mI/h)
(85 mg/dl)

= = = = = =

patients

APACHE
17 potential

Ill

score

by recording

and

Biirubmn
Glucose Acid-base

(2.4 g/dl) (3.3 mg/cU) (246 mg/cU) (pH7.24;

measurements, age, and the chronic health evaluation. if multiple comorbid conditions are present, score only the one condition with the highest risk points (scoring is not performed with elective postoperative patients). Points for the 17 potential physiologic measurements reflect the worst (most abnormal) value during the initial 24 h of ICU treatment only. Daily updates of physiologic measurements also represent the most abnormal value within subsequent 24-h periods if a physiologic measurement is not obtained during this
physiologic

forthe

Pco2=26

mm

Hg)

Neurologic
obeys

(opens eyes, confused/converses,

verbal command)
= =

3points
107 points

Total
APACHE

III

first-day 4.5575) HI Score

hospital + Hospital (107

natural
X .0537

risk floor
=

equation:
admission
5.7459)=

Aspiration
(+ .2744)
+ 1.4628

pneumonia
+ APACJIE

(-

log

odds

initial

24-h

period,

no risk

points

are

assigned.

The

most

abnormal arterial blood gas measurement is the one associated with the widest P(A-a)O2 or the lowest PaO2. Ifa patient is heavily sedated and/or paralyzed, so that his neurologic status cannot be evaluated, and no reliable evaluation prior to sedation is available, the neurologic status should be recorded as normal.
Step 4

of death. The 4.318 r/(1-r). Solving ity=81.2 percent.

=

antilogarithm
=

for r

.8120,

of + 1.4628 = risk of hospital mortal-

Example

2: Operative
man

Admission
is admitted to the ICU

A 79-year-old

from

recovery
revealed Major

room following an exploratory laparotomy an obstructed colon secondary to colon cancer. Disease Category: GI obstruction room
(surgery

the that

(postoperative) was performed

Take the coefficients gory and treatmentlocation

with the total APACHE

for the patients major disease cateprior to ICU admission, together

III score, and use them as part of

Treatment on emergency
APACHE

Location: basis)

Recovery

the first-day APACHE III risk equation to calculate a predicted risk of hospital mortality. Copies of regression coefficients and detailed definitions for the 78 diagnostic categories and other components of the APACHE III equalion are available 1634 for research and independent confirmation

III Scoring:
=

Age (79 years) Chronic health

(colon cancer,
abnormalities

not metastatic)

17 points Opoints

Acute

physiologic

APACHE III Prognostic System (Knaus etaQ

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

Pulse Mean

rate (110

beats/mm)
(82 mm Hg)

5 points 0 points 0 points

Medicine.

Crit JE, DP

Care Knaus The

Med
use

1983; 14:466-72
Sharpe
JAMA

blood

pressure

10 Zimmerman EA, Wagner RWS. units.

WA, and

MD,
1986;

Anderson
of do not

AS,

Draper

Temperature
Respiratory

(37.5#{176}C) rate (10/mm ventilated)

. -=

implications

resuscitate

Opoints

PaO2/P(A-a)02
#{176}2

(Pa02=95

mm
ventilated;

Hg;
since 0 points
3 points

orders 11 Chang
care vors

mintensive
Individual Lancet WC,

care units.
outcome 2:143-46 PL, patterns 1989; Appel

255:351-56 models for intensive trial goals of sum-

prediction Waxman

F1o2 White

0.40, mechanically <0.5, use Pa02.

(32%)

12 Shoemaker

K. Clinical

cardiorespiratory JS, of acute subsets: DE, Caronna Dianost part

as therapeutic

in critically

Hematocrit
blood Creatinine

ill post-operative

patients.
C, myocardial

cell (1.8

count

(12,000/cu

mm)

0 points 4 points 4 points

13

Forrester therapy dynamic

Crit Care Chatterjee

infarction

Med 1982; 10:398-406 K, Swan HJC. Medical

by application 1976; 295:1356-62 RH, Frydman coma. design: Crit KR. due 16:1218-21 Failure ofAPACHE II H, JAMA analytic Care APACHE to cardiogenic Med II Lapinski of hemo-

Urine
Blood

output
Urea
(142

mg/dl) (1,800 mi/h) (22 mg/dl)

=
= = = = = =

I. N Engl

J
BH,

Med

Nitrogen
mmol/L)

points

14

Levy 1985; plan 1989;

JJ,

Singer

Sodium Albumin
Biirubin Glucose

0 points 0 points 5 points 0 points 0 points 0 points 45 points

Plum F. Predictingoutcome
253:1420-24 JE, evaluation AJ, Swinburne and mortality Crit CL, of ed. The 15 Zimmerman for

from APACHE

hypoxic-ischemic III study

(2.8 g/dl)
(2.5 (190
(pH

mg/cU) mg/dl)
7.42;

of severity AJ, Wah in respiratory Care

and GW,

outcome. Bisby

Acid-base
Neurologic obeys Total

Pco236
but

mm
opens

17(suppl):5169-221
failure 1988; VM.

Hg) eyes, and

16 Fedullo
score

(intubated commands)

pulmonary edema. 17 Hopefi AW, Taaffe

alone parenteral as a predictor

Med

Herrmann

APACHE +APACHE

III first-day

five obstruction
III

hospital risk equation: GI postopera4.6974) + Emergency surgery (+ .0752) Score (45x .0537= +2.4167)= -2.2051=

(-

18 Cerra
predict

FB,

nutrition. Negro

mortality in patients receiving total Crit Care Med 1989; 17:414-17 F, Abrams J. APACHE II score does not
organ failure or mortality in WA, acute JE. Crit models
N,

multiple

postoperative Wagner physiology (in press) APACHE Care with

Walker

log odds
0. 1102 tality
=
=

of death. The r/(1-r). Solving

percent.

natural antilogarithm of - 2.2051 = for r = .0992. Risk of hospital mor-

surgical
19 Damiano Reliability

patients.
AM,

Arch Surg
Bergner M,

1990;
Draper II.

125:519-22
EA, Knaus DP of illness:

9.92

of a measure

of severity Wagner DP,

and chronic

health
Draper

evaluation:
EA,

J Clin

Epidemiol system.

Comment Predictions
variable

20

Knaus

WA,

Zimmerman

after
updates

the
the

initial
and are article. discuss
Within

day

contain
APACHE

an additional
III score for on 21 contributions different

II: A severity

of disease 5, Simon

classification

Med cubic

that and the the

patients also

1985; 13:818-828
Durrleman

R. Flexible

regression

changes
of age

in physiologic
comorbidifies

status,

the relative slightly

subsequent
beyond contact
experimental

days.
scope authors

Further
of this to

explication

of the
Research

equations
teams or should collateral
and

is

Stat Med 1989; 8:551-61 22 Daley J, Jencks S. Draper D, Lenhart Predicting hospital-associated mortality
a method infarction, 24 for patients and 5, Daley with stroke, heart congestive failure.

splines.

G, Thomas for Medicare acute

J.

patients: myocardial 260:3617-

pneumonia, JAMA N,

replication
accepted

1988;

approaches.

experimental

ethical
tional

protocols,
materials

the
as they

authors

plan

to make

available

addi-

23 Jencks

J, Draper

D, Thomas data:

are developed.
REFERENCES

Interpreting

hospital

mortality

the

Lenhart C, Walker J. role of clinical risk

E, Jacobsen outcome

adjustment. JAMA 1988; 260:3611-24 24 Dragsted L, Jorgensen J, Jensen NH,

Knaus WA, intensive et al. Interhospital care: 17:418-22 importance MA. Admission comparisons for clinical Ann medicine: Intern Med from

Bansing

5,

of patient

1 Feinstein
I. The

AR.

An additional

basic

science paradigms.

of lead-time
source

bias.
to the

Crit Care
medical

constraining

fundamental

Med 1989; 25 Escarce intensive

APACHE
26

1983; 99:939-97 2 Wasson JH, Sox HC, NeffRK, Goldman L. Clinical prediction rules: applications and methological standards. N Engl J Med 1985; 313:793-99 3 Silverstein MD. Prediction instruments and clinical judgement
in critical care. JAMA D, 1989; APACHE, systems. 1988; 260:1758-59 PE. patient MPM, Intensive Clinical outcome Care versus from actuarial intensive and other care 245:168-74

JJ, Kelley care unit

II score. Fisher Sepsis al.

predicts
JAMA 1990;

hospital
264:2389-94

death

independent
GJ, Metz entity. CA, Crit

of
Balk Care epide-

Bone RA,

RC, et

CJ, Clemmer syndrome: DR

TP, Slotman a valid clinical

4 Dawes RM, Faust judgement. Science

5 Seneff M , Knaus care: a guide to WA.

Meehi

Med 1989; 17:389-93 27 Knaus WA, Wagner

miology 28 Solomkin and JS, prognosis. Dellinger trial

Multiple Crit EP, Care Christou

systems Clin 1989;

organ

failure:

5:221-32

Predicting

NU,

Busuttil infections.

RW Results
to tobraAnn Surg

SAPS, strategies AG, Thibault

PRISM,

of a multicenter

comparing

imipenem/cilastatin

prognostic

scoring

Med

1990; 5:33-52

6 KaIb PE, Miller DH. Utilization units. JAMA 1989; 261:2389-95

7 Detsky AS, and Stricker the care SC, expenditure unit. Malley N Engi

for intensive GE.

mycin/clindamycin 1990; 212:581-91 29 Ziegler

E, Fisher ha-ia human 1991; of N EngI WA, et al. Treatment with EngI

for

intra-abdominal
C, Straube antibody Geston 316:134-37 DP, PR. care:

C, Sprung monoclonal 394:429-36 UE,

R, SadoffJ, and against Accurate a new

Prognosis, for patients its 1990; Care 31 30

of gram-negative

bacteremia

survival,

of hospital

resources

Foulke CE, septic shock endotoxin. N

prediction quantitative

in an intensive

8 Schneiderman
meaning and Consensus

U,
ethical

Jecker

J Med 1981; 305:667-72 NS, Jonsen AR. Medical futility:

Ann Conference Intern on Med Critical

J Med
MM, outcome

Pollack of method.

Ruttimann pediatric EA,

implications.

intensive Wagner

112:948-54
9 NIH Development

J Med 1987;

Knaus

Draper

Zimmerman

JE.

An

CHEST

I 100 I 6 I DECEMBER,

1991

1635

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

evaluation

of

outcome

from

centers.
32 Knaus
Campos

Ann Intern WA, LeGall

BA, Lancet

Med
TR, 1982;

intensive care in major medical 1986; 104:410-18 Wagner DP, Draper EA, Loirat P.
ofintensive care in the U.S.A.

et al. A comparison

and France. 33 Reynolds

Impact shock 34 in a

HN,
ofcritical
university

2:642-44 Haupt MT, Thill-Baharizian MC, Carison RW care physician staffing on patients with septic hospital medical intensive care unit. JAMA
D, Avrunin SJ. Gage RW Refining

house officers prognostic judgement Ihr critically ill Arch Intern Med 1990; 150:1874-78 39 Englehardt LIT, Rie MA. Intensive care units, scarce resources, and conflicting principles ofjustice. JAMA 1986; 255:1159-64 40 Knaus WA, Bauss A, Alperovitch A, LoGaII JR. Loirat I Patois E, et al. Do objective estimates ofcbances for survival influence decisions to withhold or withdraw teatment? Med Decis Making

averaging

patients.

1990;

10:163-71

1988; 260:3446-50
Lemesbow
care

S. Teres

intensive

41 Knaus WA, Wagner DP, Lynn D. Short-term for critically ill hospitalized adults: science
(in press) 42

mortality estimates and ethics. Science

35

36

37

38

unit outcome prediction by using changing probabilities of mortaht Crit Care Med 1988; 16:470-77 Kruse JA, Thill-Baharizian MC, Carlson RW Comparison of clinical assessment with APACHE II for predicting mortality risk in patients admitted to a medical intensive care unit. JAMA 1988; 260:1734-42 Brannen AL, Godfrey U, Coetter WE. Prediction of outcome from critical illness: a comparison of clinical judgement with a prediction rule. Arch Intern Med 1989; 149:1083-86 MCCIISh DK, Fbwell SH. How well can physicians estimate mortality in a medical intensive care unit? Med Decis Making 1989; 9:125-32 Ebses RM, Bekes C, Winkler RL, Scott WE, Copare FJ. Are two (inexperienced) heads better than one (experienced) head?

Chang RWS, Jacobs 5, Lee B. Predicting outcome among intensive care unit patients using computerized trend analysis of daily APACHE II scores corrected for organ system failure.
Intensive Care Med 1988; 14:558-66

43 MoskowitzAJ, Kuipers BT, KassirerJP DealingWith uncertaint risks and trade-offs in clinical decisions: a cognitive science approach. Ann Intern Med 1988; 108:435-49 44 DetSlCy AS, Redelmeier D, Abrams HB. Whats wrong with decision analysis? can the left brain influence the right? J
Chronic Dis 1987; 40:813-16 study to understand and risks of treatments.

45 Murphy

DJ, Cluff LE, eds. SUPPORT: prognoses and preferences for outcomes J Chin Epidemiol 1990; 43(suppl):1S-123S

Plan to Attend ACCPs

58th Annual Scientific Assembly Chicago October 25-29, 1992

1636

APAQIE

III Progno

Syem

(aus

etal)

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians

The APACHE III prognostic system. Risk prediction of hospital mortality for critically ill hospitalized adults WA Knaus, DP Wagner, EA Draper, JE Zimmerman, M Bergner, PG Bastos, CA Sirio, DJ Murphy, T Lotring and A Damiano Chest 1991;100;1619-1636 DOI 10.1378/chest.100.6.1619 This information is current as of November 3, 2007
Updated Information & Services Citations Permissions & Licensing Updated information and services, including high-resolution figures, can be found at: http://chestjournal.org This article has been cited by 100 HighWire-hosted articles: http://chestjournal.org Information about reproducing this article in parts (figures, tables) or in its entirety can be found online at: http://chestjournal.org/misc/reprints.shtml Information about ordering reprints can be found online: http://chestjournal.org/misc/reprints.shtml Receive free email alerts when new articles cite this article sign up in the box at the top right corner of the online article.

Reprints Email alerting service

Images in PowerPoint format Figures that appear in CHEST articles can be downloaded for teaching purposes in PowerPoint slide format. See any online article figure for directions.

Downloaded from chestjournal.org on November 3, 2007 Copyright 1991 by American College of Chest Physicians