Support Care Cancer (2003) 11:686695

DOI 10.1007/s00520-003-0498-9

Jeffrey S. Groeger
Jill Glassman
David M. Nierman
Susannah Kish Wallace
Kristen Price
David Horak
David Landsberg

Received: 21 January 2003

Accepted: 28 May 2003
Published online: 5 August 2003
Springer-Verlag 2003
Funding for initiation and coordination of
this project was provided by Memorial
Sloan Kettering Cancer Center. Participating institutions providing resources for sitespecific data collection. This paper was presented in part at the annual meeting of the
American Society of Clinical Oncology in
May, 2001 held in San Francisco, CA, USA
J. S. Groeger () D. Landsberg
Critical Care Medicine Service,
Memorial Sloan Kettering Cancer Center,
New York, NY, USA
e-mail: GroegerJ@MSKCC.org
Tel.: +1-212-6398114
Fax: +1-646-4222237
J. Glassman
Biostatistics Service,
Memorial Sloan Kettering Cancer Center,
New York, NY, USA
D. M. Nierman
The Mount Sinai Medical Center,
New York, NY, USA
D. Horak
City of Hope National Medical Center,
Duarte, CA, USA
S. K. Wallace K. Price
The University of Texas
M. D. Anderson Cancer Center,
Houston, TX, USA

O R I G I N A L A RT I C L E

Probability of mortality of critically ill cancer

patients at 72 h of intensive care unit (ICU)
management

Abstract Goals: To develop and

validate a model for probability of
hospital mortality for cancer patients
at 72 h of intensive care unit (ICU)
management. Patients and methods:
This is an inception cohort study
performed at four ICUs of academic
medical centers in the United States.
Defined continuous and categorical
variables were collected on consecutive patients with cancer admitted to
the ICU. A preliminary model was
developed from 827 patients and
then validated on an additional 415
patients. Multiple logistic regression
modeling was used to develop the
models, which were subsequently
evaluated for discrimination and calibration. The main outcome measure
is in-hospital death. Results: A probability of mortality model, which incorporates ten discrete categorical
variables, was developed and validated. All variables were collected at
72 h of ICU care. Variables included
evidence of disease progression, performance status before hospitalization, heart rate >100 beats/min,
Glasgow coma score 5, mechanical
ventilation, arterial oxygen pressure/fractional inspiratory oxygen
(PaO2/FiO2) ratio <250, platelets

Introduction
Models such as the Acute Physiology and Chronic
Health Evaluation II (APACHE) [28], Mortality Probability Model II (MPM II) [30], and the Simplified Acute

<100 k/l, serum bicarbonate

(HCO3)<20 mEq/l, blood urea nitrogen (BUN) >40 mg/dl, and a urine
output of <150 ml for any 8 h in the
previous 24 h. The p values for the
fit of the preliminary and validation
models were 0.535 and 0.354 respectively, and the areas under the receiver operating characteristic
(ROC) curves were 0.809 and 0.820.
Conclusions: We report a multivariable logistic regression model to
estimate the probability of hospital
mortality in critically ill cancer patients at 72 h of ICU care. The model
is comprised of ten unambiguous
and readily available variables.
When used in conjunction with clinical judgment, this model should improve discussions about goals of care
of these patients. Additional validation in a community hospital setting
is warranted.
Keywords ICU Cancer
Mortality Probability Models

Physiology Score II (SAPS) [29] have been shown to underestimate mortality rates of cancer patients admitted to
intensive care units (ICUs) [19, 49]. We have previously
reported a cancer-specific multivariable logistic regression model to estimate the probability of hospital mortal-

687

ity in critically ill cancer patients who are admitted to an

ICU [19]. While severity-of-illness models apply to cohorts of patients and not individuals, having a numerical
probability of survival may help foster dialogue to formulate goals for duration and intensity of ICU care.
Once a patient is admitted to the ICU, regardless of
whether they have an underlying malignancy, it is appropriate that clinical goals of care be reassessed on a regular basis. Decisions about ongoing support of critically
ill patients with cancer are formed by the prognosis of
the underlying malignancy, the patients clinical status
after a period of aggressive ICU care, and the wishes of
the patient. While many patients may improve or die in
the first 72 h of ICU management, a large percentage
will remain critically ill. With that in mind, we present
the results of a multicenter project to develop a model
for predicting survival of cancer patients remaining in
the ICU after 72 h of care. The value of this information
lies in its ability to objectify a situation for caregivers
and families, as well as for patients, when decisions of
the human cost of an intervention are being weighed
against its potential benefits. While no model is capable
of predicting the outcome of any individual patient, a
well-validated model can provide a sometimes-needed
perspective.

Patients and methods

Data were prospectively collected in four academic tertiary care
hospitals beginning 1 July 1994 to develop and validate a multivariable logistic regression model to estimate the probability of
hospital mortality among cancer patients admitted to the ICU. Participating units were the Medical/Surgical ICU of Memorial Sloan
Kettering Cancer Center (MSKCC), New York, NY, USA; the
City of Hope National Medical Center, Duarte, CA, USA; the
Medical ICU of The University of Texas, M.D. Anderson Cancer
Center, Houston, TX, USA: and the Mount Sinai Medical Center,
New York, NY, USA. The study was approved by the Institutional
Review Board of all participating sites.
All cancer patients admitted to the ICU were included, with the
exception of patients younger than 18 years, burn patients, and
coronary patients defined by a primary diagnosis of myocardial infarction (MI) or rule-out MI with no secondary diagnosis. The outcome of interest was vital status at hospital discharge. For patients
with multiple admissions to the ICU during the study period, the
most recent admission was used in the analysis.
Demographic, clinical, laboratory, and physiological variables
were obtained on consecutive cancer patients within 1 h of ICU
admission, at 24 and 72 h of ICU admission, and at ICU and hospital discharge. No specific clinical interventions were performed
on any patient for the purpose of developing the outcome model.
All categorical variables were defined prior to patient accrual,
made available to each data collector, and were previously reported [19].
Patients were classified as being in one of four tumor groups:
solid tumor, solid tumor with metastasis, leukemia or hematopoietic bone or peripheral stem cell transplant, and lymphoma or myeloma. Patients with leukemia, bone marrow transplant, lymphoma, or myeloma were classified as having nonmetastatic disease.
Patients were classified as either medical or surgical, with surgical
patients being those who had a surgical procedure prior to ICU ad-

mission during the same hospitalization. Patients with leukemia,

bone marrow transplant, lymphoma, or myeloma who had undergone open lung biopsy for diagnosis of the etiology of respiratory
insufficiency were classified as medical patients.
Patients were randomly divided into a model development set
(2/3) and a model validation set (1/3). Analyses were conducted to
examine the distribution of each variable. For variables for which
no specific data were recorded for a patient, a response of No or
Never or Normal was imputed if the variable was categorically scaled, and a value within normal limits was imputed if the
variable was continuously scaled. Univariate analyses were performed to test the significance of each variable in relation to vital
status at hospital discharge, using chi-square tests for categorical
variables and t tests for continuous variables. All data are presented as means plus or minus standard deviation unless otherwise
noted.
Upon completion of univariate analyses, variables collected at
admission, 24 hours and 72 h were eligible for entry into a logistic
regression model if their significance was less than or equal to
0.10, as well as entered or omitted by clinical evaluation (e.g., if
p0.10 but the variable was clinically important/easy to use or
p<0.10 but differences in means was not clinically meaningful or
variable infrequently collected). Preliminary models were evaluated by way of ad hoc sensitivity analysis, running stepwise logistic
regression procedures on the development database under different combinations of the following conditions: (1) substituting
normal for all missing categorical variables, (2) substituting either normal or means for missing continuous variables, (3) eliminating all continuous variables with more than one third of values
missing. This process was formalized into a bootstrap validation
procedure from which a new core group of variables emerged [9,
35, 45]. Using the appropriate dummy coding or transformations
determined for each variable, all of these variables were entered
into a logistic regression model. The significance of each variable
in the multivariate context was assessed, and variables were then
deleted, one at a time, until all variables in the model were significant, with significance defined as p 0.05.
Once the list of variables to be used in our final model was
selected, the functional form of each variable was examined. Optimal data-driven cut points were sought using exploratory plots and
by examining all possible cut points that could dichotomize a variable for the one that best separated the mortality risk groups according to the result of the chi-square statistic [33].
To evaluate models, assessments of both calibration and discrimination were made.
The Hosmer-Lemeshow [24] goodness-of-fit test was used to
evaluate model calibration. This tested the ability of the probabilities based on the model to generate expected numbers of patients
who did or did not survive to hospital discharge that closely reflected the observed numbers of such patients across several strata
of probabilities. This was done both for the development and validation data sets. The chi-square value was calculated in the usual
manner using the observed and expected frequencies in each cell.
A low value of the statistic, and a corresponding high p value, indicated that there was good agreement between the observed and
expected number of patients across all probability levels. To test
the models fit in the data set in which the model was developed,
the degrees of freedom were eight. When testing the models fit in
the independent validation data set, the degrees of freedom were
ten. For best calibration, final model coefficients were recalibrated
using a shrinkage factor, as described by Harrell [23].
The area under (AUC) the ROC curve [22] was calculated to
evaluate discrimination. This tested the models ability to discriminate between patients who lived and patients who died by
determining the percent of time that the probability of mortality
was higher for patients who died than for patients who lived,
among all possible pairings of such patients. An AUC greater
than 70% was generally considered to be evidence of good model
discrimination.

688

Results
Prospective data were collected on 3,005 consecutive patients with a malignancy diagnosis or hematopoietic
stem cell transplantation (HSCT) admitted to the ICUs of
participating institutions. At 72 h, 1,354 patients remained in the ICU. After eliminating all but the most recent admission of patients with multiple admissions
(n=20) and those admitted with a diagnosis of rule-out
MI, 1,242 patients were eligible for entry into model development and validation. Data were collected from 1
July 1994 through 30 June 1998.
Patients were randomly divided into a development
set (two thirds: n=827) and a validation set (one third:
n=415). Hospital patient mortality in the development
set was 54% (447 deaths, 380 survivors) with an average
age of 55.616.5 (median 56.8) years. These patients
were hospitalized an average of 10.023.0 (median 5.0)
days prior to ICU admission.
Variables selected for entry into preliminary stepwise
logistic regression models following univariable analysis
and clinical judgment are listed in Table 1. All preliminary modelswhether using normal for all missing
categorical variables, substituting either normal or
means for missing continuous variables, eliminating
all continuous variables with more than one third of values missing, or by including all admission disease-related categorical variables and then only data collected at
72 hrevealed acceptable calibration and discrimination
with the lowest p value for goodness of fit equal to 0.256
as well as an AUC equal to 0.818.
Based on bootstrap results and our preliminary stepwise logistic regression analysis, a model was generated
where normal values were substituted for missing continuous data elements and normal or never for missing
categorical variables. Based on clinical judgment, only
variables measured at 72 h were entered into the model,
along with initial admission categorical data. Results of
selected analyses of categorical and continuous data are
presented in Tables 2 and 3.
After clinical review, cut points were used to categorize heart rate (>100 beats/min), Glasgow coma score
(5), blood urea nitrogen (BUN) (>40 mg/dl), arterial
oxygen pressure/fractional inspiratory oxygen (PaO2/
FiO2) (<250), serum bicarbonate (<20 mEq/l) and platelet count (<100 k/l) into dichotomous variables. Note
also that there were three categories of the performance
status variable: Patients were coded either as having a
good performance status (Zubrod 0 or 1), assistance
required (Zubrod 2 or 3), or bedridden (Zubrod 4). Two
variables that were statistically significant in the regression analysistotal body irradiation and duration of hospitalization prior to ICU admissionwere entered as
continuous and dichotomized variables respectively;
however, they did not add to the robustness of the final
model and were eliminated.

Table 1 Variables entered into development model after preliminary statistical and clinical analysis. BUN blood urea nitrogen,
LDH lactic dehydrogenase, PaO2 /FiO2arterial oxygen pressure/
fractional inspiratory oxygen, PT prothrombin time, PTT partial
thromboplastin time, ECOG Eastern Cooperative Oncology Group
Laboratory evaluations
Albuminadmission, 24 h
Bilirubinadmission, 24 h, 72 h
BUNadmission, 24 h, 72 h
Creatinine24 h, 72 h
Serum bicarbonate24 h, 72 h
Lactate24 h, 72 h
LDH24 h, 72 h
Sodium72 h
Potassium24 h
PaO2/FiO2,admission, 24 h, 72 h
Absolute neutrophil count24 h
Plateletsadmission, 24 h, 72 h
PTadmission, 24 h
PTTadmission
Vital signs
Heart rateadmission, 24 h, 72 h
Respirationsadmission, 72 h
Blood pressure72 h
Clinical variables
Tracheal intubation with mechanical ventilationadmission, 24 h,
72 h
Urine output <150 ml over any 8 hr in preceding 24 h24 h, 72 h
Vasopressor or inotropic agents24 h, 72 h
Urine output/24 h24 h, 72 h
Glasgow coma scoreadmission, 24 h,72 h
Acute renal failureadmission
Dialysis24 h, 72 h
Pulmonary artery catheter24 h, 72 h
Cardiopulmonary resuscitation24 h, 72 h
Disseminated intravascular coagulation24 h, 72 h
Cancer-related variables
Allogenic or autologous transplant
Zubrod (ECOG) performance status prior to hospitalization
Recent chemotherapy
Prior surgical history
Total body irradiation
Tumor group
Disease progression or recurrence

The final model is presented in Table 4. The coefficient for each variable and its associated standard error,
as well as the corresponding odds ratio and 95% confidence interval for the odds ratio, are shown.
Evaluation of the models goodness-of-fit in the developmental sample is presented in Table 5, which
shows the correspondence between observed and expected mortality based on the model probabilities with a
Hosmer-Lemeshow chi-squared statistic of 7.01 and corresponding p=0.525. The AUC was 0.809.
Calibration and discrimination of the model in a validation sample of 415 patients is shown in Table 6. The
fit was very good in the validation sample, with a Hos-

689

Table 2 Results of selected univariate analyses on random two thirds development data set of cancer patients in intensive care unit
(ICU) for 72 h.NED no evidence of disease, TBI total body irradiation, Y yes, N no,
Died
Number
Tumor
Leukemia
Lymphoma/myeloma
Solid metastatic
Solid

Lived
Percent

P value

Number

Percent

169
77
112
82

66
57
52
41

87
59
103
114

34
43
48
58

0.001

Hematopoietic stem cell transplantation

Allogenic
Autologous
Never
Chemotherapy
>4 weeks ago
24 weeks
<2 weeks
Never

88
24
332

73
47
51

32
27
315

27
53
49

0.001

142
50
134
112

61
57
59
45

91
38
94
137

39
43
41
55

0.003

Cirrhosis/veno-occlusive disease
Y
N

9
436

69
54

4
369

31
46

0.279

Coronary artery disease

Y
N

44
401

49
55

46
327

51
45

0.266

Chronic obstructive pulmonary disease

Y
N
Cancer status
Progression
New diagnosis
NED

78
367

57
54

60
312

43
46

0.595

220
127
93

60
52
48

147
119
101

40
48
52

0.014

Acute renal failure

Y
N

59
386

65
53

32
341

35
47

0.034

Heart failure
Y
N

32
413

53
54

28
345

47
46

0.863

ICU in past 6 months

Y
N

41
404

64
54

23
350

36
46

0.106

Insulin-dependent diabetes mellitus

Y
N

16
429

46
55

19
354

54
45

0.292

Comaa
Y
N

98
344

83
50

20
348

17
50

0.001

Cardiopulmonary resuscitationa
Y
N

3
439

100
54

0
368

0
46

0.113

Dialysisa
Y
N

34
406

68
54

16
352

32
46

0.047

690

Table 2 (continued)
Died
Number

Lived
Percent

P value

Number

Percent

Do not resuscitatea
Y
N

54
387

78
52

15
353

22
48

0.001

Intubateda
Y
N

317
125

71
34

127
241

29
66

0.001

Eastern Cooperative Oncology Group

Bedridden (4)
Assistance (23)
Normal/symptoms (01)

19
111
314

79
59
52

5
77
288

21
41
48

0.012

Surgical history
Palliative
Curative
None this admission

37
42
361

46
42
58

43
58
263

54
58
42

0.004

Limits placed on carea

Y
N

48
394

83
52

10
358

17
48

0.001

Radiation
TBI
Isolated port
None

80
103
258

75
53
51

27
91
245

25
47
49

0.001

Gender
Male
Female

265
177

56
53

211
158

44
47

0.424

Urine outputb
Y
N

80
359

77
51

24
339

23
49

0.001

Vasopressorsa
Y
N

126
316

70
50

53
313

30
50

0.001

a 72 h
b Urine

output <150 ml/any 8-h period in preceding 24 h

mer-Lemeshow statistic of 11.03 and p=0.354. The AUC

was 0.820 in the validation sample.
Using the model coefficients from Table 4, the data
for a cancer patient still in the ICU at 72 h can be used to
calculate the logit as
where D is the constant and ixi is the estimated coefficient for the ith variable times the value of the ith variable, with i taking on values from 1 to k, and k being the
number of terms in the model.
Once the logit has been obtained by multiplying each
appropriately coded variable value by its corresponding
coefficient and summing, a simple calculation using the
logit produces the probability of hospital mortality for
the patient as follows:

To illustrate, Table 7 is provided to calculate the probability of hospital mortality for a hypothetical cancer patient at
72 h of ICU care. In this example, the patient is admitted
to the ICU with evidence of disease progression and requiring some assistance for daily functioning. The patient
was intubated and mechanically ventilated. The heart rate
was 109, Glasgow coma score 12, PaO2/FiO2 ratio
178 mmHg, platelet count 85,000, BUN 19 mg/dl, and serum bicarbonate (HCO3) 19 mEq/dl. Urine output was between 10 and 50 ml/h over the past 24 h; however, at no
time was the urine output <150 ml for any 8-h period.
For each variable in the model, the coefficient is
shown under the column in Table 7 labeled , and the

691

Table 3 Mean, and standard

deviation (SD) of select continuous variables for patients who
lived and patients who died.
All collected at 72 h.
PaO2 /FiO2 arterial oxygen
pressure/ fractional inspiratory
oxygen, PT prothrombin time,
INR international normalised
ratio, PTT partial thromboplastin time, LDH lactic dehydrogenase, BUN blood urea nitrogen, HCO3 serum bicarbonate

Table 4 Coefficients, standard

errors (SE) of the coefficients,
odds ratios, and 95% confidence intervals (CI) for the
odds ratios for model variables.
PaO2 /FiO2 arterial oxygen
pressure/fractional inspiratory
oxygen, BUN blood urea nitrogen

Table 5 Hosmer-Lemeshow
goodness-of-fit table for the
model in the developmental
sample (n=827)

* Predicted probability of mortality degrees of freedom=8

Hosmer-Lemeshow (H-L), chisquared statistic=7.01, p=0.535

Variable

Died

Age (years)
Hospital days prior to ICU
Months since cancer diagnosis
Glasgow coma score
Heart rate (beats/min)
Respiratory rate (breaths/min)
Systolic blood pressure (mmHg)
PaO2/FiO2 ratio
White blood cell count (1,000)
Platelets (1,000)
PT (sec)
INR
PTT (sec)
Bilirubin (mg/dl)
LDH (IU)
Albumin (g/dl)
Creatinine (mg/dl)
BUN (mg/dl)
Urine output (l/24 h)
Sodium (mEq/l)
Potassium (mEq/l)
HCO3 (mEq/1)

Variable

Performance status
Assistance required (Zubrod 2 or 3)
Bedridden (Zubrod 4)
Heart rate >100 beats/min
Glasgow coma score 5
Mechanically ventilated
PaO2/FiO2 <250
Platelets <100 k/l
Bicarbonate <20 mEq/L
BUN >40 mg/dl
Urine output <150 ml/any 8 h
Constant

1
2
3
4
5
6
7
8
9
10

Prob*

0.12
0.20
0.30
0.39
0.51
0.60
0.69
0.79
0.87
0.95

P value

Mean

SD

Mean

SD

54.9
13.2
25.6
12
103
19
121
202
9.97
90.7
15.3
1.48
33.2
4.03
1680
2.66
1.83
43
2.710
139
3.9
24.1

16.4
27.8
40.6
4
21
7
24
104
12.32
96.3
11.1
0.71
11.7
4.76
3252
0.52
1.43
28
2.079
6
0.6
5.4

56.4
8.3
31.7
14
96
20
125
269
10.57
148.1
17.5
1.32
32.3
2.57
1084
2.74
1.32
27
3.046
138
3.9
26.7

16.4
15.4
55.8
2
20
7
22
122
11.30
123.7
21.7
0.49
13.5
5.56
1857
0.52
1.20
24
1.888
5
0.5
8.4

Coefficient

Evidence of disease progression

Group

Lived

0.21
<0.01
0.002
<0.001
<0.001
0.11
0.01
<0.001
0.48
<0.001
0.13
0.003
0.44
0.004
0.02
0.13
<0.001
<0.001
0.02
0.003
0.67
<0.001

SE

Odds ratio

95%

0.4535

0.1693

1.57

1.13

2.19

0.4775
1.4949
0.4012
1.3178
1.1764
0.9131
0.8886
0.5281
0.6609
0.8293
2.3854

0.2009
0.5754
0.1671
0.5172
0.1739
0.1729
0.1757
0.2565
0.1927
0.2863
0.2200

1.61
4.46
1.49
3.74
3.24
2.49
2.43
1.70
1.94
2.29

1.09
1.44
1.08
1.41
2.31
1.78
1.73
1.03
1.33
1.32

2.40
13.77
2.07
10.29
4.57
3.50
3.44
2.80
2.83
4.02

Died

Lived

CI

Total

Observed

Expected

Observed

Expected

9
20
31
29
37
49
57
67
71
77

9.7
16.2
26.9
32.7
44.1
49.5
58.5
62.5
71.9
76.3

72
59
59
54
50
31
27
12
12
4

71.3
62.8
63.1
50.3
42.9
31.5
25.5
16.5
11.1
4.7

81
79
90
83
87
80
84
79
83
81

692

Table 6 Hosmer-Lemeshow
(H-L) goodness-of-fit table for
the model in the validation
sample (n=415)

* Predicted probability of mortality degrees of freedom=10

Hosmer-Lemeshow (H-L),
chi-squared statistic=11.03,
p=0.354
Table 7 Calculations for each
variable used to generate the
probability of hospital mortality for a hypothetical cancer patient at 72 h X observed variable value for the patient,
x coded model value of the
variable for the patient.
PaO2 /FiO2 arterial oxygen
pressure/fractional inspiratory
oxygen, BUN blood urea
nitrogen

* The value of the constant is

applied to all patients in calculating the logit

Group

1
2
3
4
5
6
7
8
9
10

Prob*

0.11
0.20
0.27
0.35
0.43
0.52
0.60
0.69
0.79
0.89

Died

Lived

Total

Observed

Expected

Observed

Expected

4
3
16
12
15
23
24
31
35
40

4.6
8.1
11.0
14.4
17.5
21.1
24.7
28.3
32.2
41.0

37
38
25
29
26
18
17
10
6
6

36.4
32.9
30.0
26.6
23.5
19.9
16.3
12.7
8.8
5.0

Variable
Evidence of disease progression
Performance status
Assistance required(Zubrod 2 or 3)
or Bedridden (Zubrod 4)
Glasgow coma score 5
Mechanically ventilated
PaO2/FiO2 ratio <250
Heart rate> 100 beats/min
Platelets <100 k/l
Serum bicarbonate <20 mEq/l
BUN >40 (mg/dl)
Urine output <150 ml/any 8 h
Constant
Logit: g(x)=0+ixi, I=1,...k
Predicted probability of mortality

observed value for the hypothetical patient is entered in

the table under the column labeled X. The code or value
to be used in the calculations is shown under the column
labeled x, and the result of multiplying the code or value
by the coefficient is shown under the column labeled ..
The values of
are summed to obtain the logit, so that
g(x)=2.453. The patients probability of hospital mortality is then calculated as:

Of 100 patients having this profile, 92 would be expected by the model to die prior to hospital discharge and
eight would be expected to be discharged from the hospital alive.

Discussion
We have previously presented a model to assess probability of hospital mortality in cancer patients admitted to
the ICU on the basis of variables readily obtained on admission [19]. We now apply the same principles to development and validation of a 72-h model distinguishing

41
41
41
41
41
41
41
41
41
46

0.4535

Yes

0.4535

0.4775
1.4949
1.3178
1.1764
0.9131
0.4012
0.8886
0.5281
0.6609
0.8293
*.2.3854
2.4530
0.92

Yes
No
12
Yes
178
yes
85
19
19
No

1
0
0
1
1
1
1
1
0
0

0.4775
0.00000
0.00000
1.1764
0.9131
0.4012
0.8886
0.5281
0.00000
0.00000
*2.3854

those variables now of greater or lesser importance, as

defined by our development cohort and adjusting the coefficients of our model to reflect these new weights. We
were motivated to develop this model as it reflects a period of time for clinicians to attempt to reverse a lifethreatening complication, for patients and families to adjust to and reflect on the global impact of a potentially
lethal change in clinical status, as well as to offer time
for the critical care staff to develop a rapport in order to
comfort and counsel family regarding goals of care.
The question of who will live and who will die is the
fundamental query being addressed by these models. It
has been shown that experienced clinicians fare quite
well against calculations previously designed to answer
this question [32]. Cancer patients admitted to the ICU
for critical illness continue to experience mortality rates
higher than their non-cancer-bearing, severity-of-illnessscored matched cohorts [53]. Published mortality rates
range from 50 to 80% in this population [3, 7, 12, 13, 14,
15, 16, 17, 20, 25, 40, 42, 46, 47, 48, 54, 55]. We observed a 54% mortality rate in our development group,
falling in line with the above ranges; however, our cohort is different, as they have already survived 72 h of
ICU care. Table 2 demonstrates mortality-rate differences among tumor categories, with hematologic malig-

693

nancies experiencing a 63% mortality rate as compared

to that of solid tumors, which shows a rate of 46%. This
difference is, in part, accounted for by the well-documented, discordantly high though improving, mortality
rate associated with allogeneic bone marrow transplant
patients requiring ICU admission and, especially, mechanical ventilation [1, 4, 13, 14, 15, 16, 20, 25, 41, 42,
43, 44, 54, 55].
Much work has surfaced in recent years looking at
prognosis in cancer patient and the possibility of objectifying the mystique that surrounds this diagnosis. Scoring
systems have been developed to aid in this task but generally lack the sensitivity and specificity required to answer
such complex questions [8, 11, 27, 34, 49, 53]. Some investigators have looked at the applicability of standard
scoring systems to this venue. It has been shown in recent
work that the probability of ICU mortality is more closely
associated with the acute physiologic processes necessitating ICU admission than the underlying cancer diagnosis or its stage [19, 21, 49]. In support of this contention,
one study concluded that APACHE II and SAPS II predicted hospital and ICU mortality well, but the writers
qualified this assertion by stating that it did not predict
these endpoints well enough to be useful in management
[49]. Many other authors have repeatedly demonstrated
the inability to interchange scoring systems between populations in whom the system was not originally developed [2, 5, 31, 36, 37, 51, 56]. It is on the basis of this assertion that we present this cancer-specific, 72-h model
so that those patients who remain in the ICU may be restratified at 3 days according to their new health status.
The 72-h time interval was chosen as it is a commonly
accepted time frame for trials of critical care for those
patients whose indications for ICU admission may not be
as clear. This is often related to questions of a
patients severity of illness or prognosis. The authors endorse the idea that survival should always be given the
benefit of the doubt in these patients, and it is this principle that is at the root of allowing patients to declare
themselves independent of any model that identifies a
probability of mortality upon ICU admission. The
thought that different variables would emerge as important after a trial of critical care is supported by our results.
Tumor burden, as manifested by recurrent or refractory
disease or with a poor performance status, purports a poor

prognosis for the cancer patient. These findings are supported by previously documented prognostic factors for
survival in advanced non-small-cell lung cancer, with extent of disease and Karnofsky score listed as the primary
and secondary determinants of survival [38]. A separate
study in small-cell lung cancer also revealed extent of disease as the most important prognostic factor and additionally found that Karnofsky score was the only predictive independent variable as time passed [39]. This mirrors quite
well our findings at 72 h as well as at ICU admission.
Presence and degree of respiratory failure continues
to be an important variable at 72 h, as manifested by an
increased risk of dying if the patient is intubated for any
reason as well as having a low PaO2/FiO2. Renal dysfunction is strongly supported by the literature as a valuable prognosticator of outcome and remains so in our 72h model, with the addition of a new parameter to not only reflect azotemia but also urinary flow as defined by
the presence of outputs of less than 150 cc for any 8-h
period [6, 10, 18, 26, 50, 52]. Tachycardia and metabolic
acidosis were not found to be discriminators on admission, but those patients who remained tachycardic or had
an HCO3 20 mEq/dl at 72 h were more likely to die.
Naturally the continued presence of these findings at
72 h describes a patient with persistently altered hemodynamics and cellular metabolism, portending a worse
outcome than the patient not so altered.
This new model used alone or in concert with the previously validated admission model may serve to better
characterize the probable course for the majority of patients stratified. This has great utility in discussions with
patients and their families to try and quantitate for them
what their chances might be. Further, these models
may prove to be very useful in processes related to ICU
bed utilization and review. These models have no validity as a criterion to deny admission for any individual patient. The authors hope that this information might be
used to better clinical care delivered to the ICU patient
with cancer by more accurately and more rapidly identifying the patient at high risk of death so that their care
might be handled most appropriately, as dictated by the
individual clinical scenario.
Acknowledgment We are indebted to Peter B. Bach for his review and assistance with this manuscript.

References
1. Afessa B, Tefferi A, Hoagland HC,
Letendre L, Peters SG (1992) Outcome
of recipients of bone marrow transplants who require intensive-care unit
support. Mayo Clin Proc 67:117122

2. Apolone G, Bertolini G, DAmico R,

Iapichino G, Cattaneo A, De Salvo G
et al (1996) The performance of SAPS
II in a cohort of patients admitted to 99
Italian ICUs: results from GiViTI.
Gruppo Italiano per la Valutazione
degli interventi in Terapia Intensiva.
Intensive Care Med 22:13681378

3. Ashkenazi YJ, Kramer BS, Harman E

(1986) Short-term outcome among
patients with leukemia and lymphoma
admitted to a medical intensive care
unit. South Med J 79:10861088

694

4. Bach PB, Schrag D, Nierman DM,

Horak D, White P Jr, Young JW et al
(2001) Identification of poor prognostic features among patients requiring
mechanical ventilation after hematopoietic stem cell transplantation. Blood
98:32343240
5. Bastos PG, Sun X, Wagner DP, Knaus
WA, Zimmerman JE (1996) Application of the APACHE III prognostic
system in Brazilian intensive care
units: a prospective multicenter study.
Intensive Care Med 22:564570
6. Brivet FG, Kleinknecht DJ, Loirat P,
Landais PJ (1996) Acute renal failure
in intensive care units--causes, outcome, and prognostic factors of hospital mortality: a prospective, multicenter
study. French Study Group on Acute
Renal Failure. Crit Care Med
24:192198
7. Brunet F, Lanore JJ, Dhainaut JF,
Dreyfus F, Vaxelaire JF, Nouira S et al
(1990) Is intensive care justified for
patients with haematological malignancies? Intensive Care Med16:291297
8. Carson SS, Bach PB (2001) Predicting
mortality in patients suffering from
prolonged critical illness: an assessment of four severity-of-illness measures. Chest 120:928933
9. Chen CH, George SL (1985) The bootstrap and identification of prognostic
factors via Coxs proportional hazards
regression model. Stat Med 4:3946
10. Chertow GM, Christiansen CL, Cleary
PD, Munro C, Lazarus JM (1995)
Prognostic stratification in critically
ill patients with acute renal failure
requiring dialysis. Arch Intern Med
155:15051511
11. Consensus statement of the Society
of Critical Care Medicines Ethics
Committee regarding futile and other
possibly inadvisable treatments (1997)
Crit Care Med 25:887891
12. Cox SC, Norwood SH, Duncan CA
(1985) Acute respiratory failure:
mortality associated with underlying
disease. Crit Care Med 13:10051008
13. Crawford SW, Petersen FB (1992)
Long-term survival from respiratory
failure after marrow transplantation
for malignancy. Am Rev Respir Dis
145:510514
14. Crawford SW, Schwartz DA, Petersen
FB, Clark JG (1988) Mechanical
ventilation after marrow transplantation. Risk factors and clinical outcome.
Am Rev Respir Dis 137:682687
15. Denardo SJ, Oye RK, Bellamy PE
(1989) Efficacy of intensive care for
bone marrow transplant patients with
respiratory failure. Crit Care Med
17:46

16. Estopa R, Torres MA, Kastanos N,

Rives A, Agusti-Vidal A, Rozman C
(1984) Acute respiratory failure in
severe hematologic disorders. Crit Care
Med 12:2628
17. Ewer MS, Ali MK, Atta MS, Morice
RC, Balakrishnan PV (1986) Outcome
of lung cancer patients requiring mechanical ventilation for pulmonary
failure. JAMA 256:33643366
18. Groeger JS, Aurora RN (2001) Intensive care, mechanical ventilation,
dialysis, and cardiopulmonary resuscitation. Implications for the patient with
cancer. Crit Care Clin 17:791803
19. Groeger JS, Lemeshow S, Price K,
Nierman DM, White P Jr, Klar J et al
(1998) Multicenter outcome study of
cancer patients admitted to the intensive care unit: a probability of mortality model. J Clin Oncol 16:761770
20. Groeger JS, White P Jr, Nierman DM,
Glassman J, Shi W, Horak D et al
(1999) Outcome for cancer patients
requiring mechanical ventilation. J Clin
Oncol 17:991997
21. Guiguet M, Blot F, Escudier B, Antoun
S, Leclercq B, Nitenberg G (1998)
Severity-of-illness scores for neutropenic cancer patients in an intensive
care unit: Which is the best predictor?
Do multiple assessment times improve
the predictive value? Crit Care Med
26:488493
22. Hanley JA, McNeil BJ (1982) The
meaning and use of the area under
a receiver operating characteristic
(ROC) curve. Radiology 143:2936
23. Harrell FE Jr, Lee KL, Mark DB
(1996) Multivariable prognostic
models: issues in developing models,
evaluating assumptions and adequacy,
and measuring and reducing errors.
Stat Med15:361387
24. Hosmer DW, Lemeshow S (1989)
Applied logistic regression. Wiley:
New York
25. Jackson T, McKeag N, Stuart J (1997)
Care after bone marrow transplantation. Nurs Times 93:4849
26. Kapoor M, Chan GZ (2001) Malignancy and renal disease. Crit Care Clin
17:571598
27. Klastersky J, Paesmans M, Rubenstein
EB, Boyer M, Elting L, Feld R et al
(2000) The Multinational Association for
Supportive Care in Cancer risk index: A
multinational scoring system for identifying low-risk febrile neutropenic cancer
patients. J Clin Oncol18:30383051
28. Knaus WA, Draper EA, Wagner DP,
Zimmerman JE (1985) Apache II: a severity of disease classification system.
Crit Care Med 13:818829
29. Le Gall J-R, Lemeshow S, Sauinier F
(1993) A new Simplified Acute Physiology Score (SAPS II) based on a European/North American Multicenter
Study. JAMA 270:29572963

30. Lemeshow S, Teres D, Klar J, Avrunin

JS, Gehlbach SH, Rapoport J (1993)
Mortality Probability Models (MPM
II) based on an international cohort of
intensive care unit patients. JAMA
270:24782486
31. Lemeshow S, Klar J, Teres D,
Avrunin JS, Gehlbach SH, Rapoport J
et al (1994) Mortality probability
models for patients in the intensive
care unit for 48 or 72 hours: a prospective, multicenter study. Crit Care Med
22:13511358
32. Marcin JP, Pollack MM, Patel KM,
Ruttimann UE (2000) Combining
physicians subjective and physiologybased objective mortality risk predictions. Crit Care Med 28:29842990
33. Mazumdar M, Glassman JR (2000)
Categorizing a prognostic variable:
review of methods, code for easy implementation and applications to
decision making about cancer treatments. Stat Med 19:113132
34. Moreno R, Miranda DR, Fidler V,
Van Schilfgaarde R (1998) Evaluation
of two outcome prediction models
on an independent database. Crit Care
Med 26:5061
35. Motzer RJ, Mazumdar M, Bacik J,
Berg W, Amsterdam A, Ferrara J
(1999) Survival and prognostic
stratification of 670 patients with
advanced renal cell carcinoma. J Clin
Oncol 17:25302540
36. Murphy-Filkins R, Teres D, Lemeshow
S, Hosmer DW (1996) Effect of changing patient mix on the performance of
an intensive care unit severity-of-illness model: how to distinguish a general from a specialty intensive care
unit. Crit Care Med 24:19681973
37. Nouira S, Belghith M, Elatrous S,
Jaafoura M, Ellouzi M, Boujdaria R et
al (1998) Predictive value of severity
scoring systems: comparison of four
models in Tunisian adult intensive care
units. Crit Care Med 26:852859
38. Paesmans M, Sculier JP, Libert P,
Bureau G, Dabouis G, Thiriaux J et al
(1995) Prognostic factors for survival
in advanced non-small-cell lung cancer: univariate and multivariate analyses including recursive partitioning
and amalgamation algorithms in 1,052
patients. The European Lung Cancer
Working Party. J Clin Oncol
13:12211230
39. Paesmans M, Sculier JP, Lecomte J,
Thiriaux J, Libert P, Sergysels R et al
(2000) Prognostic factors for patients
with small cell lung carcinoma: analysis of a series of 763 patients included
in 4 consecutive prospective trials with
a minimum follow-up of 5 years.
Cancer 89:523533

695

40. Paz HL, Crilley P, Weinar M, Brodsky

I (1993) Outcome of patients requiring
medical ICU admission following bone
marrow transplantation. Chest
104:527531
41. Paz HL, Garland A, Weinar M, Crilley
P, Brodsky I (1998) Effect of clinical
outcomes data on intensive care unit
utilization by bone marrow transplant
patients. Crit Care Med 26:6670
42. Peters SG, Meadows III JA, Gracey
DR (1988) Outcome of respiratory
failure in hematologic malignancy.
Chest 94:9999
43. Price KJ, Thall PF, Kish SK, Shannon
VR, Andersson BS (1998) Prognostic
indicators for blood and marrow transplant patients admitted to an intensive
care unit. Am J Respir Crit Care Med
158:876884
44. Rubenfeld GD, Crawford SW (1996)
Withdrawing life support from mechanically ventilated recipients of bone
marrow transplants: a case for evidence-based guidelines. Ann Intern
Med 125:625633

45. Sauerbrei W, Schumacher M (1992) A

bootstrap resampling procedure for
model building: application to the Cox
regression model. Stat Med
11:20932109
46. Schapira DV, Studnicki J, Bradham
DD, Wolff P, Jarrett A (1993) Intensive
care, survival, and expense of treating
critically ill cancer patients. JAMA
269:783786
47. Schuster DP, Marion JM (1983)
Precedents for meaningful recovery
during treatment in a medical intensive
care unit. Outcome in patients with
hematologic malignancy. Am J Med
75:402408
48. Sculier JP, Markiewicz E (1993)
Cardiopulmonary resuscitation in medical cancer patients: the experience of a
medical intensive-care unit of a cancer
centre. Support Care Cancer 1:135138
49. Sculier JP, Paesmans M, Markiewicz
E, Berghmans T (2000) Scoring systems in cancer patients admitted for
an acute complication in a medical
intensive care unit. Crit Care Med
28:27862792
50. Shusterman N, Strom BL, Murray TG,
Morrison G, West SL, Maislin G
(1987) Risk factors and outcome of
hospital-acquired acute renal failure.
Clinical epidemiologic study. Am J
Med 83:6571
51. Sirio CA, Tajimi K, Tase C, Knaus
WA, Wagner DP, Hirasawa H et al
(1992) An initial comparison of intensive care in Japan and the United
States. Crit Care Med 20:12071215

52. Spiegel DM, Ullian ME, Zerbe GO,

Berl T (1991) Determinants of survival
and recovery in acute renal failure
patients dialyzed in intensive-care
units. Am J Nephrol 11:4447
53. Staudinger T, Stoiser B, Mullner M,
Locker GJ, Laczika K, Knapp S et al
(2000) Outcome and prognostic factors
in critically ill cancer patients admitted
to the intensive care unit. Crit Care
Med 28:13221328
54. Todd K, Wiley F, Landaw E, Gajewski
J, Bellamy PE, Harrison RE et al
(1994) Survival outcome among 54
intubated pediatric bone marrow
transplant patients. Crit Care Med
22:171176
55. Welch HG, Larson EB (1989) Cost
effectiveness of bone marrow transplantation in acute nonlymphocytic
leukemia. NEJM 321:807812
56. Zhu BP, Lemeshow S, Hosmer DW,
Klar J, Avrunin J, Teres D (1996)
Factors affecting the performance of
the models in the Mortality Probability
Model II system and strategies of customization: a simulation study. Crit
Care Med 24:5763