Allan Garland

Kendiss Olafson
Clare D. Ramsey
Marina Yogendran
Randall Fransoo
Distinct determinants of long-term
and short-term survival in critical illness
Received: 30 December 2013
Accepted: 19 May 2014
Published online: 11 July 2014
Springer-Verlag Berlin Heidelberg and
ESICM 2014
Take-home message: Short-term mortality
after onset of critical illness is determined
mainly by type and severity of the acute
illness. Age and comorbid conditions exert
small inuences on short-term mortality, but
are the main determinants of long-term
mortality among those who survive in the
short term.
Electronic supplementary material
The online version of this article
(doi:10.1007/s00134-014-3348-y) contains
supplementary material, which is available
to authorized users.
A. Garland (
)
) K. Olafson
C. D. Ramsey
Department of Internal Medicine,
University of Manitoba, 820 Sherbrook St.,
Room GF-222, Winnipeg, MB R3A1R9,
Canada
e-mail: agarland@hsc.mb.ca
A. Garland R. Fransoo
Department of Community Health Sciences,
University of Manitoba, Winnipeg, MB,
Canada
A. Garland M. Yogendran R. Fransoo
Department of Manitoba Center for Health
Policy, University of Manitoba,
408-727 McDermot Ave., Winnipeg,
MB R3E3P5, Canada
Abstract Purpose: To identify
the determinants of short-term and
long-term survival in adult patients
admitted to intensive care units
(ICUs). Methods: This population-
based, observational cohort study
included all eleven adult ICUs in the
Winnipeg Health Region of Mani-
toba, Canada, analyzing initial ICU
admissions during the period
19992010 of all Manitobans
C17 years old. Analysis included
KaplanMeier survival curves and
multivariable regression models of
30-day mortality and post-90-day
survival among those who survived to
day 90. We used likelihood ratios to
compare the predictive power of
clusters of variables in these models.
Results: After 33,324 initial ICU
admissions, mortality rates within 30
and 90 days were 15.9 and 19.5 %,
respectively. The survival curve
demonstrated an early phase with a
high rate of death, followed by a
markedly lower death rate that was
only clearly established after several
months. 30-day mortality was pre-
dominantly determined by
characteristics of the acute illness;
with its relative contribution set at
1.00, the next largest contributors
were age (0.19) and comorbidity
(0.16). In contrast, post-90-day mor-
tality was mainly determined by age
(relative contribution 1.00) and
comorbidity (0.95); the next largest
contributor was characteristics of
acute illness (0.28). Conclu-
sions: We observed two phases of
survival related to critical illness.
Short-term mortality was mainly
determined by the acute illness, but
its effect decayed relatively rapidly.
Mortality beyond 3 months, among
those who survived to that point, was
mainly determined by age and
comorbidity. Recognition of these
ndings is relevant to discussions
with patients and surrogates about
achievable goals of care.
Keywords Prognosis Critical care
Outcome assessment Mortality
End of life care
Introduction
Among the most important things done by physicians car-
ing for critically ill patients in intensive care units (ICUs) is
discussing treatment plans with patients and their
surrogates. A necessary part of devising care plans with
achievable goals is accurately prognosticating the chance
of survival. However, despite decades of development of
ICU risk stratication systems such as SAPS, APACHE,
and MPM [1], their usefulness for this purpose is limited.
Intensive Care Med (2014) 40:10971105
DOI 10.1007/s00134-014-3348-y
ORIGINAL
Such systems require complex calculations, and their
ability to predict outcomes for individual patients is poor
[2]. Also, they were only designed to address short-term
outcomes, such as in-hospital mortality, which can be
misleading because it is heavily inuenced by inter-hos-
pital and post-hospital transfer patterns [3]. Most
importantly, short-term outcomes are not patient cen-
tered because most individuals place more value on
long-term than short-term survival [4].
Also, even for the short-term outcomes for which
prediction schemas exist, physicians feel they receive
inadequate training in prognostication [5, 6]. Perhaps
reecting this, White et al. [7] found that in more than
one-third of discussions, ICU physicians did not discuss
the survival prognosis. In the absence of approaches to
short-term prognostication that are easy to understand and
use at the bedside, ICU practitioners commonly rely on
gestalt based upon their personal experiences with short-
term outcomes [8].
For all of these reasons, we sought to separately
identify the factors that determine long-term and short-
term survival after initial episodes of critical illness.
A subtle issue in identifying the basic phenomena at
work is to distinguish between long-term survival starting
at the onset of critical illness and long-term survival
among those who survive in the short term. Prior studies
have failed to clearly separate these phenomena [912].
Most studies of long-term survival from critical illness
have been limited as they evaluate specic patient subsets
and adjust for few confounding variables; none have been
population-based. Furthermore, we are unaware of studies
that assessed short-term and long-term mortality in the
same cohort.
In this hypothesis-driven study, we sought to over-
come all the limitations of prior investigations by
evaluating a large, population-based ICU cohort. We
hypothesized that there are two elemental phenomena at
work, with the primary determinants of short-term sur-
vival from the onset of critical illness being distinct from
the main determinants of long-term survival among those
who survive in the short term. Specically, we hypothe-
sized that the former are the type and severity of acute
illness, while the latter are age and chronic medical
conditions.
This work was the subject of a published conference
abstract [13].
Methods
We performed a population-based analysis of adults
admitted to 11 adult ICUs in the Winnipeg Regional Health
Authority of the Canadian province of Manitoba over
130 months during 19992010. Four of the ICUs are within
community hospitals; the other seven are in tertiary
teaching hospitals. ICU types are: ve medical-surgical,
two coronary care, one medical, one surgical/trauma/neu-
rosurgical, one cardiac surgical, and one respiratory. All are
closed-model ICUs staffed by intensivists or cardiologists.
These represent all level 1 and level 2 adult ICUs [14] in
Manitoba, except for a nine-bed medical-surgical ICU in
the city of Brandon, 124 miles away. Manitobas popula-
tion was 1.23 million in 2010, and 697,000 (57 %) of those
resided in the Winnipeg region [15].
Data for this analysis came from the Manitoba Inte-
grated Critical Care Database (MICCDB), which has been
described previously [16, 17], and links a clinical ICU
database to the provinces administrative health data. The
ICU database contains detailed information about all
admissions to the 11 included ICUs. Manitoba has uni-
versal, single-payer health insurance, and the
administrative data include hospital discharge abstracts
and vital status data for all residents. We identied sur-
vival status up to a censoring date of November 30, 2011.
Creation of the MICCDB included identifying inter-ICU
and inter-hospital transfers and merging records that were
part of a single episode of ICU care [17]; these episodes
are the units of analysis in this study.
We assessed survival after ICU admission for Mani-
toba residents C17 years of age admitted to any of the 11
Winnipeg ICUs. In order to study a more homogeneous
phenomenon by evaluating outcomes after initial ICU
episodes, and to avoid biased mortality rates due to
individuals being represented multiple times, we only
considered the rst ICU episode of each person during the
study period, and excluded those who had been admitted
to any level 1 or level 2 ICU during the prior 10 years.
To assess mortality over time, we rst created a
KaplanMeier survival curve. For comparison, we
applied standard methods [18] to Manitoba life tables [19]
to create the survival curve of an age- and sex-matched
general population cohort. To assess whether the ICU
patients mortality rate eventually became similar to that
of the general population, we used Fishers Z-test to
compare the slopes from years 6 to 12.
We then created three multivariable regression mod-
els. For short-term mortality, we used logistic regression
of death within 30 days after ICU admission. For long-
term mortality, we used Cox proportional hazards
regression of time to death[90 days after ICU admission
among those who survived to day 90 and remained in the
province. Our choice to begin the long-term phase at
90 days after ICU admission was informed by the Kap-
lanMeier curve (Fig. 1). Last, to test the expectation that
a mixture of short-term and long-term determinants would
be seen in analysis spanning the early and late phases, we
performed Cox regression of time to death starting from
ICU admission. The three regression models included
identical explanatory variables, including demographics,
comorbid conditions, ICU admission details, and type and
severity of acute illness.
1098
Demographics were age, sex, and socioeconomic sta-
tus (SES). SES was divided into categories based on
average household income in the postal code of residence
[20]. Ten of the categories were quintiles, with rural and
urban residents categorized separately. The 11th SES
category, referred to as Not Calculated, included peo-
ple living in areas where the census does not calculate
average household income (mainly nursing homes and
other chronic care facilities, with a smaller portion in
prisons).
For comorbidities, the 31 conditions described by
Elixhauser et al. [21] were identied fromhospital abstracts
using ICD-9-CM and ICD-10 coding [22]. These were
derived using all diagnoses listed in the index hospitaliza-
tion and all hospitalizations up to 1 year prior to it [23]. We
collapsed the categories of diabetes with and without
chronic complications because this distinction was awed
in Manitoba before 2006. Also, the HIV/AIDS category
was not included due to small numbers (n = 61, 0.2 %).
ICU admission details were year (categorized into
approximately equal time intervals as 19992002,
200306, or 200709), timing (weekday vs. weekend;
days vs. nights, dened as 6pm8am), hospital type
(community vs. tertiary/teaching), and pre-ICU location
(emergency department, ward, or operating room/post-
anesthesia unit).
Primary ICU admission diagnosis was derived from
that eld in the ICU database. This custom system of
2,065 diagnoses was categorized by mapping to the cor-
responding ICD-9-CM chapters. The eight chapters with
\100 observations each were collapsed into a single
miscellaneous category containing 187 observations.
We used ve measures of severity and type of acute ill-
ness: (a) Glasgow Coma Scale score (GCS), categorized
as 35, 68, 912, 1314, and 15; (b) APACHE II Acute
Physiology Score [1], excluding its neurologic subscore
(APSminusNeuro), since GCS was included separately;
and use during the initial two calendar days in ICU of
(c) invasive mechanical ventilation; (d) any vasoactive
agents; and (e) any form of renal dialysis.
To allow for nonlinear relationships between mortality
and the continuous covariates of age and APSminus-
Neuro, these were included in the models as 4-knot
restricted cubic splines; the resulting relationships were
evaluated statistically and graphically [24]. We ensured
the absence of multicollinearity among the independent
variables using variance ination factors [25]. Propor-
tional hazards for Cox regression was veried by log-
minus-log plots and time-dependent covariates [26];
variables in violation were included as time-dependent
covariates.
The central analysis for this study quantied the incre-
mental predictive power of ve clusters of variables: (1) age,
(2) sex and SES, (3) comorbid conditions, (4) acute illness
characteristics (diagnosis, GCS, APSminusNeuro, pre-ICU
location, need for mechanical ventilation, vasoactive drugs,
and dialysis), and (5) the remaining model variables
(admission timing, admission year, hospital type). Because
of variable scaling and other issues, regression coefcient
magnitudes and p-values do not provide this information.
Instead, we adapted the approach espoused by Harrell [27]
and used likelihood ratios of nested models to calculate a
unitless index of the incremental predictive value of model
variables [27]. For eachof the three models, for each variable
cluster, we calculated the likelihood ratio of the model
including all explanatory variables and the likelihood ratio
for the reduced nested model excluding only the variable
cluster of interest. The quotient of the likelihood ratio (QLR)
of the reduced model to the full model is a measure of the
incremental explanatory power of the variable cluster of
interest added to all the other variables in the model [27]. To
facilitate comparison, we scaled these difference ratios so
that, within each model, the scaled value for the most
inuential variable cluster equaled 1.000.
For analysis, we used SAS 9.2 (SAS Institute, Cary,
NC, USA). Unless indicated otherwise, the values pre-
sented are the mean standard deviation (SD). p-
values \0.05 were considered statistically signicant.
This study was approved by the Health Research Ethics
Board of the University of Manitoba and the Manitoba
Health Information Privacy Committee.
Results
During the study period, 34,306 individuals had initial
ICU episodes. Of these, 982 (2.9 %) had missing pre-ICU
locations and were excluded, leaving 33,324 episodes for
analysis (see Table 1 and eTable 1 in the Electronic
supplementary material, ESM). Sixty percent were male,
average age was 63.5 16.2 years, urban residents and
0 2 4 6 8 10 12
0.00
0.25
0.50
0.75
1.00
Observed ICU cohort
Age- & sex-matched
general population
30 days
90 days
S
u
r
v
i
v
a
l

F
r
a
c
t
i
o
n
Years after ICU Admission
Fig. 1 KaplanMeier curve of survival after rst admission to
intensive care units, including 95 % condence intervals, with a
comparison of survival for an age- and sex-matched general
population cohort
1099
lower SES categories were more common, and most
patients were admitted from emergency departments into
tertiary hospital ICUs. The most common admission
diagnosis category was cardiovascular disorders, followed
by the category including sepsis and its complications,
and respiratory disorders. Average APACHE II score was
15.3 7.9 and, during the initial two ICU days, 57 %
received invasive mechanical ventilation, 46 % needed
vasoactive drugs, and 5 % received renal dialysis. Median
(interquartile range) ICU and hospital lengths of stay
were, respectively, 2.4 days (1.14.6) and 11 days (624).
In this cohort, 5,285 (15.9 %) died within 30 days and
6,484 (19.5 %) died within 90 days. Excluding 283
patients who survived to 90 days but had left the prov-
ince, 26,557 patients were included in the long-term
analysis (Table 1, eTable 1 of the ESM); of these, 1,432
(5.05 %) were still in the hospital and 42 (0.16 %) were
still in ICU at day 90.
The survival curve (Fig. 1) demonstrates two distinct
phases of mortality associated with critical illness. The
early phase has a high rate of death, which is followed by
a markedly lower death rate that is only clearly estab-
lished several months after ICU admission. Median
survival was 9.2 years (95 % CI, 8.99.5). For years 6 to
Table 1 Patient characteristics
for selected variables
Variable All patients
(n = 33,324)
Survivors to 90 days
after ICU admission
(n = 26,557)
Female sex 13,506 (40.5) 10,448 (39.3)
Age (years), mean SD 63.5 16.2 62.4 16.0
Comorbid conditions
Hypertension, uncomplicated 12,498 (37.5) 10,118 (38.1)
Arrhythmia 9,338 (28.0) 7,228 (27.2)
Diabetes 8,447 (25.3) 6,699 (25.2)
Congestive heart failure 7,355 (22.1) 5,332 (20.1)
Chronic pulmonary disease 4,159 (12.5) 3,072 (11.6)
Fluid and electrolyte disorders 3,033 (9.1) 2,075 (7.8)
Valvular heart disease 2,762 (8.3) 2,384 (9.0)
Peripheral vascular disease 2,658 (8.0) 2,011 (7.6)
Renal failure 2,564 (7.7) 1,691 (6.4)
Other neurologic disorders 2,424 (7.3) 1,385 (5.2)
Solid tumor without metastasis 2,241 (6.7) 964 (3.6)
Hypertension, complicated 1,927 (5.8) 1,386 (5.2)
Alcohol abuse 1,793 (5.4) 1,267 (4.8)
Pre-ICU location
Emergency department 17,532 (52.6) 14,163 (53.3)
Operating room or post-anesthesia unit 9,860 (29.6) 8,670 (32.6)
Ward 5,932 (17.8) 3,724 (14.0)
Hospital type
Tertiary/teaching hospital 21,197 (63.6) 17,105 (64.4)
Community hospital 12,127 (36.4) 9,452 (35.6)
Admission diagnosis category
Circulatory system 20,140 (60.4) 17,256 (65.0)
Symptoms, signs, and ill-dened conditions
a
3,925 (11.8) 2,281 (8.6)
Respiratory system 3,557 (10.7) 2,588 (9.7)
Injury and poisoning 2,542 (7.6) 2,084 (7.8)
Digestive system 1,143 (3.4) 757 (2.9)
Nervous system and sense organs 634 (1.9) 478 (1.8)
Endocrine, nutritional, metabolic, immunity 525 (1.6) 433 (1.6)
Genitourinary system 361 (1.1) 279 (1.1)
Others (see eTable 1 of the ESM) 497 (1.5) 401 (1.5)
APACHE II score, mean SD 15.3 7.9 13.3 6.2
Glasgow Coma Scale score
35 1,910 (5.7) 461 (1.7)
68 1,456 (4.4) 804 (3.0)
912 2,303 (6.9) 1,507 (5.7)
1314 2,965 (8.9) 2,096 (7.9)
15 24,690 (74.1) 21,689 (81.7)
Mechanical ventilation ICU day 1 and/or 2 19,091 (57.3) 13,897 (52.3)
Vasoactive drugs ICU day 1 and/or 2 15,387 (46.2) 11,204 (42.2)
Renal dialysis ICU day 1 and/or 2 1,789 (5.4) 1,213 (4.6)
See eTable 1 of the ESM for full results
Values are given as number (%) unless otherwise indicated
a
Includes sepsis-related diagnoses
1100
12, the survival curve slope for ICU patients was not
different from that of the age- and sex-matched general
population cohort (p = 0.99).
Numerous variables were associated with 30-day and
post 90-day mortality (Table 2, eTable 2 of the ESM;
Fig. 2, eFigure 1 of the ESM). The effects of three
comorbid conditions violated the proportional hazards
assumption and were included in the post-90-day model
as time-dependent covariates (eTable 2 of the ESM). The
relationship of age to mortality was signicant and non-
linear in both models, with a rapidly rising risk of death
above 70 years of age (Fig. 2). While APSminusNeuro
was signicant in both models, it was much more strongly
related to 30-day than to post-90-day mortality (eFigure 1
of the ESM). A number of comorbid conditions were
associated with higher mortality in both models. The
association of a few chronic conditions with lower mor-
tality likely represents the effect of coding bias, wherein
milder chronic conditions are less likely to be coded for
more severely ill patients [21].
There were notable differences between determinants
in the 30-day and post-90-day mortality models. Sex was
only signicant in the post-90-day model, with a higher
hazard for death among men, as in the general population.
After adjusting for other variables, mortality to 30 days
was associated with the use while in ICU of mechanical
ventilation and vasoactive agents, but not dialysis, while
the reverse was true for post-90-day mortality. GCS only
had a consistent doseresponse relationship with 30-day
mortality. While many acute diagnostic categories were
signicant in both models, there were some notable dif-
ferential effects, e.g., compared to the reference group of
cardiovascular diagnoses, patients admitted for genito-
urinary disorders had signicantly lower 30-day mortality
but signicantly higher post-90 day mortality.
However, the magnitude and p-values of coefcients in
regression equations do not directly indicate their predic-
tive power. Instead, Table 3 shows that death at 30 days
after ICU admission was predominantly determined by the
characteristics of the acute illness; the other four variable
clusters had much lower predictive power. In contrast,
mortality beyond 90 days after ICU admission, among
patients who survived to that time point, was mainly
determined by age and comorbidity. As expected, factors
that were predictive of death in a model that included both
the short-term and long-term phases were a mixture of
those from the two purer models: acute illness, age, and
comorbidity were the major determinants.
Discussion
Since people generally place more value on long-term than
short-term survival [4], conversations with ICU patients or
their surrogates about achievable goals and care plans are
inadequate if they focus only on the acute illness and its
short-term consequences without explicitly addressing
subsequent survival. By better delineating, in a simple
fashion, the factors relevant to short-term vs. long-term
survival, our ndings can help inform such discussions.
Though several prior studies attempted to evaluate
determinants of long-term outcomes, we are not aware of
any that avoided mixing short-term with long-term
determinants. For some, this occurred because they
evaluated long-term survival starting from the time of
ICU admission [9, 10]. Even those that attempted to
exclude the short-term period used starting points that
were still within the envelope of the short-term determi-
nants of mortality. Specically, Fig. 1 indicates that it
takes approximately 3 months after ICU admission for
most of the inuence exerted by the acute illness to dis-
appear, while those earlier studies began their long-term
analyses at the time of ICU or hospital discharge [11, 12,
28].
For these reasons, only approximate comparisons can
be made between our study and two previous studies of
long-term mortality that also assessed the predictive
power of different variables. Those studies, which used
starting points either 5 days after ICU admission [12] or
at ICU discharge [11], both reported the following trend
for strength of predictive power: age [comorbidities [
severity of acute illness. In contrast, our model of mor-
tality starting from ICU admission showed the following
trend: characteristics of acute illness C age C comor-
bidity. The reduced contribution of acute illness in the
two prior studies likely reects the beginnings of its
decaying importance, even when starting from a time
point less than 1 week after onset of critical illness.
Among other comparisons with existing literature,
Williams et al. [28] reported determinants of post-hospital
mortality similar to those seen in our post-90-day mor-
tality model (eTable 2 of the ESM), though further
comparisons are difcult because those investigators did
not perform analysis assessing the predictive power of the
variables. In contrast to our ndings for unselected ICU
patients, traumatic brain injury patients were found to
have higher long-term mortality than that seen for a
matched general population cohort [29]. And, as in Fig. 2,
prior studies have also found that the inuence of age on
long-term mortality becomes extreme among the oldest
patients [3032].
We can more readily compare our model of 30-day
mortality to existing studies of short-term mortality [33
36]. Like our study, those that assessed the predictive
power of different variables all found that the strongest
predictors were the characteristics of the acute illness,
with lesser contributions from age and comorbidity [33
35]. In a multicenter study of ICU survivors admitted for
infections, Azoulay et al. [36] found determinants of in-
hospital mortality that were similar to our 30-day mor-
tality results.
1101
Our study has strengths and limitations. It is a large,
population-based study assessing consecutive, unselected
patients admitted over a substantial time span to all types
of ICUs. Although we did not have detailed clinical data
from one of the 12 ICUs in the province, this would not be
expected to substantially change the overall results, since
that ICU accounts for only 6.5 % of all ICU bed-days in
Manitoba [37]. We evaluated a wide and robust range of
potential determinants of mortality. This is the rst study
with a methodology that permitted a clear delineation of
short-term from long-term inuences on outcome, and it
did so in the same patient cohort. Two major limitations
are that our dataset did not allow us to address the
prognosis for functional recovery, or the fact that some
acute illnesses result in new chronic illnesses (e.g., major
stroke). The former is relevant because functional and
quality of life outcomes are important inuences on
patients treatment decisions [38], and are known to be
degraded among ICU survivors [39]. The latter is relevant
because such new chronic illnesses may substantially
Table 2 Regression models of mortality after initial ICU admission for selected variables
Independent variable Death within 30 days after ICU
admission
Death [90 days after ICU admission for
90-day survivors
OR 95 % CI HR 95 % CI
Age * (See Fig. 2, text) * (See Fig. 2, text)
APSminusNeuro
a
* (See eFigure 1 of the ESM,
text)
* (See eFigure 1 of the ESM, text)
Female sex 0.956 0.887, 1.031 0.854* 0.815, 0.894
Socioeconomic status * *
Urban 1 (lowest urban) (reference) 1.000 1.000
Urban 5 (highest urban) 0.877 0.758, 1.016 0.719* 0.658, 0.787
Rural 1 (lowest rural) 0.870 0.727, 1.042 0.974 0.875, 1.085
Rural 5 (highest rural) 0.930 0.777, 1.113 0.785* 0.701, 0.879
Not calculated 0.647* 0.523, 0.801 1.405* 1.255, 1.574
Comorbid conditions
Metastatic cancer 2.880* 2.453, 3.381 4.679* 4.205, 5.207
Liver disease 2.475* 2.124, 2.884 1.715* 1.509, 1.950
Other neurologic disorders 1.943* 1.728, 2.185 1.234* 1.120, 1.360
Pulmonary circulation disorders 1.480* 1.240, 1.766 1.142* 1.007, 1.295
Congestive heart failure 1.247* 1.140, 1.363 1.611* 1.532, 1.695
Chronic pulmonary disease 1.132* 1.018, 1.257 1.510* 1.424, 1.602
Renal failure 1.079 0.934, 1.248 1.659* 1.516, 1.817
Diabetes 0.986 0.901, 1.078 1.386* 1.317, 1.459
Alcohol abuse 0.838* 0.714, 0.983 1.264* 1.135, 1.408
Hypertension, uncomplicated 0.799* 0.734, 0.871 0.733* 0.680, 0.790
Year of ICU admission (reference: 19992002) * *
20072009 0.878* 0.798, 0.966 0.705* 0.655, 0.759
Timing of ICU admission (reference: weekday daytime) * *
Weekday evening 1.143* 1.049, 1.247 1.077* 1.022, 1.134
Weekend daytime 1.154* 1.015, 1.313 1.086* 1.004, 1.175
Pre-ICU location (reference: emergency department) * *
Operating room or post-anesthesia unit 0.474* 0.424, 0.530 0.693* 0.648, 0.740
Ward 1.189 1.085, 1.301 1.043 0.980, 1.110
Admission diagnosis (reference: circulatory system) * *
Digestive system 1.629* 1.361, 1.949 1.491* 1.330, 1.672
Respiratory system 1.400* 1.237, 1.584 1.547* 1.432, 1.671
Symptoms, signs, and ill-dened conditions
b
1.318* 1.180, 1.472 1.392* 1.287, 1.506
Endocrine, nutritional, metabolic, immunity 0.818 0.605, 1.105 1.227* 1.042, 1.445
Genitourinary system 0.628* 0.444, 0.887 1.375* 1.145, 1.651
Glasgow Coma Scale score (reference: 15) * *
912 1.983* 1.751, 2.246 1.183* 1.076, 1.301
35 5.271* 4.619, 6.016 0.994 0.829, 1.192
Mechanical ventilation ICU day 1 and/or 2 1.415* 1.284, 1.560 0.991 0.938, 1.047
Vasoactive drugs ICU day 1 and/or 2 1.201* 1.103, 1.308 0.952 0.904, 1.002
Renal dialysis ICU day 1 and/or 2 1.157 0.994, 1.347 1.356* 1.221, 1.504
See eTable 2 of the ESM for full results
OR odds ratio, HR hazard ratio
* p \0.05
a
APACHE II Acute Physiology Score excluding its neurologic
subscore
b
Includes sepsis-related diagnoses
1102
inuence post-90-day survival. Thus, by limiting consid-
eration to mortality outcomes, our analysis does not
include the full range of personal values that patients use
in making treatment decisions. Also, we only considered
each patients initial ICU admission. Finally, some known
determinants of mortality were not present in our data,
such as prehospital living site, prehospital functional
status, post-hospital discharge location [40], and bio-
chemical markers [41].
Conclusions
We demonstrated that there are two distinct phases of
survival related to critical illness. Death in the month
after admission to ICU was mainly determined by the
type and severity of acute illness, but this factor was less
important when examining mortality after 90 days of
survival. Age and comorbidity were only minor con-
tributors to short-term survival. However, among those
who survived to 90 days after illness onset, subsequent
survival was primarily determined by comorbid condi-
tions and age, as in the general population. While
survival to 3090 days is not unimportant, existing at the
transition between the two clearcut phases, its determi-
nants will necessarily be a combination of those of the
short-term and long-term phasesit does not have its
own distinct determinants.
Acknowledgments This work was funded by a grant from the
University of Manitoba Research Grants Program.
Conicts of interest On behalf of all authors, the corresponding
author states that there is no conict of interest.
20 40 60 80 100
0
1
2
3
4
5
6
7
8
p-values
overall <0.0001
nonlinearity <0.0001
p-values
overall <0.0001
nonlinearity <0.0001
Hazard ratio for death >90 days
after ICU admission,
among those who survived to day 90
Odds ratio for death within 30 days
after ICU admission
O
d
d
s

R
a
t
i
o

o
r

H
a
z
a
r
d

R
a
t
i
o
(
r
e
l
a
t
i
v
e

t
o

v
a
l
u
e

a
t

m
e
d
i
a
n

a
g
e
)
Age (years)
20 40 60 80 100
0
1
2
3
4
5
6
7
8
Fig. 2 Relationship between
age and mortality from
regression models. Odds ratio
and hazard ratio are relative to
those at the median value of age
Table 3 Predictive power of variables or clusters of variables, as indicated by the quotient of likelihood ratios (QLR)
Variable or variable cluster Regression model
Death within 30 days after
ICU admission
Post-90-day time to death, for
90-day survivors
Time to death, from
ICU admission
QLR Scaled QLR* QLR Scaled QLR* QLR Scaled QLR*
Acute illness characteristics 0.548 1.000 0.082 0.280 0.313 1.000
Age 0.101 0.185 0.293 1.000 0.214 0.684
Comorbid conditions 0.090 0.164 0.279 0.952 0.174 0.556
Sex, socioeconomic status 0.004 0.008 0.024 0.080 0.012 0.038
Admission year, admission timing, hospital type 0.004 0.006 0.014 0.048 0.015 0.048
* Scaled relative to the largest QLR value, separately within each regression model
1103
References
1. Rosenberg AL (2002) Recent
innovations in intensive care unit risk-
prediction models. Curr Opin Crit Care
8:321330
2. Rogers J, Fuller HD (1994) Use of daily
acute physiology and chronic health
evaluation (APACHE) II scores to
predict individual patient survival rate.
Crit Care Med 22:14021445
3. Vasilevskis EE, Kuzniewicz MW, Dean
ML, Clay T, Vittinghoff E, Rennie DJ,
Dudley RA (2009) Relationship
between discharge practices and
Intensive Care Unit in-hospital
mortality performance evidence of a
discharge bias. Med Care 47:803812
4. Mazur D, Hickam D (1990)
Interpretation of graphic data by
patients in a general medicine clinic.
J Gen Intern Med 5:402405
5. Stevens L, Cook D, Guyatt G, Grifth
L, Walter S, McMullin J (2002)
Education, ethics, and end-of-life
decisions in the intensive care unit. Crit
Care Med 30:290296
6. Christakis NA, Iwashyna TJ (1998)
Attitude and self-reported practice
regarding prognostication in a national
sample of internists. Arch Intern Med
158:23892395
7. White D, Engelberg R, Wenrich M, Lo
B, Curtis J (2007) Prognostication
during physicianfamily discussions
about limiting life support in intensive
care units. Crit Care Med 35:442448
8. Sinuff T, Adhikari N, Cook D,
Schunemann H, Grifth L, Rocker G,
Walter S (2006) Mortality predictions
in the intensive care unit: comparing
physicians with scoring systems. Crit
Care Med 34:878885
9. Timmers T, Verhofstad M, Moons K,
Leenen L (2011) Long-term survival
after surgical intensive care unit
admission: fty percent die within
10 years. Ann Surg 253:151157
10. Ulvik A, Kvale R, Wentzel-Larsen T,
Flaatten H (2007) Multiple organ
failure after trauma affects even long-
term survival and functional status. Crit
Care 11:R95
11. Baldwin MR, Narain W, Wunsch H,
Lederer DJ, Bach PB (2012) Predictors
of 6 month mortality for elderly
survivors of critical illness. Am J Respir
Crit Care Med 185:A6787
12. Ho KM, Knuiman M, Finn J, Webb SA
(2008) Estimating long-term survival of
critically ill patients: the PREDICT
model. PLoS One 3:e3226
13. Garland A, Olafson K, Ramsey C,
Yogendren M, Fransoo R (2013)
Distinct determinants of long-term and
short-term survival in critical illness.
Am J Respir Crit Care Med 187:A5075
14. Haupt MT, Bekes CE, Brilli RJ, Carl
LC, Gray AW, Jastremski MS, Naylor
DF, Rudis M, Spevetz A, Wedel SK,
Horst M, Task Force of the American
College of Critical Care Medicine and
the Society of Critical Care Medicine
(2003) Guidelines on critical care
services and personnel:
recommendations based on a system of
categorization of three levels of care.
Crit Care Med 31:26772683
15. Manitoba Health (2011) Manitoba
health population report: June 1, 2011.
https://www.gov.mb.ca/health/
population/pr2011.pdf. Accessed 16 Jul
2012
16. Garland A, Yogendran M, Olafson K,
Scales DC, McGowan K-L, Fransoo R
(2012) The accuracy of administrative
data for identifying the presence and
timing of admission to Intensive Care
Units in a Canadian province. Med Care
50:e1e6
17. Fransoo R, Yogendran M, Olafson K,
Ramsey C, McGowan K-L, Garland A
(2012) Constructing episodes of
inpatient care: data infrastructure for
population-based research. BMC Med
Res Methodol 12:133
18. Finkelstein DM, Muzikansky A,
Schoenfeld DA (2003) Comparing
survival of a sample to that of a
standard population. J Natl Cancer Inst
95:14341439
19. Complete life table, Manitoba (2000 to
2002) Statistics Canada.
http://www.statcan.gc.ca/pub/84-537-x/
4064441-eng.htm. Accessed 25 Feb
2014
20. Statistics Canada (2007) Postal Code
Conversion File (PCCF), reference
guide. Statistics Canada, Ottawa
21. Elixhauser A, Steiner C, Harris DR,
Coffey RM (1998) Comorbidity
measures for use with administrative
data. Med Care 36:827
22. Quan H, Sundararajan V, Halfon P,
Fong A, Burnand B, Luthi J, Saunders
LD, Beck CA, Feasby TE, Ghali WA
(2005) Coding algorithms for dening
comorbidities in ICD-9-CM and ICD-
10 administrative data. Med Care
43:11301139
23. Stukenborg G, Wagner D, Connors A Jr
(2001) Comparison of the performance
of two comorbidity measures, with and
without information from prior
hospitalizations. Med Care 39:727739
24. Marrie RA, Dawson NV, Garland A
(2009) Quantile regression and
restricted cubic splines are useful for
exploring relationships between
continuous variables. J Clin Epidemiol
62:510516
25. Fox J (1991) Regression diagnostics.
Sage, London
26. Kleinbaum DG (1996) Survival
analysis: a self-learning text. Springer,
New York
27. Harrell Jr FE (2001) Regression
modeling strategies. Springer, New
York, pp 202206
28. Williams TA, Dobb GJ, Finn JC,
Knuiman MW, Geelhoed E, Lee KY,
Webb SAR (2008) Determinants of
long-term survival after intensive care.
Crit Care Med 36:15231530
29. Shavelle RM, Strauss D, Whyte J, Day
SM, Yu YL (2001) Long-term causes of
death after traumatic brain injury. Am J
Phys Med Rehabil 80:510516
30. Fuchs L, Chronaki CE, Park S, Novack
V, Baumfeld Y, Scott D, McLennan S,
Talmor D, Celi L (2012) ICU admission
characteristics and mortality rates
among elderly and very elderly patients.
Intensive Care Med 38:16541661
31. Biston P, Aldecoa C, Devriendt J, Madl
C, Chochrad D, Vincent J-L, De Backer
D (2014) Outcome of elderly patients
with circulatory failure. Intensive Care
Med 40:5056
32. deRooij SE, Govers A, Korevaar JC,
Abu-Hanna A, Levi M, deJonge E
(2006) Short-term and long-term
mortality in very elderly patients
admitted to an intensive care unit.
Intensive Care Med 32:10391044
33. Render ML, Kim HM, Welsh DE,
TImmons S, Johnston J, Hui S, Connors
AF, Wagner D, Daley J, Hofer TP, VA
ICU Project (VIP) Investigators (2003)
Automated intensive care unit risk
adjustment: results from a National
Veterans Affairs study. Crit Care Med
31:16381646
34. Johnston JA, Wagner DP, Timmons S,
Welsh D, Tsevat J, Render ML (2002)
Impact of different measures of
comorbid disease on predicted mortality
of Intensive Care Unit patients. Med
Care 40:929940
35. Knaus WA, Wagner DP, Zimmerman
JE, Draper EA (1993) Variations in
mortality and length of stay in intensive
care units. Ann Intern Med
118:753761
36. Azoulay E, Alberti C, Legendre I,
Buisson CB, Gall JRL, European Sepsis
Group (2005) Post-ICU mortality in
critically ill infected patients: an
international study. Intensive Care Med
31:5663
1104
37. Garland A, Fransoo R, Olafson K,
Ramsey C, Yogendran M, Chateu D,
McGowan K (2011) The epidemiology
and outcomes of critical illness in
Manitoba. Manitoba Centre for Health
Policy, Winnipeg
38. Fried TR, Bradley EH, Towle VR,
Allore H (2002) Understanding the
treatment preferences of seriously ill
patients. N Engl J Med 346:10611068
39. Cuthbertson BH, Roughton S,
Jenkinson D, MacLennan G, Vale L
(2010) Quality of life in the ve years
after intensive care: a cohort study. Crit
Care 14:R6
40. Wunsch H, Guerra C, Barnato AE,
Angus DC, Li G, Linde-Zwirble WT
(2010) Three-year outcomes for
medicare beneciaries who survive
intensive care. JAMA 303:849856
41. Yende S, DAngelo G, Mayr F, Kellum
J, Weissfeld L, Kaynar A, Young T,
Irani K, Angus D, GenIMS
Investigators (2011) Elevated
hemostasis markers after pneumonia
increases one-year risk of all-cause and
cardiovascular deaths. PLoS One
6:e22847
1105