RESEARCH ARTICLE

Prognostication of Long-Term Outcomes

after Subarachnoid Hemorrhage:
The FRESH Score
Jens Witsch, MD,1 Hans-Peter Frey, PhD,1 Sweta Patel, PA-C, MHS,1
Soojin Park, MD,1 Shouri Lahiri, MD,1 J. Michael Schmidt, PhD,1
Sachin Agarwal, MD, MPH,1 Maria Cristina Falo, PhD,1 Angela Velazquez, MD,1
Blessing Jaja, MD, PhD,2 R. Loch Macdonald, MD, PhD,2
E. Sander Connolly, MD,3 and Jan Claassen, MD, PhD1

Objective: To create a multidimensional tool to prognosticate long-term functional, cognitive, and quality of life out-
comes after spontaneous subarachnoid hemorrhage (SAH) using data up to 48 hours after admission.
Methods: Data were prospectively collected for 1,619 consecutive patients enrolled in the SAH outcome project July
1996 to March 2014. Linear models (LMs) were applied to identify factors associated with outcome in 1,526 patients
with complete data. Twelve-month functional, cognitive, and quality of life outcomes were measured using the modi-
fied Rankin scale (mRS), Telephone Interview for Cognitive Status, and Sickness Impact Profile. Based on the LM
residuals, we constructed the FRESH score (Functional Recovery Expected after Subarachnoid Hemorrhage). Score
performance, discrimination, and internal validity were tested using the area under the receiver operating characteris-
tic curve (AUC), Nagelkerke and Cox/Snell R2, and bootstrapping. For external validation, we used a control popula-
tion of SAH patients from the CONSCIOUS-1 study (n 5 413).
Results: The FRESH score was composed of Hunt & Hess and APACHE-II physiologic scores on admission, age, and
aneurysmal rebleed within 48 hours. Separate scores to prognosticate 1-year cognition (FRESH-cog) and quality of
life (FRESH-quol) were developed controlling for education and premorbid disability. Poor functional outcome
(mRS 5 4–6) for score levels 1 through 9 respectively was present in 3, 6, 12, 38, 61, 83, 92, 98, and 100% at 1-year
follow-up. Performance of FRESH (AUC 5 0.90), FRESH-cog (AUC 5 0.80), and FRESH-quol (AUC 5 0.78) was high.
External validation of our cohort using mRS as endpoint showed satisfactory results (AUC 5 0.77). To allow for con-
venient score calculation, we built a smartphone app available for free download.
Interpretation: FRESH is the first clinical tool to prognosticate long-term outcome after spontaneous SAH in a multi-
dimensional manner.
ANN NEUROL 2016;00:000–000

S pontaneous subarachnoid hemorrhage (SAH) is a

devastating condition with commonly reported in-
hospital mortality rates of about 20%,1 30-day case fatal-
ity of 29 to 38%,2,3 and markedly impaired long-term
outcome in up to half of survivors.4 Compared to other
stroke subtypes, the incidence of SAH is lower, but it
tends to affect people at a younger age, when permanent
disability has a particularly strong impact on patients and
their families.5 Ideally, prognostic models accurately esti-
mate the degree of long-term disability to which patients
and relatives will have to adjust. Primarily, such tools
serve the purpose of supporting families and physicians
to make ethically challenging decisions and secondarily
of guiding rational utilization of limited resources to

View this article online at wileyonlinelibrary.com. DOI: 10.1002/ana.24675

Received Aug 10, 2015, and in revised form Apr 19, 2016. Accepted for publication Apr 19, 2016.

Address correspondence to Dr Claassen, Division of Critical Care Neurology and Comprehensive Epilepsy Center, Neurological Institute,
Columbia University, 177 Fort Washington Avenue, MHB 8 Center, Room 300, New York, NY 10032. E-mail: jc1439@cumc.columbia.edu

From the 1Division of Critical Care Neurology, Department of Neurology, Columbia University, College of Physicians and Surgeons, New York, NY;
2
Division of Neurosurgery, St Michael’s Hospital, Labatt Family Centre of Excellence in Brain Injury and Trauma Research, Keenan Research Centre for
Biomedical Science and Li Ka Shing Knowledge Institute of St Michael’s Hospital, Institute of Medical Science, Department of Surgery, University of
Toronto, Toronto, Ontario, Canada; and 3Department of Neurosurgery, Columbia University, College of Physicians and Surgeons, New York, NY.

C 2016 American Neurological Association

V 1
ANNALS of Neurology
those patients who need them most.6 For patients with
spontaneous SAH, there is no widely accepted clinical
prognostication tool.7 The main limitations of previous
SAH scores are 2-fold: the expectation bias and limited
outcomes measures. First, in the majority of patients who
die in intensive care units (ICUs), care is withdrawn
before death,8 which strongly interferes with prognostic
models via the so-called self-fulfilling prophecy mecha-
nism or expectation bias.1,9–12 That means that physi-
cians may make the decision to withdraw care based on
certain diagnostic observations (eg, absent pupillary
reflexes), which then in prognostic models turn out to be
significantly associated with in-hospital mortality. How-
ever, in reality, it is primarily the physician’s assumption
that these observations indicate the presence of an irre-
versible condition that will eventually lead to the death
of the patient. Whether this assumption is true is dis-
guised by the deliberate change of the course of the ill-
ness, namely, withdrawal of care (WOC). In the present
study we therefore conducted our score performance
analyses twice, first on the complete patient cohort, and
then on a cohort excluding those patients for whom care
was withdrawn.
Second, in previous scores outcome measures were
heterogeneous, mostly being limited to the modified
Rankin scale (mRS).7 Although the mRS is a recom-
mended stroke-specific National Institutes of Health
common data element, single stroke scales measure lim-
ited aspects of outcome and should be complemented by
a quality of life–specific measure.13,14 Taking into
account strengths and weaknesses of previous models, we
propose a 12-month outcome prognostication tool for
patients with SAH, including 3 separate scores to prog-
nosticate outcomes in the dimensions of disability, cogni-
tion, and quality of life. All 3 scores are based on data
obtained within the first 48 hours of hospital stay,
because major decisions regarding goals of care are fre-
quently made within the first 2 days. Because our func-
tional prognostication score had several of its items
overlapping with the recently published HAIR score,1
designed to prognosticate in-hospital mortality after
SAH, we provide a direct comparison of both scores in
the Results section.
Patients and Methods
Study Population, Data Collection, and Clinical
Management
A total of 1,619 consecutive patients with spontaneous nontrau-
matic SAH admitted to the Columbia University Medical Cen-
ter (CUMC) Neurological Intensive Care Unit between July
1996 and March 2014 were prospectively enrolled in the
Columbia University SAH Outcomes Project (SHOP). SAH
2 Volume 00, No. 00
was diagnosed by computed tomography (CT) or xanthochro-
mia of the cerebrospinal fluid. Exclusion criteria were SAH
from trauma or rupture of an arteriovenous malformation,
admission >14 days after onset, and age < 18 years. Clinical
management followed the guidelines by the American Heart
Association as described previously.15 We developed the Func-
tional Recovery Expected after Subarachnoid Hemorrhage
(FRESH) score from patients who had complete data necessary
for score calculation (derivation cohort; n 5 1,526; 94% of all
patients in SHOP). For the 93 patients included in SHOP but
not included in the functional FRESH score derivation cohort,
the APACHE-II score was missing. In 24% and 23% of
patients, respectively, the mRS data at 3 and 12 months was
missing because patients and caretakers could not be reached
via telephone. Either death or change of address and telephone
number in the dynamic process of recovery may be a reason for
patients being lost to follow-up. We assessed mRS at discharge,
3 months, and 12 months after bleed onset. If a patient was
deceased at any given time point, an mRS value of 6 was
assigned for this and all subsequent outcome assessment time
points. For the purpose of score derivation, missing mRS data
were imputed for the time points 3 and 12 months after the
SAH.
Demographics, clinical variables, and functional, cogni-
tive, and quality of life outcomes were prospectively recorded
by a dedicated investigator. Hospital complications, entered in
the database, included “infections,” “neuropathology” (eg,
aneurysmal rebleed), “other pathology” (eg, cardiac/gastrointes-
tinal complications), and “abnormal values” (eg, hyponatremia).
Prospective data collection included the APACHE-II Physio-
logic score, which consists of 12 items with a maximum of 56
possible points. For the purpose of development of the FRESH
score, we used the APACHE-II Physiologic score without its
Glasgow Coma Scale (GCS) component, as described
previously.16,17
The institutional review board (IRB) at CUMC approved
the study protocol. Written informed consent was obtained
from patients or, if they were neurologically impaired, from
family members. A data sharing agreement was in place and
IRB-approved by both CUMC and St Michael’s Hospital.
Outcomes
The functional outcome measure was the mRS 12 months after
the SAH.13 The mRS is a global disability scale with scores
that range from 0 (no symptoms) to 6 (death). We used a lin-
ear model (LM) to prognosticate 12-month mRS. In addition,
we dichotomized the mRS to quantify score performance using
the area under the receiver operating characteristic curve (AUC)
and to facilitate clinical interpretation. For dichotomization,
poor outcome was defined as mRS of 4 to 6. Cognition 1 year
after the SAH was assessed using the Telephone Interview for
Cognitive Status (TICS). The TICS is a 5- to 10-minute test
administered via telephone. It assesses orientation, attention,
language, long-term memory, motor function, and verbal mem-
ory. It is scored from 0 (worst) to 51 (best). For score modeling
purposes the TICS was used as a linear outcome measure. To

quantify the performance of the cognitive prognostication score
(FRESH-cog), we dichotomized the TICS (to calculate the
AUC), and for clinical purposes we defined 3 TICS score
ranges to represent normal cognition, and moderate and severe
impairment of cognition. These ranges were based on previous
studies: we defined TICS-values of 32 and higher as “normal”18
and the range between 26 and 31 as “moderately impaired,”
and the cutoff for “severely impaired” cognition was set at 25
points.19
Long-term quality of life was assessed using the Sickness
Impact Profile (SIP).20 The SIP consists of 3 dimensions (Phys-
ical/Psychosocial/Independent categories) including a total of 12
categories with overall 136 statements, for example, “I am very
clumsy in body movements.” Here, we analyzed SIP scores in
the dimension Physical (SIP-phys; categories: Ambulation/
Mobility/Body Care and Movement). In parallel to the TICS,
the SIP was used as linear endpoint during the score modeling
process, in a dichotomized fashion to measure score perform-
ance, and categorically to illustrate clinical outcome. Based on
previous recommendations, we determined 3 SIP-phys score
ranges: good quality of life (SIP-phys 5 0–6),21 moderate qual-
ity of life impairment (SIP-phys 5 7–22), and severe quality of
life impairment (SIP-phys  23).22
Model Selection
For the prognostication of the mRS at 12 months we developed
a LM, in all patients with complete data (derivation cohort,
n 5 1,526). To determine the variables that best prognosticate
outcome after SAH, we used a mix of knowledge-based and
data-driven approaches. We selected 35 variables from demo-
graphics as well as early hospital complications. This selection
was based on previously described outcome predictors, espe-
cially considering those included in previously SAH scoring sys-
tems. These were identified by means of an extensive literature
search. We searched PubMed up to June 1, 2015, with the
terms “SAH” OR “subarachnoid hemorrhage” AND
“prediction scale” or “outcome scale” or “prediction score” or
“outcome score” or “risk stratification” for studies on dedicated
outcome prediction tools (combination of at least 2 predictor
variables, for example, GCS and age) to predict long-term out-
come (at least 3 months after the index event). The search pro-
duced 1,996 results; abstracts were screened for relevance, and
in cases of relevant or ambiguous or missing abstracts, the full
papers were read. We identified 14 studies reporting on clinical
SAH scoring systems (please see the Discussion section for
more details).
Examples of the selected variables include age, conditions
in the patient’s past medical history (myocardial infarction, con-
gestive heart failure, hypertension, diabetes mellitus), admission
Hunt & Hess, GCS, and APACHE-II Physiologic scores,
admission pupillary reactivity, features of initial or follow-up
brain imaging (ictal infarction, SAH blood volume, hydroceph-
alus, generalized cerebral edema, intraventricular hemorrhage
(IVH), intracerebral hemorrhage component, location/size of
the ruptured aneurysm), mode of aneurysm treatment (clipping
vs coiling of the aneurysm), and aneurysmal rebleed within 48
Month 2016
Witsch et al: FRESH SAH Score
hours of hospital admission. On these variables, we performed
an exhaustive search on all possible models using the R package
bestglm. This package enumerates all possible LMs using the
method of Tatar and Morgan and determines the Bayesian
Information Criterion (BIC) for each of these models.23 The
optimal model is determined as the one with the lowest sum of
squares and BIC. This optimal model underestimates the prob-
ability values for the regression coefficients. However, this
should not be a problem in the current application, because we
only selected the optimal variables using this method and
adjusted the variables for our final model based on data cluster-
ing using k-means. On these adjusted variables, a final LM was
run and the contribution of each variable to the FRESH score
was determined proportional to its explained variance. We
repeated all score derivation analyses treating variables as ordi-
nal instead of linear, which yielded very similar results (data
not shown).
We based our model on the amount of explained var-
iance and not on the beta weights of the LM, because this
resulted in a more stable score. The score contribution of each
variable was based on the explained variance in the LM. The
modeling process was done separately and independently for 3
scores with the endpoints function, cognition, and quality of
life.
Internal validation was performed on the point predic-
tions of the LM, as the FRESH scores are based on the variance
of each variable in the LM. We performed nonparametric boot-
strap using 500 repetitions. The mean absolute error (MAE)
for each repetition was obtained. We then determined the 95%
confidence interval (CI) to examine whether the observed MAE
was within these limits.
Statistical Analysis
Analyses were performed using R (v3.0.2, R Project). Data are
expressed as median and interquartile range. The functional
FRESH score was derived from the entire cohort (derivation
cohort). Score performance was evaluated both in the derivation
cohort and in those patients excluding WOC patients. The
AUC was calculated to assess score sensitivity and specificity for
prognostication of poor outcome. Additionally, the goodness-
of-fit measures Nagelkerke R2 and Cox/Snell R2 are reported
for this binary endpoint. For cognitive outcome and quality of
life subscores, AUCs were calculated for different cutoff values
within the TICS and SIP-phys scores. CIs of the AUCs were
generated through bootstrapping (1,000 repetitions). A proba-
bility value of <0.05 was considered statistically significant.
Twelve-month functional outcome was available in 77% of
patients, and 3-month outcome was available in 76%.
We performed multiple imputation to account for mRS
lost to follow-up.24 Using SPSS (IBM, Armonk, NY), we
imputed 10 complete data sets based on a subset of our data
(which included the variables: age, gender, ethnicity, aneurysm
size, cerebral infarction as hospital complication, hydrocephalus
as hospital complication, WOC, GCS on admission, pupillary
response on admission, IVH on admission CT, and discharge
mRS). We imputed the missing values of the 2 mRS variables
3
ANNALS of Neurology

(at 3- and 12-month follow-up) separately but based on the

TABLE 1. Baseline Characteristics of 1,526
same data subset. In the data subset, the respective mRS vari-
Patients with Spontaneous Subarachnoid Hemor-
able was the only one with larger amounts of missing data
rhage (Score Derivation Cohort)
(23.6% at 3 months and 22.7% at 12 months). All of the other
variables had very few missing data (<2%). Little’s MCAR test Characteristic Value
was not significant for the 3-month and the 12-month data sets
(p 5 0.2 and p 5 0.2, respectively), allowing the assumption Age, yr, mean 6 SD 55.3 6 14.5
that the data was missing at random. Under this assumption,
SPSS uses the Markov Chain Monte Carlo Method using linear Female, No. (%) 1,032 (67.8)
regression for continuous variables and logistic regression for Ethnicity, No. (%)
categorical variables to impute data. We used a total of 100 White 681 (44.6)
iterations for a total of 10 imputed data sets as previously rec-
ommended.25 Complete TICS and SIP-phys data were available
Black 269 (17.6)
in 56% and 59% of surviving patients, respectively. Imputation Asian 79 (5.2)
of data on that scale is not consensus, which is why we Hispanic 456 (29.9)
addressed the issue of missing TICS and SIP data by conduct-
Other 41 (2.7)
ing a follow-up bias analysis (please refer to Results section).
Transfers, No. (%) 1,253 (82.1)
External Validation Fisher Scale grade
We externally validated our score derivation cohort using a pro-
spectively collected cohort of patients from the “Clazosentan to 1 222 (14.5)
Overcome Neurological Ischemia and Infarction Occurring after 2 324 (21.2)
Subarachnoid Hemorrhage” (CONSCIOUS-1) study (valida- 3 748 (49.0)
tion cohort), which includes data from 413 patients with spon-
taneous SAH.26 Patients for this study were recruited in a total 4 219 (14.4)
of 52 centers in Israel, Europe, and North America. In contrast Hunt & Hess grade
to the derivation cohort, in the validation cohort clinical out- 1 359 (23.5)
come had been assessed at 3 months and the World Federation
of Neurosurgeons Score (WFNS) was determined on admission 2 300 (19.7)
instead of the Hunt & Hess score. In CONSCIOUS-1, data on 3 393 (25.8)
cognition and quality of life had not been collected. To com- 4 222 (14.5)
pare AUCs for the FRESH and HAIR scores, we used the
method described by DeLong et al.27 Because 12-month mRS
5 252 (16.5)
scores were not available in the external validation cohort, we GCS
could not externally validate the FRESH score itself, which was 3–8 429 (28.1)
designed to prognosticate 12-month outcomes. However, to
illustrate the generalizability of the cohort in which the FRESH
9–12 155 (10.2)
score was created, we created an additional score prognosticat- 13–15 938 (61.5)
ing 3-month mRS scores, which was then externally validated APACHE-II, median [IQR] 11 [7–19]
in the CONSCIOUS-1 cohort.
Aneurysm size, mm, mean 6 SD 7.9 6 5.3
Comparison with the HAIR Score Aneurysm > 10.0mm, No. (%) 227 (14.9)
We compared the FRESH and HAIR scores and their perform-
Management/aneurysm treatment
ances in both the derivation and validation cohorts.1 HAIR
consists of the 4 items Hunt & Hess grade on admission EVD 570 (37.4)
(grades 1–3, 0 points; grade 4, 1 point; grade 5, 4 points), age Clipping 842 (55.1)
(<60 years, 0 points; 60–79 years, 1 point; 80 years, 2 Coiling 299 (19.6)
points), IVH on admission CT (presence, 1 point; absence, 0),
EVD 5 external ventricular drain; GCS 5 Glasgow Coma
and aneurysmal rebleed within 24 hours (presence, 1 point;
Scale; IQR 5 interquartile range; SD 5 standard deviation.
absence, 0). HAIR scores range from 0 (best) to 8 (worst).

Results
Score Development
The FRESH score was developed in the derivation
cohort (n 5 1,526), that is, those patients of the SHOP
database with complete data necessary for score deriva-
tion. Cohort characteristics are shown in Table 1. The
in-hospital mortality was 291 (18%). In 10%, care was
withdrawn. The combination of Hunt & Hess grade,
age, APACHE-II Physiologic score (without the GCS

4 Volume 00, No. 00


FIGURE 1: Relative distribution of modified Rankin scale (mRS) by respective score values at 12 months after subarachnoid
hemorrhage. (A) FRESH, (B) HAIR, and (C) Hunt & Hess score values in all patients of the derivation cohort (A1–C1) and in an
analysis excluding patients in whom care was withdrawn (A2–C2).
component), and aneurysmal rebleed yielded the highest
explained variance in the LM of mRS at 12 months. All
other factors entered in the model were inferior in com-
parison, including mode of aneurysm treatment, which
had LM weights significantly different from 0 (patients
undergoing coiling had significantly worse functional
outcome than those who underwent clipping), but did
not improve the overall model performance (sum of
squares) when entered as a fifth variable.
The LM explained 46.6% of the variance. Internal
validation of the LM point prediction for mRS at 12
months was good. MAE in the full cohort was 1.171;
the 95% CI from nonparametric bootstrapping was from
1.166 to 1.189.
In the k-means cluster analysis, one prominent clus-
ter consisted of patients with high Hunt & Hess scores
(>3), high age (>70 years), and poor 12-month out-
Month 2016
Witsch et al: FRESH SAH Score
come. Another cluster consisted of patients with similar
Hunt & Hess scores, but of lower age. The latter patients
generally had a more favorable outcome. This finding
was used to fine-tune the variables in the FRESH score,
namely setting an age threshold of 70 years.
Based on the explained variance, we assigned score
values to the 4 items, allowing for a final sum score
between 1 and 9 according to the following formula:
ðH &H Þ2 1 APACHEphys
10 1a1rÞ
FRESH score ¼ 11 (1)
5
where H&H 5 Hunt & Hess score on admission, a 5
age (>70 years 5 9; 70 years 5 0), r 5 aneurysmal
rebleed (occurred 5 4; did not occur 5 0), and APA-
CHEphys 5 APACHE Physiologic score without GCS
component.16,17
5
ANNALS of Neurology
TABLE 2. Discrimination and Goodness-of-Fit Measures of FRESH, HAIR, and Hunt & Hess
FRESH HAIR
(derivation (derivation
cohorta) cohorta)
Discrimination
AUC 89.8% 88.3%
(88.1–91.6)b (86.4–90.2)b
Goodness of fit
Nagelkerke R2 0.50 0.45
2
Cox/Snell R 0.35 0.32
Prognostication of poor functional outcome 12 months after spontaneous subarachnoid hemorrhage. Poor functional outcome is
defined as modified Rankin scale 5 4–6.
a
Excluding patients that had care withdrawn
b
95% confidence intervals based on bootstrapping (1,000 repetitions).
AUC 5 area under the receiver operating characteristic curve.
The weighting of the score items in the formula is
a proportionate reflection of the explained variance of
each variable (eg, factors for age and rebleed, and divi-
sion of the APACHE score by 10). Hunt & Hess grade
was squared to reflect the exponential relationship
between incremental increase of Hunt & Hess scores and
frequency of poor outcome, apparent in raw data plots
of the derivation cohort (data not shown). The score was
divided by 5 for purely practical reasons, to allow for
score grades in the single-digit range.
Smartphone App
In exchange for the highest possible accuracy of the prognos-
tic model, the FRESH score sacrifices ease of calculation.
We therefore created a FRESH score application, that may
be downloaded to smartphones or tablets for free (https://
itunes.apple.com/us/app/fresh-score/id1015675236?mt=8).
It allows for score calculation within a few seconds and vis-
ualizes the likelihood of poor long-term functional, cogni-
tive, and quality of life outcomes. In addition to the mRS
score distribution for a given FRESH score, point estimates
with CIs for each possible combination of input are dis-
played in the smartphone app.
The FRESH Score in the Derivation and
Validation Cohorts
In an analysis of our 1,526 patients with the 9-point
FRESH score (Fig 1A1), poor 12-month outcome was
present in 3, 6, 12, 38, 61, 83, 92, 98, and 100%. We
compared our score to the recently published in-hospital
mortality prediction score HAIR1; patients of our cohort
with HAIR scores of 0 to 8 had poor outcome in 4, 11,
31, 66, 84, 81, 96, 94, and 100% (Fig 1B1). Finally, cat-
egorization according to initial Hunt & Hess scores of 1
6 Volume 00, No. 00
Hunt & Hess FRESH HAIR
(derivation (validation (validation
cohorta) cohort) cohort)
85.3% 73.2% 71.8%
(83.2–87.4)b (67.3–79.1)b (66.0–77.5)b
0.40 0.20 0.17
0.29 0.13 0.11
to 5 corresponded to poor outcome in 6, 9, 22, 50, and
88% of cases (Fig 1C1).
Correlation between FRESH score and mRS (Pearson
coefficient 5 0.67, p < 0.0001) and discrimination between
favorable and poor mRS grades were excellent; the latter was
also the case for the HAIR score (statistical comparison
according to DeLong et al27 for correlated ROC curves of
FRESH and HAIR: p 5 0.6). All performance measures of
FRESH, HAIR, and Hunt & Hess scores are shown in
Table 2. AUCs of the functional FRESH score in the deriva-
tion and validation cohorts are shown in Figure 2A.
To illustrate generalizability of the derivation
cohort, we created a 3-month score (which was only
used for this purpose and does not constitute the final
FRESH score). In an attempt to best compare our cohort
to the CONSCIOUS-1 cohort, we built the 3-month
FRESH using WFNS instead of Hunt & Hess. Other-
wise, the 3-month score was identical to the 12-month
score with the addition of IVH on admission CT. In con-
trast to the 12-month score, k-means clustering of features
revealed a large cluster of patients with poor outcome and
somewhat lower age than for the 12-month score. We
therefore used a binary variable of age > 65 years. The 3-
month FRESH score was therefore defined as:
FRESH validation score 5
ðWFNSÞ2 1 APACHEphys 1IVH 1a1r
10
5
where WFNS 5 World Federation of Neurosurgeons
Score on admission, IVH 5 intraventricular hemorrhage
on admission CT (present 5 2; absent 5 0), a 5 age
(>65 5 9; 65 5 0), r 5 aneurysmal rebleed (occurred
5 5; did not occur 5 0)

FIGURE 2: Receiver operating characteristic curves for prog-
nostication of poor outcome. (A) Poor functional outcome
(modified Rankin scale 5 4–6) by FRESH score in the deriva-
tion (upper panels) and in the validation (lower panel)
cohorts. Area under the curve (AUC) percentages are given
with confidence intervals. (B) Poor cognitive outcome by
FRESH cognition (FRESH-cog) score. AUC percentages are
given for different dichotomization cutoff values (below
respective value 5 poor cognitive outcome) of the Tele-
phone Interview for Cognitive Status (TICS) score (TICS 0 5
worst, TICS 51 5 best outcome). (C) Poor quality of life by
FRESH quality of life (FRESH-quol) score. AUC percentages
are given for dichotomization cutoff values (above respective
value 5 poor quality of life) of the Sickness Impact Profile–
Physical (SIP-phys) score (SIP-phys 0 5 best, SIP-phys 45 5
worst). All confidence intervals were generated from boot-
strapping (1,000 repetitions). WOC 5 withdrawal of care.
Using this score, we obtained a mean AUC of
0.878 in the derivation cohort. For the external valida-
tion cohort, we obtained an AUC of 0.769 (95% CI 5
Month 2016
Witsch et al: FRESH SAH Score
0.713–0.823). This AUC was significantly higher than
the one obtained using the HAIR score (AUC 5 0.722;
DeLong’s test for correlated ROC curves: Z 5 2.3865,
p 5 0.01701). It is of note that the DeLong tests com-
paring AUCs of FRESH and HAIR have to be inter-
preted with caution, because HAIR was optimized to
prognosticate in-hospital mortality, but not 3-month or
12-month mRS.
Regarding 12-month mRS, the Hunt & Hess score
alone yielded a reasonable to high overall performance at
predicting dichotomized favorable versus unfavorable
outcome. However, goodness of fit measures were poorer
than those of both the HAIR and the FRESH scores
(Table 2).
An external cohort including documentation on
long-term cognition and quality of life using the TICS
and SIP, respectively, was not available at this point.
Internal validations for these endpoints were achieved
using nonparametric bootstrapping, in the same manner
as for the functional FRESH score.
Expectation Bias Analysis
As the parameters included in our prognostication model
were all assessed within 48 hours of admission, it may be
argued that the bias introduced by the WOC practice
might have a negligible impact on our model. However,
a remaining bias, in particular of old age and high Hunt
& Hess score on admission, is likely present, independ-
ent of the timing of WOC relative to score assessment.
To address the bias associated with WOC in 10%
of the cohort, we report score distributions after exclu-
sion of WOC patients (Fig 1A2–C2). Patients with
respective FRESH scores 1 through 9 had poor outcome
in 3, 5, 10, 28, 60, 72, 86, 92, and 100%. Overall score
performance was nearly identical in the analysis of the
cohort from which WOC patients were excluded com-
pared to the entire cohort (Fig 2A; Cox/Snell R2 5 0.3,
Nagelkerke R2 5 0.5). Patients with HAIR scores of 0 to
8 in the WOC exclusion cohort had poor outcome in 4,
9, 23, 58, 76, 70, 93, 90, and 100% (AUC 5 0.83,
Cox/Snell R2 5 0.26, Nagelkerke R2 5 0.4). Patients
with initial Hunt & Hess scores of 1 to 5 had poor out-
come in 5, 7, 19, 42, and 77%.
Subgroup Analysis: False-Negative Predictions
Whereas 50% with Hunt & Hess grade 4 and 12% with
Hunt & Hess grade 5 had favorable outcome, only 5%
of patients with high FRESH scores (7–9) had favorable
outcome at 12 months (FRESH grade 7, n 5 4; grade
8, n 5 1; grade 9, n 5 0). We further analyzed patients
in whom the FRESH score clearly failed to prognosticate
outcome (poor outcome despite low FRESH score),
7
ANNALS of Neurology

hypothesizing that hospital complications accounted for

much of the poor outcome. Among patients with
FRESH scores of 1 to 3, 7% had poor outcome at 12
months (n 5 68); we assessed these patients for premor-
bid disability and clinical courses; they had higher pre-
morbid disability than the remainder of the cohort
(premorbid mRS > 0 in 16% vs 8%, Mann–Whitney U
test, p 5 0.02) and higher mRS scores at hospital dis-
charge (medians 5 vs 4, Mann–Whitney U test, p 5
0.004). The overall rate of hospital complications was
not significantly different between the 68 miscategorized
patients and the remaining cohort (94% vs 92%; p 5
0.9). However, the following severe complications were
significantly more frequent in the miscategorized
patients: transfusion-requiring anemia (p < 0.001), brain
herniation (p 5 0.007), cardiac arrest (p 5 0.001),
newly occurring congestive heart failure (p 5 0.04), cere-
bral infarction (p 5 0.006), high fever (>38.58C; p 5
0.02), hepatic failure (p 5 0.006), and newly occurring
subdural hematoma (p < 0.001).

Prognostication of Cognition and Quality of Life

We developed additional scores for prognostication of
cognitive outcome (FRESH-cog) and quality of life
(FRESH-quol) at 12 months. These scores were devel-
oped in all patients with available score derivation and
follow-up data at 12 months (n 5 699 for FRESH-cog
and n 5 401 for FRESH-quol). Score performances
were quantified in the same cohorts.
Although there is a large overlap between the varia-
bles of the 3 scores, the cognitive and quol subscores
both represent models generated independently from the
functional outcome model. In the final prognostic mod-
els, the addition of education to the FRESH-cog and
both education and premorbid disability to the FRESH-
quol yielded the highest explained variances, correspond-
ing to the formulas:
ðH &H Þ2 1 APACHEphys 1a1r22  eÞ
10
FRESH cog ¼ (2)
3
ðH &H Þ2 1 APACHEphys 1a1r2e2GOSÞ
10
FRESH quol ¼
5
(3)
where H&H 5 Hunt & Hess score on admission, a 5
age constant (>70 years 5 9; 70 5 0), r 5 aneurys-
mal rebleed constant (occurred 5 4; did not occur 5 0),
APACHEphys 5 APACHE Physiologic score without
GCS component,16,17 e 5 education in years (maximum
5 24 years), and GOS 5 premorbid Glasgow Outcome
Scale (range 5 1–5).
FIGURE 3: (A) Correlation of FRESH cognition (FRESH-cog)
values and Telephone Interview for Cognitive Status (TICS)
12 months after subarachnoid hemorrhage. Hexagonal bin-
ning plot with multiple counts coded by shading (see plot
legend). (B) Correlation of FRESH quality of life (FRESH-
quol) scores with Sickness Impact Profile–Physical (SIP-phys).
Negative values on the FRESH-cog and FRESH-quol scores
are possible, resulting from the subtraction of education
years from the FRESH score in a given study individual.
We found a negative linear correlation between
FRESH-cog and TICS at 12 months (Fig 3A, Pearson
coefficient 5 20.47, p < 0.001) and a positive correla-
tion between FRESH-quol and SIP-phys scores (Fig 3B,
Pearson coefficient 5 0.43, p < 0.001).
Overall, 13% had poor cognitive outcome (TICS
 25) and 11% had poor quality of life (SIP-phys 
23). In the modeling process TICS and SIP were treated
as linear scales; however, for clinical applicability we
defined outcome categories, and to quantify score per-
formances, as with the functional FRESH score, we
dichotomized the FRESH-cog and FRESH-quol scores.
Dichotomization is also preferred by some for clinical
categorization, because it provides a clear cutoff between
favorable and unfavorable outcomes. Based on the
respective distributions of FRESH-cog and FRESH-quol
scores in our cohort, we defined 3 score ranges (favor-
able, moderate, and poor outcome), for which we calcu-
lated the frequencies of 3 TICS and SIP-phys score

8 Volume 00, No. 00

 Witsch et al: FRESH SAH Score

TABLE 3. Prognostication of 12-Month Cognitive Status per FRESH-cog Score

FRESH-cog Score Severely Impaired Moderately Impaired Favorable

Cognition, Cognition, TICS 5 Cognition,
TICS £ 25, No. (%) 26–31, No. (%) TICS  32,
No. (%)

Favorable, score range -16 to -9 3 (1.8) 18 (10.6) 149 (87.6)

Moderate, score range -8 to 0 73 (12.0) 117 (37.3) 308 (50.7)
Poor, score range 1 to 8 15 (48.4) 10 (32.3) 6 (19.3)
FRESH-cog 5 FRESH cognition; TICS 5 Telephone Interview for Cognitive Status.

ranges, respectively, to facilitate clinical categorization of
patients (Tables 3 and 4). Depending on the dichotomi-
zation cutoff point between favorable and poor outcome,
the AUC for both FRESH-cog and FRESH-quol scores
varied (Fig 2B, C). When dichotomized at previously
recommended cutoff points19,22 (TICS  25 and SIP-
phys > 23), AUCs were 79.7% for FRESH-cog (95%
CI 5 75.2–84.2%) and 78.2% for FRESH-quol (95%
CI 5 71.3–85.2%).
The LMs, which our scores were based on,
explained 23.5% of the variance for SIP and 26.2% for
TICS. Internal validation of the LM point prediction for
TICS and SIP at 12 months was good. The MAEs in
the full cohort were 4.69 and 8.87, and the estimated
95% CIs were 4.65 to 4.8 and 8.31 to 9.19, respectively.
Cognitive and Quality of Life Outcomes:
Follow-up Bias Analysis
A univariate comparison of patients with completed and
missing cognitive and quality of life assessments (follow-up
bias analysis) was conducted. Patients with missing cogni-
tive outcome assessment were older (median 5 55 vs 52
years), were less frequently of Caucasian ethnicity (40% vs
46%), had a higher Hunt & Hess score on admission
(median 5 3 vs 2), had a higher APACHE-II score
(median 5 11 vs 8), had a lower GCS score on admission
(median 5 14 vs 15), and had a higher SAH sum score on
admission (median 5 14 vs 13). Among these patients the
presence of IVH on admission CT was more frequent
(50% vs 44%), they had a higher occurrence of delayed
cerebral ischemia (34% vs 25%), and they had a higher
mRS at discharge (median 5 4 vs 3). Similarly, patients
with missing 12-month quality of life outcome assessment
were less frequently Caucasian (37% vs 57%), had a higher
admission Hunt & Hess score (both medians 5 2) and
APACHE-II score (median 5 10 vs 8), had a lower admis-
sion GCS score (both medians 5 15), underwent aneu-
rysm clipping less frequently (61% vs 71%), and had
higher rates of delayed cerebral ischemia (32% vs 24%).
Discussion
Summary
Based on clinical information available within 48 hours
after admission for spontaneous SAH, FRESH prognosti-
cates long-term outcome 12 months later. FRESH is the
first SAH prognostication tool to combine functional
outcome13 with cognitive and quality of life outcomes,
the necessity for which has been pointed out previously.14
FRESH was developed in a large cross-sectional, prospec-
tively collected study population, and was validated both
internally and externally. Despite a high burden of
delayed hospital complications in patients with SAH,

TABLE 4. Prognostication of 12-Month QOL per FRESH-quol Score

FRESH-quol Score Severe QOL Moderate QOL Good QOL,

Impairment, Impairment, SIP-phys 5
SIP-phys  23, No. (%) SIP-phys 5 0–6, No. (%)
7–22, No. (%)

Favorable, score range 26 to 22 6 (3.4) 35 (19.6) 138 (77.1)

Moderate, score range 21 to 11 27 (14.7) 51 (27.9) 105 (57.4)
Poor, score range 2 to 8 16 (41.0) 8 (20.5) 15 (38.5)
FRESH-quol 5 FRESH quality of life; QOL 5 quality of life; SIP-phys 5 Sickness Impact Profile–Physical.

Month 2016 9
ANNALS of Neurology
FRESH demonstrated excellent performance in all of its
3 outcome dimensions: function, cognition, and quality
of life. We addressed the problem of expectation bias by
including an analysis from which WOC patients were
excluded. This analysis yielded comparable results to the
analysis in the entire derivation cohort, illustrating the
robustness of the FRESH score against bias that might
have arisen from the WOC practice (also known as
expectation bias). Clinical applicability of the FRESH
score is facilitated by a smartphone app.
Rationale for Choosing 9 FRESH Grades
Clinicians often think and communicate about their
patients’ clinical outcome using the categories “no or
mild,” “moderate,” and “severe” impairment. This intui-
tive categorization is contained in the number of FRESH
score grades, which are a multiple of 3. However, it was
important to the authors to differentiate further between
these 3 main outcome categories. The percentage with
poor 12-month outcome extracted from the raw data
illustrates that this approach might be reasonable.
Although the difference of patients with poor 12-month
outcome in the lower FRESH grades 1 to 3 (3, 6, and
12% respectively) and the higher FRESH grades 7 to 9
(92, 98, and 100%) is not overwhelming, we believe that
these differences are crucial in communications with
patients and relatives, and help to give a sense of the
degree of clinical acuity in low-grade SAH patients and a
sense of the chance of favorable outcome in high-grade
patients. These patient groups at the extremes of the
spectrum are the most challenging ones regarding clinical
decision making, whereas there is usually less debate
about moderately ill patients (FRESH grades 4–6), who
tend to receive all care and resources available in our
experience.
Previous Scores
To date, there is no generally accepted score to determine
the prognosis in patients after spontaneous SAH.
Although outcome may be approximated by admission
WFNS or Hunt & Hess grades, neither grading scale
was developed for that purpose. Although the overall per-
formance of the Hunt & Hess score in our cohort
seemed satisfactory by statistical measures, a significant
number of patients were miscategorized, illustrating the
inadequacy of Hunt & Hess grading as a prognostication
tool. Our results include a head-to-head comparison
between FRESH and the recently published HAIR scores,
because of several overlapping items between the two
scores, albeit with notably different weighting. The
weighting in HAIR does not distinguish between Hunt &
Hess grades 1 to 3, and includes the factor IVH on
10
admission CT, which was inferior to several other factors
in our prognostic model, particularly to the APACHE-II
score. In HAIR—for the sake of simplicity—both IVH
on admission CT and aneurysmal rebleed after admission
are weighted 0 (absent) and 1 (present), which does not
reflect the proportions demonstrated by our LM analysis.
HAIR is a priori limited by several factors. It was designed
to prognosticate in-hospital mortality, an endpoint that is
almost meaningless to both caregivers and the patients’ rel-
atives, because it is strongly influenced by the deliberate
decision to withdraw care. When using HAIR to deter-
mine the risk of 12-month poor outcome, its overall per-
formance was high, but discrimination between high
HAIR scores (5–8) was poor.9 Thus, HAIR did not allow
accurate prognostication in those patients whose clinical
courses usually pose the greatest ethical challenges. Hence,
the main limitation of HAIR, similar to Hunt & Hess
grading alone, is its high rate of false positively prognosti-
cated poor outcomes, which may entail fatal consequences
if clinical management is based on these prognoses. Hunt
& Hess grading alone showed the strongest association
with long-term outcome and demonstrated good perform-
ance at prognostication of dichotomized outcome, with
the upper 95% CI of the AUC even overlapping with the
lower 95% CI of the HAIR score. However, goodness of
fit of Hunt & Hess was poor (in comparison to both the
HAIR and the FRESH scores). These findings are in line
with previous studies from the 1990s showing that both
the Hunt & Hess and WFNS scales perform poorly at
differentiating between some of their adjacent scale
grades.28–31
Previous scores to approximate long-term outcome
prognosis (systematically reviewed by Jaja et al7) were
developed mostly in cohorts of <500 patients. Study
designs were single-center retrospective, single-center pro-
spective, or multicenter prospective, respectively.7,32 Most
SAH scores were assembled of 3 to 6 variables, usually
containing a combination of age with either Hunt &
Hess, WFNS, or GCS scores on admission. Further vari-
ables included Fisher grade or IVH on admission CT,
aneurysm size or location, and vasospasm on vessel imag-
ing, among others; all of these variables were considered
during FRESH score development, and found to be
clearly inferior to both APACHE-II score on admission
and aneurysmal rebleed. One previous prognostic score
was solely based on physiologic parameters collected dur-
ing the first 24 hours after admission, resulting in a good
overall performance at prognosticating poor outcome at
3 months (mRS 5 4–6, AUC 5 0.79, 95% CI 5 0.74–
0.85).17 However, no previous prognostic model inte-
grated both SAH-specific variables and an established
ICU assessment scale, such as the APACHE-II score.
Volume 00, No. 00

Among the larger studies, in cohorts of >500
patients,32–35 only the work by Rosen and Macdonald
unambiguously reported its applied methods and score
performance measures. This score was developed using
data from the landmark international study by the Inter-
national Subarachnoid Aneurysm Trial group.33 Its over-
all performance at prognosticating poor outcome at 3
months (AUC 5 0.78) was significantly stronger than
prognostication by WFNS alone (AUC 5 0.74). Despite
satisfactory performance, the score is not being widely
used, possibly because of its relative complexity, consist-
ing of a total of 8 variables.
There are different ways of displaying the output of
a clinical score. Whereas some highly cited previous
work, such as the CRASH and IMPACT score papers for
prognostication of outcome after traumatic brain injury
and a recent score paper for prognostication of long-term
aneurysm rupture risks, use point estimates with CIs to
display score outputs, we chose to display outcome
ranges for mRS, TICS, and SIP scores, to facilitate com-
munication with patients and relatives, and among physi-
cians.36–38 However, knowing that some researchers and
clinicians prefer point estimates as score output, we
included them in the FRESH smartphone app.
External Validation
We found excellent performance and internal validity of
FRESH in the derivation cohort and satisfactory per-
formance in the validation cohort. None of the prior
studies attempted internal and external validation proce-
dures. Performance differences between the derivation
and validation cohort include the finding that the clinical
scale to assess SAH severity differed, the former using the
Hunt & Hess and the latter the WFNS scales, respec-
tively. These scales are not exchangeable,28 but Hunt &
Hess performed better in the study cohort to prognosti-
cate outcome, and it was not feasible to retrospectively
generate Hunt & Hess scores for the external validation
cohort. In contrast to the derivation cohort, patients in
the validation cohort were selected in the setting of a
randomized controlled trial, which for example excluded
patients with vasospasm present on admission angiogra-
phy or individuals with cardiac failure requiring iono-
tropic support. Patients of the validation cohort were
younger (median age 5 51 vs 55 years in the derivation
cohort); the percentage of Caucasians was higher (89%
vs 44%), the rate of aneurysm clipping was lower (45%
vs 59%), the occurrence of aneurysmal rebleed was low-
er(<1% vs 9%), and IVH (36% vs 54%) and hydro-
cephalus on admission CT were less frequent (30% vs
38%).26 Finally, in the validation cohort the mRS was
assessed at 3, not at 12 months, a difference that is not
Month 2016
Witsch et al: FRESH SAH Score
negligible, and which we addressed by creating a separate
3-month mRS prognostication score in the derivation
cohort.39 Despite all these major differences in the
patient characteristics, successful external validation illus-
trates the generalizability of the FRESH score.
Additional FRESH Scores to Prognosticate
Cognition and Quality of Life
Crude functional outcome scales such as the mRS do not
adequately represent outcomes that are considered mean-
ingful to patients and families. Cognitive and quality of
life outcomes are increasingly recognized as fundamen-
tally important outcomes for acute brain injury
patients.14 To our knowledge, no previous study has pro-
posed a comprehensive scoring system to prognosticate
cognition or quality of life after spontaneous SAH. Varia-
bles associated with cognitive and quality of life out-
comes are partly overlapping with those associated with
disability, which explains the similarity of the 3 FRESH
scores, despite being developed independently. Additional
variables important for cognition include length of aca-
demic education and psychiatric comorbidity and cogni-
tion for quality of life.40,41 Performance of the FRESH
cognitive and quality of life scores on prognosticating
poor outcomes was excellent, comparable to the prognos-
tic accuracy of prior functional outcomes SAH scores.33
However, in communications with patients and relatives,
strictly dichotomized prognostic models are not helpful;
therefore, we introduced moderate outcome categories
for both FRESH-cog and FRESH-quol (Tables 3 and 4,
and smart phone app) to allow for a hierarchical range of
outcomes.
Limitations
Limitations of our study include a recruitment period for
the study cohort, resulting in variations of treatment pro-
tocols over the years.15 All patients were treated at a sin-
gle tertiary care center that receives most admissions
from outside hospitals, which might impact on the repre-
sentativeness of our cohort. Aneurysm clipping predomi-
nated in the study cohort, which may differ from other
centers. However, we replicated our results in the valida-
tion cohort, which consisted of a multicenter patient
population, recruited during a brief time period, and
which mostly received coiling as aneurysm treatment.
Additional limitations include incomplete cognitive and
quality of life follow-up data in approximately half of
surviving patients. We addressed this limitation by con-
ducting follow-up bias analyses; in our cohort, patients
with available data were less severely affected on average
(eg, they had lower admission Hunt & Hess and
APACHE-II scores), which indicates that overall long-
11
ANNALS of Neurology
term cognition and quality of life are probably overesti-
mated here. Clearly, we cannot exclude that follow-up
bias might have influenced the cognitive and quality of
life models that we calculated.
The FRESH-cog and FRESH-quol scores are more-
over limited by some of the circumstances that are also
limitations of previous functional SAH scores, such as
relatively high rates of false-positive and false-negative
predictions. This is because both cognition and quality
of life are by far more difficult to capture than functional
outcomes. Sophisticated and lengthy questionnaires are
necessary to reduce the amount of subjective judgment
and to make these outcomes comparable between indi-
viduals. Therefore, comprehensive and high-quality out-
come assessments in large enough cohorts are extremely
hard to obtain. However, we are hopeful that the results
obtained with our cognitive and quality of life FRESH
scores may be reproduced, and that these scores may
serve as a solid basis, open to modifications and improve-
ments in future studies.
Furthermore, the external cohort did not allow for
validation of the cognitive and quality of life FRESH
scores, which should also be the subject of future study.
Although we are confident that successful external valida-
tion of these additional scores is feasible, as this has not
been done at this point, we recommend refraining from
using the cognitive and quality of life measures in clinical
practice.
Acknowledgment
Supported by the Deutsche Forschungsgemeinschaft
(Research Fellowship Wi 4300/1-1, J.W.), Physicians
Services Incorporated Foundation (R.L.M.), Brain Aneu-
rysm Foundation (R.L.M.), Canadian Institutes for
Health Research (R.L.M.), and Heart and Stroke Foun-
dation of Canada (R.L.M.).
Author Contributions
J.W., H.-P.F., and J.C. contributed to conception and
design of the study; J.W., H.-P.F., S.Pat., S.L., J.M.S.,
S.A., M.C.F., A.V., B.J., R.L.M., E.S.C., and J.C. con-
tributed to data collection and analysis; J.W., H.-P.F.,
S.Par., and J.C. contributed to drafting the manuscript
and figures; S.Pat. designed the smartphone app.
Potential Conflicts of Interest
R.L.M. is Chief Scientific Officer of Edge Therapeutics.
Edge Therapeutics develops products intended to
improve outcomes after subarachnoid hemorrhage. J.C.
received honoraria from serving on the Advisory Board
of Actelion for study development. Actelion develops
12
products intended to improve outcomes after subarach-
noid hemorrhage.
References
1. Lee VH, Ouyang B, John S, et al. Risk stratification for the in-
hospital mortality in subarachnoid hemorrhage: the HAIR score.
Neurocrit Care 2014;21:14–19.
2. Gonzalez-Perez A, Gaist D, Wallander MA, et al. Mortality after
hemorrhagic stroke: data from general practice (the Health
Improvement Network). Neurology 2013;81:559–565.
3. Lovelock CE, Rinkel GJ, Rothwell PM. Time trends in outcome of
subarachnoid hemorrhage: population-based study and systematic
review. Neurology 2010;74:1494–1501.
4. Al-Khindi T, Macdonald RL, Schweizer TA. Cognitive and func-
tional outcome after aneurysmal subarachnoid hemorrhage. Stroke
2010;41:e519–e536.
5. Connolly ES Jr, Rabinstein AA, Carhuapoma JR, et al. Guidelines
for the management of aneurysmal subarachnoid hemorrhage: a
guideline for healthcare professionals from the American Heart
Association/American Stroke Association. Stroke 2012;43:1711–
1737.
6. Weinstein MC, Skinner JA. Comparative effectiveness and health
care spending—implications for reform. N Engl J Med 2010;362:
460–465.
7. Jaja BN, Cusimano MD, Etminan N, et al. Clinical prediction mod-
els for aneurysmal subarachnoid hemorrhage: a systematic review.
Neurocrit Care 2013;18:143–153.
8. Treece PD, Engelberg RA, Crowley L, et al. Evaluation of a stand-
ardized order form for the withdrawal of life support in the inten-
sive care unit. Crit Care Med 2004;32:1141–1148.
9. Witsch J, Lahiri S, Meyers E, et al. Comments on the risk stratifica-
tion for the in-hospital mortality in subarachnoid hemorrhage: the
HAIR score. Neurocrit Care 2015;23:142–143.
10. Naval NS, Kowalski RG, Chang TR, et al. The SAH score: a com-
prehensive communication tool. J Stroke Cerebrovasc Dis 2014;
23:902–909.
11. Lantigua H, Ortega-Gutierrez S, Schmidt JM, et al. Subarachnoid
hemorrhage: who dies, and why?. Critical care 2015;19:309.
12. Becker KJ, Baxter AB, Cohen WA, et al. Withdrawal of support in
intracerebral hemorrhage may lead to self-fulfilling prophecies.
Neurology 2001;56:766–772.
13. Saver JL, Warach S, Janis S, et al. Standardizing the structure
of stroke clinical and epidemiologic research data: the National Insti-
tute of Neurological Disorders and Stroke (NINDS) Stroke Common
Data Element (CDE) project. Stroke 2012;43:967–973.
14. Kasner SE. Clinical interpretation and use of stroke scales. Lancet
Neurol 2006;5:603–612.
15. Lord AS, Fernandez L, Schmidt JM, et al. Effect of rebleeding on
the course and incidence of vasospasm after subarachnoid hemor-
rhage. Neurology 2012;78:31–37.
16. Knaus WA, Draper EA, Wagner DP, Zimmerman JE. APACHE II: a
severity of disease classification system. Crit Care Med 1985;13:
818–829.
17. Claassen J, Vu A, Kreiter KT, et al. Effect of acute physiologic
derangements on outcome after subarachnoid hemorrhage. Crit
Care Med 2004;32:832–838.
18. Knopman DS, Roberts RO, Geda YE, et al. Validation of the tele-
phone interview for cognitive status-modified in subjects with nor-
mal cognition, mild cognitive impairment, or dementia.
Neuroepidemiology 2010;34:34–42.
Volume 00, No. 00

19. Welsh KA, Breitner JCS, Magruder-Habib KM. Detection of demen-
tia in the elderly using telephone screening of cognitive status.
Neuropsychiatry Neuropsychol Behav Neurol 1993;6:103–110.
20. Bergner M, Bobbitt RA, Kressel S, et al. The sickness impact pro-
file: conceptual formulation and methodology for the development
of a health status measure. Int J Health Serv 1976;6:393–415.
21. MacKenzie EJ, Burgess AR, McAndrew MP, et al. Patient-oriented
functional outcome after unilateral lower extremity fracture.
J Orthop Trauma 1993;7:393–401.
22. van Straten A, de Haan RJ, Limburg M, van den Bos GA. Clinical
meaning of the Stroke-Adapted Sickness Impact Profile-30 and
the Sickness Impact Profile-136. Stroke 2000;31:2610–2615.
23. Morgan JA, Tatar FA. Calculation of the residual sum of squares
for all possible regressions. Technometrics 1972;14:317–325.
24. Little RJ, D’Agostino R, Cohen ML, et al. The prevention and
treatment of missing data in clinical trials. N Engl J Med 2012;
367:1355–1360.
25. Stuart EA, Azur M, Frangakis C, Leaf P. Multiple imputation with
large data sets: a case study of the Children’s Mental Health Initi-
ative. Am J Epidemiol 2009;169:1133–1139.
26. Macdonald RL, Kassell NF, Mayer S, et al. Clazosentan to over-
come neurological ischemia and infarction occurring after subar-
achnoid hemorrhage (CONSCIOUS-1): randomized, double-blind,
placebo-controlled phase 2 dose-finding trial. Stroke 2008;39:
3015–3021.
27. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas
under two or more correlated receiver operating characteristic
curves: a nonparametric approach. Biometrics 1988;44:837–845.
28. Hirai S, Ono J, Yamaura A. Clinical grading and outcome after
early surgery in aneurysmal subarachnoid hemorrhage. Neurosur-
gery 1996;39:441–446; discussion 446–447.
29. Ogilvy CS, Carter BS. A proposed comprehensive grading
system to predict outcome for surgical management of
intracranial aneurysms. Neurosurgery 1998;42:959–968; discussion
968–970.
30. Gotoh O, Tamura A, Yasui N, et al. Glasgow Coma Scale in the
prediction of outcome after early aneurysm surgery. Neurosurgery
1996;39:19–24; discussion 24–25.
Month 2016
Witsch et al: FRESH SAH Score
31. Takagi K, Tamura A, Nakagomi T, et al. How should a subarach-
noid hemorrhage grading scale be determined? A combinatorial
approach based solely on the Glasgow Coma Scale. J Neurosurg
1999;90:680–687.
32. Jabbarli R, Reinhard M, Roelz R, et al. Early identification of indi-
viduals at high risk for cerebral infarction after aneurysmal subar-
achnoid hemorrhage: the BEHAVIOR score. J Cereb Blood Flow
Metab 2015;35:1587–1592.
33. Rosen DS, Macdonald RL. Grading of subarachnoid hemorrhage:
modification of the World Federation of Neurosurgical Societies
scale on the basis of data for a large series of patients. Neurosur-
gery 2004;54:566–575; discussion 575–576.
34. Niskanen MM, Hernesniemi JA, Vapalahti MP, Kari A. One-year
outcome in early aneurysm surgery: prediction of outcome. Acta
Neurochir (Wien) 1993;123:25–32.
35. Germanson TP, Lanzino G, Kongable GL, et al. Risk classification
after aneurysmal subarachnoid hemorrhage. Surg Neurol 1998;49:
155–163.
36. Steyerberg EW, Mushkudiani N, Perel P, et al. Predicting outcome
after traumatic brain injury: development and international valida-
tion of prognostic scores based on admission characteristics. PLoS
Med 2008;5:e165; discussion e165.
37. MRC CRASH Trial Collaborators, Perel P, Arango M, et al. Predict-
ing outcome after traumatic brain injury: practical prognostic
models based on large cohort of international patients. BMJ
2008;336:425–429.
38. Tominari S, Morita A, Ishibashi T, et al. Prediction model for 3-
year rupture risk of unruptured cerebral aneurysms in Japanese
patients. Ann Neurol 2015;77:1050–1059.
39. Korja M, Silventoinen K, Laatikainen T, et al. Cause-specific mor-
tality of 1-year survivors of subarachnoid hemorrhage. Neurology
2013;80:481–486.
40. Springer MV, Schmidt JM, Wartenberg KE, et al. Predictors of
global cognitive impairment 1 year after subarachnoid hemor-
rhage. Neurosurgery 2009;65:1043–1050; discussion 1050–1051.
41. Wong GK, Lam SW, Ngai K, et al. Development of a short form
of Stroke-Specific Quality of Life Scale for patients after aneurys-
mal subarachnoid hemorrhage. J Neurol Sci 2013;335:204–209.
13