Q J Med 2003; 96:67–74
doi:10.1093/qjmed/hcg008
The prognostic value of the components of the Glasgow
Coma Scale following acute stroke
C.J. WEIR, A.P.J. BRADFORD and K.R. LEES
From the Department of Medicine and Therapeutics, University of Glasgow, UK
Received 17 July 2002 and in revised form 21 October 2002
Summary
Background: The Glasgow Coma Scale (GCS) is
widely used in assessing level of consciousness. The
GCS verbal component may be misleading in acute
stroke: a focal neurological deficit leading to
dysphasia could affect the score, independently of
level of consciousness.
Aim: To investigate the relationship, in all strokes
and in dysphasic patients, between stroke outcome
and total GCS (with and without the verbal score)
and its components, to assess their relative values.
Study design: Retrospective analysis following pro-
spective data collection in an acute stroke unit and
follow-up.
Methods: Outcomes studied were 2-week mortality
and 3-month recovery (survival, subject living at
home). We used area under the receiver operating
characteristic curve (AUC) to compare versions of
the GCS and multivariate logistic regression to
Introduction
The Glasgow Coma Scale (GCS) was developed
to describe consciousness level in head-injured
patients.1 It measures the best eye, motor and verbal
responses, and is a widely used and accepted
prognostic score2 for both traumatic3 and non-
traumatic altered consciousness levels.4 The score
has been validated for its inter-observer reliability,5
which improves with training and experience.6
The assessment of consciousness level in acute
stroke is important for clinical management and as
an indicator of prognosis. Presence or absence of
coma is also used in selecting patients for clinical
Address correspondence to Dr C.J. Weir, Department of Medicine and Therapeutics, University of Glasgow,
Gardiner Institute, Western Infirmary, Glasgow G11 6NT. e-mail: c.j.weir@clinmed.gla.ac.uk
ß Association of Physicians 2003
identify which subset of GCS components best
predicted outcome.
Results: Of 1517 patients with acute stroke, 1217
had complete clinical and follow-up data; 349 were
dysphasic. Total GCS had greater AUC than the
GCS without the verbal score, for mortality (all
patients 0.78 vs. 0.76, p = 0.021; dysphasics 0.72
vs. 0.71, p = 0.52) and recovery (all patients 0.71
vs. 0.67, p = 0.0001; dysphasics 0.74 vs. 0.70,
p = 0.055). Verbal and eye scores independently
provided prognostic information, for each patient
group and outcome measure.
Conclusions: The GCS contains valuable predictive
information. Regardless of whether dysphasia is
present, the verbal score should be assessed since it
adds prognostic information to that from the eye
component, and has greater value than the motor
score.
trials because of its relationship with outcome. As
stroke may cause localized motor, speech or
language deficits, the accuracy of the GCS as a
measure of consciousness level may be affected. In
turn, its prognostic value may be impaired.
Conversely, in patients with a language disorder,
the verbal score may reflect stroke severity in
addition to its measurement of consciousness
level, and for that reason it may retain useful
prognostic information.
In head injury, the verbal component was shown
to be inaccurate in up to 10% of patients due to
68 C.J. Weir et al.
injury or intubation and provided no additional
prognostic information to motor and pupil res-
ponses.3 A recent evaluation of the admission GCS
in 275 acute stroke patients compared three alter-
native strategies for verbal scoring in intubated or
dysphasic patients.7 The total GCS score predicted
acute mortality with 88% accuracy, and the verbal
component could be excluded from the total GCS
score without loss of predictive value.
We sought to determine the prognostic value of
the total GCS score and its individual components
in a large cohort of patients with acute stroke. In
particular, we assessed the contribution of the
verbal component to outcome prediction in patients
with and without dysphasia. We also aimed to
identify the subset of the eye, verbal and motor
components which best predicted outcome in all
strokes and in patients with a language disorder.
Methods
We studied patients admitted to our acute stroke
unit between July 1990 and March 1995. We
included individuals with acute stroke who had not
previously been entered in the study. In all patients,
CT scanning was performed to exclude any non-
vascular cause of the neurological deficit. Medical
staff prospectively recorded demographic data and
clinical features of patients on admission, including
measurement of the GCS eye, motor and verbal
scores. The stroke unit admission policy aims for
patients to be admitted and assessed within 48 h
of stroke onset. Medical residents and a qualified
speech and language therapist tested for the
presence of dysphasia.
Outcome follow-up by record linkage8 to death
and hospital discharge records established patient
survival and placement. Death records were
obtained from the Registrar General of Scotland.
Record linkage uses a probability-matching algo-
rithm that gives low rates of false-positive and
false-negative links (about 1%).8 Admissions to
non-National Health Service hospitals and institu-
tions outside Scotland are not detected. Placement
was used as a surrogate marker of survival and
functional outcome combined: patients who were
alive and living at home or with relatives were
defined as having a good outcome; patients who
had died or were resident in institutional care were
considered to have a poor outcome. Patient survival
and placement were recorded, blind to the GCS
score, at several time points. Two-week survival
and placement at 3 months were selected to
represent the clinically relevant outcomes of early
mortality and medium-term recovery, respectively.
Statistical methods
Receiver operating characteristic (ROC) curves
were plotted to assess the prognostic value, for
2-week mortality and 3-month placement, of the
total GCS score, the total GCS excluding the verbal
component, and the eye, verbal and motor com-
ponents individually. For 2-week mortality, we
defined sensitivity and specificity for a cut-point
on a score as the respective proportions of deaths
and survivors that were correctly predicted.
Mortality was predicted in patients with a score
below the cut-point; survival was predicted in
patients having a score greater than or equal to
the cut-point. We used an analogous interpretation
of the cut-point for 3-month placement. We also
calculated the positive predictive value (proportion
of patients predicted to die or have a poor outcome
who actually died or had a poor outcome) and the
negative predictive value (proportion of patients
predicted to survive or have a good outcome who
actually survived or had a good outcome).
For each outcome measure and combination of
GCS components, we identified the optimal cut-
point which maximized the sum of sensitivity and
specificity. The area under the curve (AUC) and
its standard error were calculated9 to measure
the prognostic information provided by each com-
bination of GCS components. The AUCs were
compared using a test appropriate for correlated
samples.10 The correlations arose since several
combinations of GCS components were being
compared on the same group of subjects.
Forward stepwise logistic regression modelling
determined the subset of the GCS eye, motor and
verbal components that best predicted 2-week
mortality and 3-month placement. Effect sizes in
the logistic regression were expressed as odds ratios
and their 95%CIs. ROC curves were used to
illustrate the performance of the model with the
optimal subset of GCS components in each case. All
analyses were repeated in the subgroup of patients
with dysphasia.
The ROC analysis was carried out by Non-
parametric Receiver Operating Characteristic
Analysis software (version 2.5)11 and the stepwise
logistic regression modelling used SAS version 8.2
(SAS Institute) on a desktop PC.
Results
During the study, 1859 individuals were admitted
to the acute stroke unit, of whom 1517 had a
diagnosis of acute stroke. All components of the
GCS were recorded for 1232 patients; outcome data
 Assessing acute stroke 69

were available for 1217 (99%). Overall, 349 were 0.78 and 0.76, respectively (p = 0.021). The
patients (29%) were dysphasic. Table 1 shows the difference in area had similar magnitude for dys-
demographic and clinical features of the patients. phasic patients (AUCs 0.72 and 0.71; p = 0.52). For
Some 235 patients (19%) died after 2 weeks; the 3-month placement endpoint in the whole
537 (44%) had a poor outcome at 3 months. The patient group, the AUCs for the GCS scores with
corresponding figures for dysphasic patients were and without the verbal score, were 0.71 and 0.67,
95 deaths (27%) and 196 poor outcomes (56%). respectively (p = 0.0001). The corresponding areas
Each of the GCS components was strongly for dysphasic patients were 0.74 and 0.70
related to outcome (Table 2). The verbal compo- (p = 0.055). Figure 2 compares the ROC curves for
nent gave the greatest AUC for each patient group the summed GCS scores, including and excluding
and outcome measure, except for 2-week mortality the verbal component.
in dysphasic patients. Differences among compo- In the whole patient group, stepwise logistic
nents were statistically significant only for 3-month regression identified first the verbal and then the eye
placement in all patients. Figure 1 gives ROC components as independent predictors of 2-week
curves for the eye, verbal and motor components mortality and 3-month placement (Table 3). In
for prediction of mortality and placement in each dysphasic patients, the eye component entered the
patient group. model for 2-week mortality first, followed by the
For 2-week mortality in the whole patient group, verbal score. The motor component, while itself
the areas under the ROC curve for the summed GCS associated with outcome, did not add statistically
scores with and without the verbal component, significant predictive information to these models.
The motor score only provided prognostic informa-
Table 1 Patient characteristics tion independently of the verbal and eye compo-
nents in the prediction of placement in dysphasic
Variable n (%) patients (Table 3).
Further ROC curve analysis compared the ver-
Age* 71 (62-79) sions of the GCS obtained at each step of the logistic
Male sex 595 (49)
regression modelling. This identified whether the
Side of symptoms
statistically significant term added to the model at
Left 563 (48)
Right 578 (50) each step substantially improved predictive accu-
Bilateral 20 (2) racy. The AUC became progressively greater as
components were added to the score, for all patient
Oxfordshire Community Stroke
groups and outcome measures (Figure 3, Table 4).
Project clinical classification12
In prediction of 3-month placement in dysphasic
Total anterior 309 (25)
circulation syndrome patients, the AUC of 0.74 after addition of the eye
Partial anterior 402 (33) component (Figure 3d) was not substantially greater
circulation syndrome than the AUC of 0.73 for the total verbal and motor
Posterior circulation 143 (12) score, although the difference was statistically
syndrome significant (p = 0.037). Table 4 shows the sensi-
Lacunar syndrome 348 (29) tivity, specificity, and positive and negative pre-
Other 11 (1) dictive values for the optimal cut-point on each
Nature of acute cerebrovascular event score. For the whole patient group the optimal cut-
Haemorrhagic stroke 158 (13) points were the maximum scores (5 for the verbal
Ischaemic stroke 1059 (87) component and 9 for the sum of verbal and eye).
Use of the verbal component on its own provided
*Median (IQR). good specificity but low sensitivity to poor outcome

Table 2 AUCs for individual GCS component ROC curves

Endpoint Patient group Eye Verbal Motor p*

2-week mortality All 0.71 0.74 0.71 0.09

Dysphasic 0.66 0.66 0.67 0.92
3-month placement All 0.63 0.68 0.64 0.0001
Dysphasic 0.64 0.69 0.67 0.20

*Tests whether or not all three AUCs are equal.

70
Figure 1. ROC curves for the individual GCS components. a 2-week mortality wall patientsx, b 2-week mortality wdysphasic
patientsx, c 3-month placement wall patientsx, d 3-month placement wdysphasic patientsx.
and mortality. Sensitivity for 2-week mortality and
poor outcome at 3 months was greatly improved
when a combined verbal and eye score was used,
although specificity was reduced. Optimal cut-
points were lower in the dysphasic subgroup,
although the patterns of predictive accuracy were
similar. Inclusion of the motor component did not
substantially improve the prediction of 3-month
placement in dysphasic patients, despite its
statistical significance in logistic regression
modelling.
Discussion
Prognostic indicators have an important role in
clinical management. It is important to identify
patients with very poor prognosis who are unlikely
C.J. Weir et al.
to benefit from an effective treatment: patients who
would otherwise have died may benefit only by
surviving, but may be completely dependent and
have a poor quality of life. Ideally, any method of
predicting outcome should be simple, accurate and
reproducible. The focal nature of acute stroke
brings into question the validity of the GCS as a
measure of level of consciousness and a predictor of
outcome. Can information from the GCS be
employed to determine the likely outcome from
acute stroke?
Various methods exist for assessing level of con-
sciousness in acute stroke clinical trials. Some trials
have used adaptations of the level of consciousness
assessments in the NIH,13,14 Scandinavian15 and
European16 stroke scales, while another defined
coma as inadequate motor response to painful
stimuli.17 The GCS was developed because of a
similar difficulty in defining level of consciousness
 Assessing acute stroke 71

Figure 2. ROC curves for total GCS scores with and without the verbal component. a 2-week mortality wall patientsx,
b 2-week mortality wdysphasic patientsx, c 3-month placement wall patientsx, d 3-month placement wdysphasic patientsx.

Table 3 Multivariate stepwise logistic regression analysis using individual GCS components

Logistic regression model Component Order of entry OR (95%CI)*

to the model

2-week mortality
All patients Verbal 1 1.52 (1.37–1.68)
Eye 2 1.62 (1.37–1.92)
Dysphasic patients Eye 1 1.58 (1.18–2.11)
Verbal 2 1.30 (1.11–1.52)
3-month placement
All patients Verbal 1 1.46 (1.33–1.60)
Eye 2 1.64 (1.34–2.01)
Dysphasic patients Verbal 1 1.29 (1.12–1.49)
Motor 2 1.45 (1.09–1.92)
Eye 3 1.77 (1.02–3.06)

*Odds ratio for survival (2-week mortality) or good outcome (3-month placement) per additional point on the score.
72 C.J. Weir et al.

Figure 3. ROC curves for combinations of GCS components at each stage of the stepwise logistic regression modelling.
a 2-week mortality wall patientsx, b 2-week mortality wdysphasic patientsx, c 3-month placement wall patientsx, d 3-month
placement wdysphasic patientsx. The numbers adjacent to the ROC curves show the best cut-point for each GCS version.

in head injury, and has subsequently been shown to
predict outcome well following head injury.
Teasdale recommends that the GCS components
should not be summed, as they are not equiva-
lent.18 Jennett and Teasdale (1977) noted, however,
that all of the combinations of eye, verbal and
motor findings which lead to a total GCS score
of -8 meet the recognized definition of coma.19
Our ROC curve analysis in all patients shows that
the total GCS score contains valuable prognostic
information following acute stroke, and predicts
early mortality more accurately than 3-month
outcome. A cut-point of 15 was best, giving a
simple interpretation: any impairment of conscious-
ness level predicts poorer outcome. Predictive
accuracy was also good in the dysphasic subgroup.
In the whole patient group, multivariate analysis
of the individual GCS components identified the
verbal and eye components as the best predictors of
clinically relevant outcomes. Addition of the eye to
the verbal component improved sensitivity although
specificity was reduced. The motor score did not
add prognostic information to the regression
models. In contrast, previous stroke studies found
that the verbal component was of little value and
the motor component predicted outcome more
accurately.7,20 Both of these studies were relatively
small and one7 measured outcome at hospital
discharge or transfer, rather than at a fixed time
point after the stroke.
Focal deficits in acute stroke may affect compo-
nent scores without reflecting the true conscious-
ness level. As with speech, the motor score may be
affected by focal deficit: severe or posterior
circulation strokes may cause bilateral limb weak-
ness. Another possible confounding factor is the
 Assessing acute stroke 73

Table 4 Predictive accuracy of the best cut-point for various GCS scores for 2-week mortality and 3-month placement

Patient Prevalence Score AUC Sensitivity Specificity PPV NPV

group of mortality/ (optimal (95%CI) (95%CI) (95%CI) (95%CI)
poor cut-point)
outcome

2-week mortality
All patients 0.19 V (5) 0.74 0.66 (0.60–0.72) 0.78 (0.76–0.81) 0.43 (0.38–0.48) 0.90 (0.88–0.92)
0.19 EqV (9) 0.78 0.74 (0.68–0.79) 0.76 (0.73–0.79) 0.42 (0.38–0.47) 0.92 (0.90–0.94)
Dysphasic 0.27 E (4) 0.66 0.44 (0.35–0.54) 0.87 (0.83–0.91) 0.57 (0.45–0.67) 0.81 (0.76–0.85)
patients 0.27 EqV (7) 0.70 0.68 (0.59–0.77) 0.65 (0.59–0.71) 0.42 (0.35–0.50) 0.85 (0.79–0.89)
3-month placement
All patients 0.44 V (5) 0.68 0.49 (0.45–0.53) 0.86 (0.83–0.88) 0.73 (0.68–0.77) 0.68 (0.65–0.71)
0.44 EqV (9) 0.70 0.55 (0.51–0.59) 0.84 (0.81–0.86) 0.73 (0.68–0.77) 0.70 (0.67–0.73)
Dysphasic 0.56 V (3) 0.69 0.58 (0.51–0.65) 0.76 (0.68–0.82) 0.76 (0.68–0.82) 0.59 (0.52–0.65)
patients 0.56 VqM (9) 0.73 0.62 (0.55–0.69) 0.74 (0.66–0.80) 0.75 (0.68–0.81) 0.60 (0.53–0.67)
0.56 EqVqM 0.74 0.69 (0.62–0.75) 0.67 (0.60–0.74) 0.73 (0.66–0.79) 0.63 (0.55–0.70)
(14)

AUC, area under curve; E, eye component of GCS; EqV, sum of eye and verbal GCS components; VqM, sum of verbal and
motor GCS components; EqVqM, sum of all GCS components; PPV, positive predictive value; NPV, negative predictive
value.

erroneous recording of the motor score from the
paretic limb, rather than observing the best
response. Because of the effect of focal deficits, it
would superficially seem reasonable to exclude the
verbal score in dysphasic stroke patients.
Our analysis for the dysphasic subgroup showed
that the verbal component provided additional
prognostic information to the combined eye and
motor scores. Our results suggest that when a
language disorder is absent, the verbal score
contributes prognostic information by measuring
level of consciousness or by acting as a marker
for confusion. In dysphasic patients, the verbal
component may reflect stroke severity and hence
predicts outcome. The motor component, whilst
itself a prognostic indicator, did not add sub-
stantially to the predictive accuracy of the com-
bined eye and verbal scores in the ROC analysis.
This finding differs from the results in head
injury.3
The GCS score predicts both 2-week mortality
and 3-month placement after stroke. The verbal
score contains valuable prognostic information and
should be recorded even for dysphasic patients. We
have shown that the maximum verbal score of 5,
corresponding to an orientated verbal response,
accurately predicts 2-week survival and good
3-month outcome. A score below 5 is an acceptable
predictor of 2-week mortality, but is not as accurate
in predicting poor 3-month outcome. A simple test
for an orientated verbal response would be effec-
tive in selecting, prior to clinical trial entry, patients
likely to survive the early effects of the stroke, since
the negative predictive value lies between 0.88 and
0.92. Summing the verbal and eye scores and using
a cut-point of 9 (unimpaired vs. impaired con-
sciousness) improves sensitivity to mortality and
poor outcome but slightly reduces specificity to
survival and good outcome. Using all three of the
GCS components is helpful in describing the
clinical status of a patient, but unnecessary for
prognostic purposes.
The optimal cut-points we identified require
confirmation in other populations, since positive
and negative predictive values depend on the
mortality and poor outcome rates in a patient
population. The 3-month outcome measure also
has limitations, as placement may be influenced
by factors unrelated to functional outcome. Availa-
bility of a family carer may enable a patient to
continue living at home. Placement in institutional
care may be a less relevant outcome in countries
other than the UK; for example, in Southern Europe,
care for a disabled relative conventionally takes
place within the family home.
We have demonstrated a strong relationship
between the verbal and eye GCS scores and
outcome in the acute stroke population. However,
the positive predictive values for 2-week mortality
and negative predictive values for 3-month place-
ment range from 0.42 to 0.70 and are not
sufficiently high to be used as the sole basis for
clinical decision-making on the individual patient.
It would thus be preferable to combine GCS data in
a model with other stroke prognostic factors if they
were to be used in patient management.
74 C.J. Weir et al.
Acknowledgements
Dr Weir was supported in this work by a Special
Training Fellowship in Health Services Research
(Medical Research Council, UK). Chris Povey of the
National Health Service (Scotland) Information and
Statistics Division performed the record linkage
analysis. The authors are grateful to Professor
J.L. Reid, Dr G.T. McInnes and Dr P.F. Semple for
permission to use data from patients under their care.
References
1. Teasdale G, Jennett B. Assessment of coma and impaired
consciousness: a practical scale. Lancet 1974; ii:81–4.
2. Teasdale G, Gentleman D. The description of ‘conscious
level’: a case for the Glasgow Coma Scale. Scott Med J 1982;
27:7–9.
3. Jagger J, Jane JA, Rimel R. The Glasgow Coma Scale: to sum
or not to sum? Lancet 1983; ii:97.
4. Levy DE, Bates MB, Caronna JJ, Cartlidge NEF, Knill-Jones
RP, Lapinski RH, Singer BH, Shaw DA, Plum F. Prognosis in
nontraumatic coma. Ann Intern Med 1981; 94:293–301.
5. Teasdale G, Knill-Jones R, van der Sande J. Observer
variability in assessing impaired consciousness and coma.
J Neurol Neurosurg Psychiatry 1978; 41:603–10.
6. Rowley G, Fielding K. Reliability and accuracy of the
Glasgow Coma Scale with experienced and inexperienced
users. Lancet 1991; 337:535–8.
7. Prasad K, Menon GR. Comparison of the three strategies of
verbal scoring of the Glasgow Coma Scale in patients with
stroke. Cerebrovascular Diseases 1998; 8:79–85.
8. Kendrick S, Clarke J. The Scottish record linkage system.
Heath Bull (Edinb) 1993; 51:1–15.
9. Hanley JA, McNeil BJ. The meaning and use of the area
under a receiver operating characteristic curve. Radiology
1982; 143:29–36.
10. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the
areas under two or more correlated receiver operating
characteristic curves: a nonparametric approach. Biometrics
1988; 44:837–45.
11. Vida S. Nonparametric Receiver Operating Characteristic
Analysis. Version 2.5. Montreal, McGill University,
1993.
12. Bamford JM, Sandercock PAG, Dennis MS, Burn J,
Warlow CP. Classification and natural-history of clinically
identifiable subtypes of cerebral infarction. Lancet 1991;
337:1521–6.
13. Lees KR, Asplund K, Carolei A, Davis SM, Diener HC,
Kaste M, Orgogozo JM, Whitehead J. Glycine antagonist
(gavestinel) in neuroprotection (GAIN International) in
patients with acute stroke: a randomised controlled trial.
Lancet 2000; 355:1949–54.
14. Bradford A, for the IMAGES Study Group. Intravenous
Magnesium Efficacy in Stroke trial and MRI substudy
(IMAGES & MR IMAGES). Cerebrovascular Diseases 2000;
10:80.
15. Lees KR, for the 256–301 Multi-centre study group UK.
UK-279,276 in acute stroke—a phase IIa safety study. Stroke
2000; 31:2866.
16. Diener HC, for the European, Australian Lubeluzole
Ischaemic Stroke Study Group. Multinational randomised
controlled trial of lubeluzole in acute ischaemic stroke.
Cerebrovascular Diseases 1998; 8:172–81.
17. Peters GR, Hwang L-J, Musch B, Brosse DM, Orgogozo JM.
Safety and efficacy of 6mg/kg/day tirilizad mesylate in
patients with acute ischemic stroke (TESS study). Stroke
1996; 27:195.
18. Teasdale G, Jennett B, Murray L, Murray G. Glasgow coma
scale: to sum or not to sum. Lancet 1983; ii:678.
19. Jennett B, Teasdale G. Aspects of coma after severe head
injury. Lancet 1977; i:878–81.
20. Diringer MN, Edwards DF. Does modification of the
Innsbruck and the Glasgow Coma Scales improve their
ability to predict functional outcome? Arch Neurol 1997;
54:606–11.