Articles

The Edinburgh CT and genetic diagnostic criteria for lobar

intracerebral haemorrhage associated with cerebral amyloid
angiopathy: model development and diagnostic test
accuracy study
Mark A Rodrigues, Neshika Samarasekera, Christine Lerpiniere, Catherine Humphreys, Mark O McCarron, Philip M White, James A R Nicoll,
Cathie L M Sudlow, Charlotte Cordonnier, Joanna M Wardlaw, Colin Smith, Rustam Al-Shahi Salman

Summary
Lancet Neurol 2018; 17: 232–40 Background Identification of lobar spontaneous intracerebral haemorrhage associated with cerebral amyloid
Published Online angiopathy (CAA) is important because it is associated with a higher risk of recurrent intracerebral haemorrhage than
January 10, 2018 arteriolosclerosis-associated intracerebral haemorrhage. We aimed to develop a prediction model for the identification
http://dx.doi.org/10.1016/
S1474-4422(18)30006-1
of CAA-associated lobar intracerebral haemorrhage using CT features and genotype.
See Comment page 197
Methods We identified adults with first-ever intracerebral haemorrhage diagnosed by CT, who died and underwent
Centre for Clinical Brain
Sciences (M A Rodrigues FRCR, research autopsy as part of the Lothian IntraCerebral Haemorrhage, Pathology, Imaging and Neurological Outcome
N Samarasekera PhD, (LINCHPIN) study, a prospective, population-based, inception cohort. We determined APOE genotype and radiologists
C Lerpiniere RGN, rated CT imaging appearances. Radiologists were not aware of clinical, genetic, and histopathological features. A
C Humphreys MBChB,
neuropathologist rated brain tissue for small vessel diseases, including CAA, and was masked to clinical, radiographic,
Prof C L M Sudlow DPhil,
Prof J M Wardlaw MD, and genetic features. We used CT and APOE genotype data in a logistic regression model, which we internally
Prof C Smith FRCPath, validated using bootstrapping, to predict the risk of CAA-associated lobar intracerebral haemorrhage, derive diagnostic
Prof R Al-Shahi Salman PhD), UK criteria, and estimate diagnostic accuracy.
Dementia Research Institute at
The University of Edinburgh
(Prof J M Wardlaw), Row Fogo Findings Among 110 adults (median age 83 years [IQR 76–87], 49 [45%] men) included in the LINCHPIN study between
Centre for Research into Ageing June 1, 2010 and Feb 10, 2016, intracerebral haemorrhage was lobar in 62 (56%) participants, deep in 41 (37%), and
and the Brain infratentorial in seven (6%). Of the 62 participants with lobar intracerebral haemorrhage, 36 (58%) were associated with
(Prof J M Wardlaw), and Usher
moderate or severe CAA compared with 26 (42%) that were associated with absent or mild CAA, and were independently
Institute of Population Health
Sciences and Informatics associated with subarachnoid haemorrhage (32 [89%] of 36 vs 11 [42%] of 26; p=0·014), intracerebral haemorrhage with
(Prof C L M Sudlow), The finger-like projections (14 [39%] of 36 vs 0; p=0·043), and APOE ε4 possession (18 [50%] of 36 vs 2 [8%] of 26; p=0·0020).
University of Edinburgh, A prediction model for CAA-associated lobar intracerebral haemorrhage using these three variables had excellent
Edinburgh, UK; Department of
Neurology, Altnagelvin
discrimination (c statistic 0·92, 95% CI 0·86–0·98), confirmed by internal validation. For the rule-out criteria, neither
Hospital, Londonderry, UK subarachnoid haemorrhage nor APOE ε4 possession had 100% sensitivity (95% CI 88–100). For the rule-in criteria,
(M O McCarron MD); Institute of subarachnoid haemorrhage and either APOE ε4 possession or finger-like projections had 96% specificity (95% CI 78–100).
Neuroscience and Institute for
Ageing, Newcastle University,
Newcastle-upon-Tyne, UK
Interpretation The CT and APOE genotype prediction model for CAA-associated lobar intracerebral haemorrhage
(Prof P M White FRCR); shows excellent discrimination in this cohort, but requires external validation. The Edinburgh rule-in and rule-out
Newcastle upon Tyne Hospitals diagnostic criteria might inform prognostic and therapeutic decisions that depend on identification of CAA-associated
National Health Services lobar intracerebral haemorrhage.
Foundation Trust,
Newcastle-upon-Tyne, UK
(Prof P M White); Clinical Funding UK Medical Research Council, The Stroke Association, and The Wellcome Trust.
Neurosciences, Clinical and
Experimental Sciences, Copyright © The Author(s). Published by Elsevier Ltd. This is an Open Access article under the CC BY 4.0 license.
University of Southampton,
Southampton, UK
(Prof J A Nicoll FRCPath); and Introduction risk of recurrent intracerebral haemorrhage and post-
Université Lille, Inserm U1171, About 85% of spontaneous intracerebral haemorrhages stroke dementia than arteriolosclerosis-associated intra­
Degenerative and Vascular have no underlying macrovascular cause and are cerebral haemorrhage,5,6 and might increase the risk of
Cognitive Disorders, CHU Lille,
Department of Neurology, Lille,
attributed to small vessel disease, mostly arteriolosclerosis intracerebral haemorrhage in patients taking anti­
France (Prof C Cordonnier PhD) with or without cerebral amyloid angiopathy (CAA).1,2 thrombotic drugs.7 Criteria to rule out CAA underlying
Correspondence to: CAA affects cortical and leptomeningeal vessels and is intracerebral haemorrhage would allow clinicians to be
Prof Rustam Al-Shahi Salman, only associated with lobar intracerebral haemorrhage,3,4 more confident about the use of antithrombotic drugs;2,7
Centre for Clinical Brain Sciences, whereas arteriolosclerosis can cause intracerebral ruling in CAA underlying intracerebral haemorrhage
University of Edinburgh,
Edinburgh EH16 4SB, UK
haemorrhage anywhere in the brain. would provide important prognostic information.
[email protected] Identification of CAA-associated intracerebral haemor­ The MRI-based modified Boston criteria have excellent
rhage is important because it is associated with a higher sensitivity and good specificity for CAA.8 However, MRI

232 www.thelancet.com/neurology Vol 17 March 2018

 Articles

Research in context
Evidence before this study Added value of this study
We did a systematic review of studies on imaging features of This diagnostic test accuracy study minimised biases, used
lobar or cerebellar intracerebral haemorrhage with pathologically masking, assessed inter-rater and intra-rater agreement,
proven cerebral amyloid angiopathy (CAA) published in standardised index test and reference standard, and adhered to
MEDLINE (from 1946 to Nov 1, 2016) and Embase (from 1974 to recommended approaches for analysis. We were able to
Nov 1, 2016) using comprehensive electronic search strategies develop a highly discriminatory and well calibrated prediction
combining terms “stroke”, “cerebrovascular disorders”, (brain$ or model using subarachnoid haemorrhage and finger-like
cerebr$ or intracerebr$) adj5 (h?emorrhag$ or h?ematoma$), projections from intracerebral haemorrhage on CT, and APOE
“amyloid beta-protein”, “cerebral amyloid angiopathy”, “vascular ε4 possession, which was internally validated. We identified
amyloidosis”, “congo red”, “pathology, clinical”, “pathology”, clinically useful probability cutoffs and two sets of diagnostic
“histo?patholog$”, “post?mortem$” and “autops$” with no criteria that can rule in or rule out CAA-associated lobar
language restriction. We identified 22 case series describing intracerebral haemorrhage.
imaging features of lobar or cerebellar intracerebral haemorrhage
Implications of all the available evidence
accompanying histopathologically proven CAA. Overall, the
Both the Boston MRI and the Edinburgh CT-based diagnostic
study quality was poor, with small sample sizes, unclear
criteria for CAA-associated lobar intracerebral haemorrhage are
definitions of predictor or outcome variables, and infrequent
now available. The Edinburgh sensitive rule-out criteria and
masking of study assessors. The most frequently reported CT
specific rule-in criteria based on CT and APOE genotype are
features of CAA-associated intracerebral haemorrhage in 21 case
potentially widely applicable for diagnostic, prognostic, and
series were subarachnoid extension and an irregular intracerebral
therapeutic decisions in everyday clinical practice if MRI is
haemorrhage border. No diagnostic test accuracy studies have
contraindicated, intolerable, or unavailable. Future research is
been done. Although the modified Boston MRI criteria are widely
required to externally validate these diagnostic criteria and
used for the identification of CAA-associated intracerebral
evaluate their clinical use.
haemorrhage, no CT-based diagnostic criteria exist for patients
who cannot tolerate or do not have access to MRI.

can be unsuitable for very unwell patients in the acute Lothian health board region of Scotland (the Lothian
setting or for those patients with contraindications, IntraCerebral Haemorrhage, Pathology, Imaging
such as non-MRI compatible implanted devices or and Neurological Outcome [LINCHPIN] study).12
claustrophobia, and might be unavailable particularly in We prospectively identified all incident cases of
low-income and middle-income countries where 75% of intracerebral haemorrhage with multiple overlapping
worldwide deaths due to haemorrhagic stroke occur.9 sources of case ascertainment.1 We included consecutive
Other tests that can diagnose CAA include CT, which is adult patients (aged ≥16 years) with first-ever intracerebral
usually the first test to diagnose intracerebral haemorrhage, haemorrhage confirmed by CT. We excluded patients
and APOE genotype. Subarachnoid haem­orrhage on CT with recurrent intracerebral haemorrhage; exclusively
occurs with 82% (95% CI 69–93) of cases of intracerebral extra-­axial intracranial haemorrhage; and intracerebral
haemorrhage accompanied by histopathologically proven haemorrhage secondary to trauma, macrovascular
CAA.10 The presence of an APOE ε4 allele is the strongest causes, structural causes, or haemorrhagic trans­form­
genetic association with histo­ pathologically confirmed ation of an ischaemic stroke.1 We collected
sporadic CAA (odds ratio [OR] 2·67, 95% CI 2·31–3·08), demographics, medical history, and drug use at
and the association is dose dependent and occurs diagnosis of intracerebral haemorrhage data by
regardless of dementia comorbidity.11 However, the interviewing patients or their families or carers at the
diagnostic use of CT features and APOE genotype—alone time of presentation and reviewing primary care and
or in combination—is unknown. hospital records.1
We aimed to develop a multivariable prediction The Scotland A Research Ethics Committee
model for identifying lobar intracerebral haemorrhage (10/MRE00/23) approved LINCHPIN. We obtained
associated with CAA using CT and genetic features, written informed consent from all participants or their
internally validate the model, and assess the diagnostic immediate next of kin when participants did not have
accuracy of different cutoffs to rule in and rule out mental capacity.
CAA-associated intracerebral haemorrhage.
Index tests
Methods Two neuroradiologists (MAR and PMW) independently
Study design and participants evaluated reformatted head CT images with a
We did a community-based inception cohort study of standardised pro forma derived from previous large-
spontaneous intracerebral haemorrhage in the scale stroke studies (appendix).13 They assessed See Online for appendix

www.thelancet.com/neurology Vol 17 March 2018 233

 Articles

extra-axial haemorrhage (in the subarachnoid, CAA (0 or 1), and vasculopathy (0–2). Two neuro­
subdural, or intra-ventricular spaces), finger-like pathologists (CS and CH) rated the presence and
projections (elongated extensions arising from the severity of non-CAA (or other small vessel disease)
haematoma, longer than they are wide, regardless of features in the left cerebral hemisphere using
whether they extended to the cortex or not [appendix], haematoxylin and eosin staining with a modified
variably referred to as lobulated or multinodular version of a published scale:16 none (very mild,
appearance in others studies10,14), and other radiographic occasional arteriolosclerosis without media splitting or
features (appendix). All raters did all assessments luminal narrowing), mild (widespread mild or focal
masked to clinical, genetic, and pathological moderate arteriolosclerosis), moderate (widespread
information. We used the initial ratings done by MAR moderate or focal severe arteriolosclerosis, with splitting
for primary analyses. of the media and narrowing of the lumen), or severe
For APOE genotype analysis, we obtained DNA from (widespread severe arteriolosclerosis, fibrinoid necrosis,
peripheral blood samples or cerebellar tissue stored in the lipohyalinosis, evidence of vascular occlusion with or
LINCHPIN brain bank with standard methods described without recanalisation; appendix). We made the CAA
in the appendix. Investigators were masked to clinical, CT, ratings similar to the other small vessel disease ratings
and pathological features during DNA extraction and by restricting them to the parenchymal CAA scores in
genotyping (appendix). We classified APOE genotype as the left cerebral hemisphere, which we summed and
APOE ε2 possession if participants had at least one graded for a CAA burden category (0=none, 1–4=mild,
ε2 allele or APOE ε4 possession if they had at least one 5–8=moderate, and 9–12=severe). Macroscopic neuro­
ε4 allele. pathological assessment could not be masked to
intracerebral haemorrhage location, but investigators
Reference test were masked to other CT features, clinical information,
One neuropathologist (CS) assessed post mortem brain and genotype. For microscopic histopathological
tissue according to a standard operating procedure assessment, investigators were masked to intracerebral
(appendix). The maximum interval from death to haemorrhage location when possible (ie, unless the
autopsy was 5 days. The same neuropathologist rated intracerebral haemorrhage was included in one of the
CAA features (appendix) with a consensus rating scale,15 prespecified sampled regions), as well as other CT
which rates the presence and severity of parenchymal features, clinical information, and genotype.
and meningeal CAA (0–3); the presence of capillary
Statistical analysis
We were unable to calculate the sample size required for
50 Neither moderate or severe CAA nor other small vessel disease
Moderate or severe other small vessel disease alone
a diagnostic test accuracy study because a systematic
Moderate or severe CAA and other small vessel disease together review10 identified only two retrospective cross-sectional
Moderate or severe CAA alone studies that compared CT features of intracerebral
40 haemorrhage associated with histopathologically proven
CAA with intracerebral haemorrhage without CAA, but
these two studies were at high risk of selection and
n=10 information biases. Therefore, after completing
Number of participants

30 recruitment to this prospective population-based study,1

we used the largest sample size possible (n=110) from
n=2 its nested brain bank and we restricted models to three
n=42 predictors (nine outcomes per variable) to reduce
20
overfitting.17 We did a post-hoc power calculation using a
two-sided Z test for a logistic regression model with
n=26 a 5% level of significance, with subarachnoid haemor­
n=24
10
rhage as the main predictor (OR 10·9) and adjusting for
the effect of other covariates (Nagelkerke R² 0·06); this
calculation showed that our maximum achievable
n=6 sample size of 62 participants with lobar intracerebral
0 haemorrhage would result in 70·6% power.
Absent or Moderate or Absent or Moderate or No data were missing from this study. We did statistical
mild CAA severe CAA mild CAA severe CAA
analyses using the R statistic package (version 3.3.2),
Non-lobar intracerebral Lobar intracerebral with the exception of post-hoc power calculation and
haemorrhage (n=48) haemorrhage (n=62) diagnostic accuracy statistics (sensitivity, specificity,
Figure 1: Pathological severity of CAA and other small vessel disease likelihood ratios, and predictive values and their 95% CIs)
For more on VassarStats Clinical
Calculator 1 see http:// according to intracerebral haemorrhage location which we obtained using G*Power 3.1.9.2 and VassarStats
vassarstats.net/clin1.html CAA=cerebral amyloid angiopathy. Clinical Calculator 1, respectively.

234 www.thelancet.com/neurology Vol 17 March 2018

 Articles

Prediction modelling and performance

The multivariable prediction model aimed to use CT
features and APOE genotype to predict lobar intracerebral Absent or Moderate or Odds ratio (95% CI) p value
haemorrhage that was associated with CAA, defined as a mild CAA severe CAA
(n=26) (n=36)
CAA burden category of moderate or severe, with or
without other small vessel disease. Age (years) 84 (78–88) 82 (79–85) NC 0·41
We assessed intra-rater and inter-rater agreement of CT Sex
features using unweighted Cohen’s κ coefficient for Men 12 (46%) 11 (31%) 0·51 (0·18–1·46) 0·27
categorical data, linear-weighted Cohen’s κ coefficient for Women 14 (54%) 25 (69%) 1·95 (0·68–5·55) 0·21
ordinal data, and intra-class correlation coefficient for Hypertension 19 (73%) 23 (64%) 0·65 (0·22–1·96) 0·45
continuous data. We excluded radiographic features with Antiplatelet use at intracerebral 15 (58%) 18 (50%) 0·73 (0·27–2·03) 0·55
haemorrhage
κ less than 0·5 from further analyses. We compared the
Anticoagulant use at 4 (15%) 5 (14%) 0·89 (0·21–3·68) 1·00
distributions of clinical, genetic, and CT characteristics in
intracerebral haemorrhage
cases of lobar intracerebral haemorrhage with or without
Dementia 2 (8%) 8 (22%) 3·43 (0·66–17·72) 0·17
moderate or severe CAA using χ² test (or Fisher’s exact
APOE ε2 possession 3 (12%) 11 (31%) 3·37(0·83–13·63) 0·077
test, where appropriate) for categorical variables and the
APOE ε4 possession 2 (8%) 18 (50%) 12·00 (2·46–58·47) 0·0004
Mann-Whitney U test for non-normally distributed
Multiple intracerebral 6 (23%) 3 (8%) 0·30 (0·07–1·35) 0·15
continuous variables. For the full model we prespecified haemorrhage
subarachnoid haemorrhage and APOE ε4 possession on Left side 14 (54%) 18 (50%) 0·86 (0·31–2·35) 0·76
the basis of systematic review data.10,11 We included finger- Intracerebral haemorrhage location
like projections from the intracerebral haemorrhage on
Frontal 10 (38%) 19 (53%) 1·79(0·64–4·99) 0·27
the basis of the strong univariate association that
Parietal 6 (23%) 8 (22%) 0·95(0·29–3·17) 0·94
we found with CAA-associated lobar intracerebral
Temporal 5 (19%) 5 (14%) 0·68(0·17–2·63) 0·57
haemorrhage. We used Firth’s penalised likelihood logistic
Occipital 5 (19%) 4 (11%) 0·53(0·13–2·18) 0·38
regression to assess the association of predictors with
Intracerebral haemorrhage 59 (23–126) 66 (22–117) NC 0·72
CAA-associated lobar intracerebral haemorrhage and volume (mL)
calculate OR with 95% CIs, because finger-like projections Strictly lobar intracerebral 22 (85%) 36 (100%) NA 0·027
showed complete separation between the two outcome haemorrhage
groups.18 We assessed overall performance using Intraventricular extension 14 (54%) 17 (47%) 0·77 (0·28–2·11) 0·61
Nagelkerke’s R², the Brier score, and the Akaike Any subarachnoid haemorrhage 11 (42%) 32 (89%) 10·91 (2·98–39·96) <0·0001
information criterion.19 We evaluated model dis­crimination Subdural extension 5 (19%) 7 (19%) 1·01 (0·28–3·64) 0·98
with the concordance (c) statistic and discrimination slope, Midline shift 18 (69%) 21 (58%) 0·62 (0·21–1·80) 0·38
and displayed the discrimination graphically using a Finger-like projections 0 14 (39%) 34·16 (1·93–605·23) 0·0003
receiver operating characteristic plot. We assessed model Cortical involvement 21 (81%) 35 (97%) 8·33 (0·91–76·28) 0·074
calibration with the Hosmer–Lemeshow goodness-of-fit Dilute or seeping 9 (35%) 15 (42%) 1·35 (0·47–3·84) 0·59
test and calibration plots. We used bootstrapping to Old vascular lesion 8 (31%) 15 (42%) 1·61 (0·55–4·66) 0·38
evaluate the internal validity of the model performance Anterior WML ·· ·· ·· 0·26
measures.20 We used 2000 random bootstrap samples with 0 2 (8%) 8 (22%) ·· ··
replacement from the full sample of participants with 1 16 (62%) 21 (58%) ·· ··
lobar intracerebral haemorrhage, constructed models on 2 8 (31%) 7 (19%) ·· ··
these bootstrap samples, and derived optimism-adjusted
Posterior WML ·· ·· ·· 0·65
performance measures to provide a realistic estimate of
0 7 (27%) 6 (17%) ·· ··
future performance.19
1 3 (12%) 6 (17%) ·· ··
2 16 (62%) 24 (67%) ·· ··
Development of diagnostic criteria
Central atrophy ·· ·· ·· 0·26
We defined three risk categories for CAA-associated lobar
0 9 (35%) 10 (28%) ·· ··
intracerebral haemorrhage according to the probability
1 17 (65%) 22 (61%) ·· ··
of CAA-associated intracerebral haemorrhage predicted
2 0 4 (11%) ·· ··
by the model: low (≤7%), medium (44–64%), and
Cortical atrophy ·· ·· ·· 0·37
high (≥95%). We used decision curve analysis, which
0 4 (15%) 11 (31%) ·· ··
assesses the use of different risk category cutoffs across
1 15 (58%) 18 (50%) ·· ··
the full range of threshold probabilities and false positive
and false negative weighting, to confirm the optimum 2 7 (27%) 7 (19%) ·· ··

risk category cutoffs for ruling CAA-associated intra­ Data are n (%) or median (IQR). CAA=cerebral amyloid angiopathy. NC=not calculable because the data are continuous.
cerebral haemorrhage either in or out.21 We evaluated the NA=not available as one or more cells contained a zero. WML=white matter lesion.
sensitivity, specificity, positive and negative likelihood
Table 1: Characteristics of lobar intracerebral haemorrhage associated with severity of CAA
ratios, and predictive values and their 95% CIs for the

www.thelancet.com/neurology Vol 17 March 2018 235

 Articles

diagnostic criteria that distinguished low versus medium Data sharing

or high risk categories, and high versus medium or low Clinical, radiographic, genetic, and pathological data
For the data used in this study risk categories. used in this study are available online, along with the
see http://dx.doi.org/10.7488/ code for logistic regression and internal validation.
ds/2230
β coefficient (SE) Odds ratio (95% CI) p value
Role of the funding source
Intercept –2·55 (0·89) ·· 0·0040 The funders of the study had no role in study design,
APOE ε4 carrier 3·11 (1·01) 22 (4–862) 0·0020 data collection, data analysis, data interpretation, or
Subarachnoid 2·31 (0·94) 10 (2–299) 0·014 writing of the report. The corresponding author had full
haemorrhage
access to all the data in the study and had final
Finger-like 3·20 (1·58) 27 (3–not reached) 0·043
projections
responsibility for the decision to submit for publication.

Table 2: Multivariable Firth’s logistic regression prediction model for Results

lobar intracerebral haemorrhage associated with moderate or severe
Between June 1, 2010, and Feb 10, 2016, 110 participants
cerebral amyloid angiopathy
underwent research autopsy after non-contrast head CT
that initially diagnosed first-ever intracerebral
A haemorrhage, after unavoidable exclusions (appendix).12
1·0 The median age was 83 years (IQR 76–87) and
49 (45%) were men (appendix). DNA for APOE genotyping
was available for all 110 participants (peripheral blood
0·8
sample for 28 participants and post-mortem cerebellar
tissue for 82 participants). The median time from
True-positive rate

0·6 intracerebral haemorrhage to CT was 5 h (IQR 3–18) and

the median time from CT to autopsy was 11 days (5–80).
For most radiographic features, intra-rater agreement was
0·4
substantial to almost perfect and inter-rater agreement
was moderate to almost perfect (appendix).
0·2 62 (56%) participants had lobar intracerebral
haemorrhage, 41 (37%) had deep intracerebral haemor­
AUC=0·92 (95% CI 0·86−0·98) rhage, and seven (6%) had infratentorial intracerebral
0
0 0·2 0·4 0·6 0·8 1·0
haemorrhage (appendix). All 48 participants with non-
False-positive rate lobar intracerebral haemorrhage had moderate or severe
B other small vessel disease: most (n=42 [88%]) had other
1·0 small vessel disease with absent or mild CAA and
six (13%) also had moderate or severe CAA, consistent
0·8 with the population prevalence of CAA in octogenarians
(figure 1, appendix).2 Therefore, we focused our further
Observed frequency

analyses on the prediction of moderate or severe CAA in

0·6
participants with lobar intracerebral haemorrhage.
Of 62 participants with lobar intracerebral haemorrhage,
0·4
26 (42%) had moderate or severe other small vessel
disease as well as moderate or severe CAA, 24 (39%) had
0·2 moderate or severe other small vessel disease alone,
ten (16%) had moderate or severe CAA alone, and
0 two (3%) had no clear underlying cause of intracerebral
0 0·2 0·4 0·6 0·8 1·0 haemorrhage (one participant had mild CAA and mild
Predicted probability other small vessel disease, and the second participant had
Figure 2: Discrimination and calibration measures of prediction model mild other small vessel disease and absent CAA, but
performance neither had a macrovascular abnormality, coagulopathy,
(A) Receiver operating characteristic curve for predicted probability of or tumour; figure 1). In univariable analyses, participants
moderate or severe CAA. The AUC is equivalent to the c statistic. The shaded
area represents the 95% CI of the AUC based on 2000 bootstrap replicates. The
with lobar intracerebral haemorrhage and moderate or
dotted line indicates a non-informative AUC of 0·50 for comparison. (B) severe CAA were significantly more likely to be
Calibration plot of predicted probability versus observed frequency of moderate APOE ε4 carriers, and to have a strictly lobar intracerebral
or severe CAA. Grey line indicates perfect calibration, the model’s calibration is haemorrhage, subarachnoid haemorrhage, and finger-like
shown by the dotted line. Triangles represent the six different moderate or
severe CAA risk groups produced by the prediction model. Vertical lines
projections from the intracerebral haemorrhage than
represent the frequency and distribution of model predicted probabilities. participants with lobar intracerebral haemorrhage and
CAA=cerebral amyloid angiopathy. AUC=area under the curve. absent or mild CAA (table 1).

236 www.thelancet.com/neurology Vol 17 March 2018

 Articles

The multivariable prediction model for CAA-associated CAA into low, medium, or high risk (appendix). When
lobar intracerebral haemorrhage included three pre­ no predictors were present (low risk), the predicted
dictors: APOE ε4 possession, subarachnoid haemorrhage, probability of moderate or severe CAA was 7%.
and finger-like projections (table 2). The model calculates Subarachnoid haemorrhage or APOE ε4 possession in
the predicted probability of moderate or severe CAA as isolation (medium risk) predicted 44–64% probability of
follows (the predictor values are 1 when present and 0 moderate or severe CAA. The presence of subarachnoid
when absent): haemorrhage and at least one other predictor (high risk)
predicted a probability of moderate or severe CAA of
1 95% or more.
Predicted probability=
1 + exp-risk score Guided by how the predictors stratified the probability
of moderate or severe CAA associated with lobar
Risk score=– 2⋅55 + 3⋅11 × (APOE ε4 positive) + intracerebral haemorrhage, we identified two sets of
2⋅31 × (subarachnoid haemorrhage) + 3⋅20 × diagnostic criteria including the three predictors
(finger-like projections) (appendix, figure 3). Subarachnoid haemorrhage or
APOE ε4 possession separated low from medium or high
All three predictors were independently associated with probability groups (appendix), and had a sensitivity of
moderate or severe CAA. The variance inflation factor 100% (95% CI 88–100; appendix) meaning that the
values (APOE ε4 carrier 1·33, subarachnoid haemorrhage absence of these two predictors with lobar intracerebral
1·34, and finger-like projections 1·03) confirmed no haemorrhage ruled out moderate or severe CAA.
evidence of multi­collinearity between predictors. We did Subarachnoid haemorrhage and APOE ε4 possession or
a sensitivity analysis excluding participants taking oral finger-like projections, separated high from low or
anticoagulants (n=9), given the high frequency of medium probability groups (appendix), and had a
finger-like projections in such cases,22 which showed specificity of 96% (95% CI 78–100; appendix) meaning
similar significance, direction, and magnitude of the that the presence of these predictors with lobar
independent associations (appendix). intracerebral haemorrhage ruled in moderate or severe
The model showed excellent discrimination CAA. These optimum risk category cutoffs for ruling
(c statistic 0·92, 95% CI 0·86–0·98) with no evidence of CAA-associated intracerebral haemorrhage either in or
poor calibration (Hosmer–Lemeshow goodness of fit test out were confirmed by decision curve analysis (appendix).
p=0·685; figure 2, appendix). Internal validation We repeated the prediction modelling and derived
identified small differences in the overall performance diagnostic criteria using CT ratings from a second
measures (eg, the Brier score increased from independent investigator (PMW). The prediction model
0·11 to 0·12 and the c statistic decreased from 0·92 to 0·91), shows similar direction and magnitude of the predictors
with no evidence of poor calibration (appendix). with moderate or severe CAA, while the sensitivity of
We used the multivariable model to select two cutoff the rule-out criteria remained 100% and specificity of
points to stratify the probability of moderate or severe the rule-in criteria increased to 100% (appendix).

Probability of moderate or severe CAA

Low Medium High

Subarachnoid
– + – + + +
haemorrhage
Predictors

APOE ε4
– – + + – +
possession
Finger-like
– – – – + +
projections

Example of
brain CT

Diagnostic Rule out sensitivity Rule in specificity

test accuracy 100% (95% CI 88–100) 96% (95% CI 78–100)

Figure 3: Categorisation of probability of lobar intracerebral haemorrhage associated with moderate or severe cerebral amyloid angiopathy according to the
three predictor variables, with example CT images
CAA=cerebral amyloid angiopathy. Adapted from Salman and Rodrigues (Creative Commons 4.0).23

www.thelancet.com/neurology Vol 17 March 2018 237

 Articles

We did a comparison of our CT and genetic criteria peripheral blood samples is not a universally available
with the modified Boston criteria8 in the seven test, the inclusion of APOE ε4 possession significantly
participants with lobar intracerebral haemorrhage who improves the prediction model. Given the drive towards
subsequently had MRI done within 6 months of the stratified medicine, these techniques are becoming
intracerebral haemorrhage, and found that all cases at increasingly cost-effective.26 Our criteria could be used in
high or intermediate probability of having moderate or other patient groups to help guide the use of more
severe CAA by our Edinburgh criteria were classified as restricted clinical resources by identifying patients who
probable CAA by the Boston criteria (appendix). might benefit from advanced imaging, such as MRI or
Since APOE genotyping might not be available PET to assess for further features of CAA and other small
worldwide, we assessed the diagnostic test accuracy of a vessel disease biomarkers.
simplified model on the basis of CT features alone This study minimised selection bias by using prospective
(appendix). Subarachnoid haemorrhage alone had a case ascertainment in one community and inviting all
sensitivity of 89% (95% CI 73–96) and the combination potentially eligible people to consent to the nested brain
of subarachnoid haemorrhage and finger-like projections bank. We minimised information bias by standardising
had 100% (95% CI 84–100) specificity (appendix). imaging format; defining and systematically assessing
However, the inclusion of APOE ε4 improved the model radiographic features; standard­ ising brain tissue
(full model χ²=59·0, Akaike information criterion=49·9 acquisition; extensively sampling brain tissue, rather than
vs simpli­­ fied CT-based model χ²=41·9, Akaike using cortical biopsy; systematically assessing pathological
information criterion=65·0, p<0·0001). features; using validated rating scales;15,16,27 and masking
assessors. No data were missing and inter-rater agreement
Discussion was moderate to substantial for the key predictors. We
In our study, we used a systematically acquired brain chose logistic regression rather than machine-learning
tissue bank nested within a prospective, population-based approaches, such as neural networks, for our prediction
cohort study to develop and internally validate a simple, model, because of its simplicity, familiarity, and
three-variable model using two CT features (subarachnoid transparency. We reduced overfitting by restricting the
haemorrhage and finger-like projections from intra­ multivariable model to three predictors, prespecified two
cerebral haemorrhage) and APOE genotype to predict of these predictors, and avoided stepwise methods of
moderate or severe CAA associated with lobar predictor selection due to the probable instability of
intracerebral haemorrhage. The model had excellent selection related to the sample size. We did internal
discrimination and calibration. The diagnostic criteria validation with bootstrapping to further reduce optimism
might inform estimates of prognosis and decisions about in study performance measures.
antithrombotic drugs after intracerebral haemorrhage. This study has some limitations. Sample size was
We identified two clinically useful diagnostic cutoffs that modest and the post-hoc power calculation shows that the
have implications for clinical practice. Neither study has 71% power and therefore is at risk of a type II
subarachnoid haemorrhage nor APOE ε4 possession was error. Our sample size left us unable to develop criteria to
100% sensitive for moderate or severe CAA with a negative distinguish lobar intracerebral haemorrhage associated
likelihood ratio of 0; a negative likelihood of less than with CAA alone, other small vessel disease alone, and
0·1 means a negative test is good at ruling out a diagnosis.24 mixed CAA and other small vessel disease, which could
Therefore, the absence of these features can rule out be clinically important. However, this study has the
CAA-associated lobar intracerebral haemorrhage, which largest sample we could achieve over 6 years in a
might identify people with a lower risk of recurrent population of about 850  000 people, with roughly
intracerebral haemorrhage,5,25 dementia,6 and susceptibility 50% consenting soon after an acute brain injury,12 which
to the effects of anti­thrombotic drugs.7 The presence of is comparable with other brain banks.28
subarachnoid haemorrhage and APOE ε4 possession or Although we tried to limit selection bias, participants
finger-like projections was 96% specific with a positive who underwent autopsy were older, had larger intra­
likelihood ratio 16·6. A positive likelihood of more than cerebral haemorrhage, more frequent intra­ ventricular
10 means a positive test is good at ruling in a diagnosis,24 so extension, and more severe posterior white matter
the presence of these features can effectively rule in CAA- lucencies, and generally died soon after their intracerebral
associated lobar intracerebral haemorrhage to identify haemorrhage compared with participants who did not
patients for studies of CAA treatment. (appendix); these differences were inevitable in standard
Our diagnostic criteria are based on features identified clinical practice, and indicate to whom our results are
on non-contrast CT, a widely available, relatively generalisable. Whilst the frequency of model predictors
inexpensive diagnostic test with few contraindications, (APOE ε4 possession, subarachnoid haemorrhage and
which is suitable in acutely unwell patients. The intra- finger-like projections) and the distribution of risk
rater and inter-rater agreement for subarachnoid categories did not vary between those participants who
haemorrhage and finger-like projections were moderate underwent autopsy and those participants who did not
to almost perfect. Although APOE genotyping using undergo autopsy (appendix), the applicability of our

238 www.thelancet.com/neurology Vol 17 March 2018

 Articles

criteria to other intracerebral haemorrhage groups—such content. JMW designed the study, collected and interpreted data, and
as younger patients with smaller intracerebral critically revised the manuscript for important intellectual content. CS
collected and interpreted data, and critically revised the manuscript for
haemorrhage—is unclear. important intellectual content. RA-SS conceived the study design,
We only identified finger-like projections in cases of interpreted the data, and critically revised the manuscript for important
lobar intracerebral haemorrhage with moderate or severe intellectual content. All authors gave final approval of the manuscript.
CAA, making them very specific for CAA-associated Declaration of interests
haemorrhage. However, finger-like projections are CH reports grants from Alzheimer’s Society, during the conduct of the
difficult to define, subjective, and potentially prone to study; and grants from Alzheimer’s Research UK, outside the
submitted work. PMW reports grants from Medical Research Council,
observer variability. We used both written and pictorial during the conduct of the study; and sits on the Stryker Neurovascular
definitions to improve reliability and showed substantial Clinical Advisory Board whose remit includes intracerebral
intra-rater and moderate inter-rater agreement. haemorrhage. CC has participated on national boards for Bayer,
Furthermore, we repeated the model­ling using ratings Medtronics, and Daichii-Sankyo; has been a clinical investigator in
trials supported by AstraZeneca, Daichii-Sankyo, and
from a second independent investigator, which resulted Boehringer Ingelheim; and was national coordinator and investigator
in rule-out criteria with 100% sensitivity and rule-in for the study A9951024 supported by Pfizer, outside the submitted
criteria with 100% specificity for CAA-associated lobar work. JMW reports grants from European Union Horizon 2020,
intracerebral haemorrhage. Previous studies10 have Fondation Leducq, Medical Research Council, Engineering and
Physical Sciences Research Council, British Heart Foundation, The
reported that irregular and lobulated intracerebral Stroke Association, The Alzheimer’s Society, The Chief Scientist Office,
haemorrhage borders are more frequently associated and the Wellcome Trust, outside the submitted work. All other authors
with CAA-associated lobar intracerebral haemorrhage. declare no competing interests.
However, these features were not defined in the studies, Acknowledgments
making them difficult to relate to our findings. This study was funded by UK Medical Research Council (MRC), The
Stroke Association, and The Wellcome Trust Edinburgh Clinical
Finally, CT features of intracerebral haemorrhage, such
Academic Track PhD Programme. We thank Rosemary Anderson,
as subarachnoid haemorrhage and finger-like projections, Aidan Hutchison, the adults included in the LINCHPIN study, and their
will evolve with time. The timecourse for this evolution in relatives and carers. Some of the imaging was done at the Brain
intracerebral haemorrhage is unknown and how it might Research Imaging Centre, Neuroimaging Sciences, Edinburgh,
University of Edinburgh, which is part of the Scottish Imaging
affect the diagnostic accuracy of these criteria is unclear. Network–A Platform for Scientific Excellence collaboration funded by
We would expect a reduction in sensitivity with time, the Scottish Funding Council and the Chief Scientist Office. Support
however our rule-out criteria remain 100% sensitive from National Health Service Lothian R&D and the Edinburgh Clinical
despite including participants who had a CT scan up to Research Facility is gratefully acknowledged. We thank the Edinburgh
Brain and Tissue Bank, part of the MRC Brain Banks Network, for
7 days after symptom onset. curating the brain tissue from donors for this study (appendix). CC is a
We are planning an external validation study to assess member of the Institut Universitaire de France.
the performance of the prediction model and diagnostic References
criteria in other settings, countries, ethnic groups, and 1 Samarasekera N, Fonville A, Lerpiniere C, et al. Influence of
patient populations, such as those patients with intracerebral hemorrhage location on incidence, characteristics,
and outcome: population-based study. Stroke 2015; 46: 361–68.
smaller intracerebral haemorrhage, and determine the 2 Charidimou A, Gang Q, Werring DJ. Sporadic cerebral amyloid
reproducibility of predictor assessment by other angiopathy revisited: recent insights into pathophysiology and
investigators (Greenberg SM, Massachusetts General clinical spectrum. J Neurol Neurosurg Psychiatry 2012; 83: 124–37.
Hospital, Boston, MA, USA, personal communication). 3 Pantoni L. Cerebral small vessel disease: from pathogenesis and
clinical characteristics to therapeutic challenges. Lancet Neurol 2010;
Further large, representative samples with a systematic 9: 689–701.
reference standard based on autopsy could attempt to 4 Samarasekera N, Smith C, Al-Shahi Salman R. The association
distinguish lobar intracerebral haemorrhage associated between cerebral amyloid angiopathy and intracerebral
haemorrhage: systematic review and meta-analysis.
with CAA alone, other small vessel disease alone, and J Neurol Neurosurg Psychiatry 2012; 83: 275–81.
mixed CAA and other small vessel disease. Comparison of 5 Charidimou A, Imaizumi T, Moulin S, et al. Brain hemorrhage
these CT and genetic criteria and the widely used modified recurrence, small vessel disease type, and cerebral microbleeds:
a meta-analysis. Neurology 2017; 89: 820–29.
Boston criteria8 against a pathological reference standard 6 Moulin S, Labreuche J, Bombois S, et al. Dementia risk after
would be noteworthy. Studies of the clinical and economic spontaneous intracerebral haemorrhage: a prospective cohort study.
impact of these criteria on prognosis and therapeutic Lancet Neurol 2016; 15: 820–29.
7 Biffi A, Halpin A, Towfighi A, et al. Aspirin and recurrent
decisions might quantify, and hopefully confirm, their intracerebral hemorrhage in cerebral amyloid angiopathy.
effect on the outcome of this devastating disease. Neurology 2010; 75: 693–98.
Contributors 8 Linn J, Halpin A, Demaerel P, et al. Prevalence of superficial
MAR did the literature search, conceived the study design, collected, siderosis in patients with cerebral amyloid angiopathy. Neurology
2010; 74: 1346–50.
analysed, and interpreted the data, made the figures, drafted the
manuscript, and had final approval of the manuscript. NS did the 9 Feigin VL, Krishnamurthi RV, Parmar P, et al. Update on the Global
Burden of Ischemic and Hemorrhagic Stroke in 1990–2013:
literature search, conceived the study design, collected data, and
the GBD 2013 study. Neuroepidemiology 2015; 45: 161–76.
critically revised the manuscript for important intellectual content. CL,
10 Samarasekera N, Rodrigues MA, Toh PS, Al-Shahi Salman R.
CH, and PMW collected data and critically revised the manuscript for
Imaging features of intracerebral hemorrhage with cerebral
important intellectual content. MOM, JARN, CLMS, and CC designed amyloid angiopathy: systematic review and meta-analysis.
the study and critically revised the manuscript for important intellectual PLoS One 2017; 12: e0180923.

www.thelancet.com/neurology Vol 17 March 2018 239

 Articles

11 Rannikmae K, Samarasekera N, Martinez-Gonzalez NA, 20 Efron B. Estimating the error rate of a prediction rule:
Al-Shahi Salman R, Sudlow CL. Genetics of cerebral amyloid improvement on cross-validation. J Am Stat Assoc 1983; 78: 316–31.
angiopathy: systematic review and meta-analysis. 21 Vickers AJ, Elkin EB. Decision curve analysis: a novel method for
J Neurol Neurosurg Psychiatry 2013; 84: 901–08. evaluating prediction models. Med Decis Making 2006; 26: 565–74.
12 Samarasekera N, Lerpiniere C, Fonville AF, et al. Consent for brain 22 Huttner HB, Steiner T, Hartmann M, et al. Comparison of ABC/2
tissue donation after intracerebral haemorrhage: a community-based estimation technique to computer-assisted planimetric analysis in
study. PLoS One 2015; 10: e0135043. warfarin-related intracerebral parenchymal hemorrhage. Stroke
13 IST-3 collaborative group. Association between brain imaging signs, 2006; 37: 404–08.
early and late outcomes, and response to intravenous alteplase after 23 Al-Shahi Salman R, Rodrigues M. The Edinburgh diagnostic
acute ischaemic stroke in the third International Stroke Trial criteria for lobar intracerebral haemorrhage associated with
(IST-3): secondary analysis of a randomised controlled trial. moderate/severe cerebral amyloid angiopathy, 2010–2016 [image].
Lancet Neurol 2015; 14: 485–96. http://dx.doi.org/10.7488/ds/2262 (accessed Dec 6, 2017).
14 Miller JH, Wardlaw JM, Lammie GA. Intracerebral haemorrhage 24 Grimes DA, Schulz KF. Refining clinical diagnosis with likelihood
and cerebral amyloid angiopathy: CT features with pathological ratios. Lancet 2005; 365: 1500–05.
correlation. Clin Radiol 1999; 54: 422–29. 25 O’Donnell HC, Rosand J, Knudsen KA, et al. Apolipoprotein E
15 Love S, Chalmers K, Ince P, et al. Development, appraisal, genotype and the risk of recurrent lobar intracerebral hemorrhage.
validation and implementation of a consensus protocol for the N Engl J Med 2000; 342: 240–45.
assessment of cerebral amyloid angiopathy in post-mortem brain 26 Zhong L, Xie YZ, Cao TT, et al. A rapid and cost-effective method
tissue. Am J Neurodegener Dis 2014; 3: 19–32. for genotyping apolipoprotein E gene polymorphism.
16 Deramecourt V, Slade JY, Oakley AE, et al. Staging and natural Mol Neurodegener 2016; 11: 2.
history of cerebrovascular pathology in dementia. Neurology 2012; 27 Charidimou A, Schmitt A, Wilson D, et al. The Cerebral
78: 1043–50. Haemorrhage Anatomical RaTing inStrument (CHARTS):
17 Vittinghoff E, McCulloch CE. Relaxing the rule of ten events per development and assessment of reliability. J Neurol Sci 2017;
variable in logistic and Cox regression. Am J Epidemiol 2007; 372: 178–83.
165: 710–18. 28 Samarasekera N, Al-Shahi Salman R, Huitinga I, et al. Brain banking
18 Heinze G, Schemper M. A solution to the problem of separation in for neurological disorders. Lancet Neurol 2013; 12: 1096–105.
logistic regression. Stat Med 2002; 21: 2409–19.
19 Harrell FE. Regression modeling strategies: with applications to
linear models, logistic and ordinal regression, and survival analysis.
Cham: Springer, 2015.

240 www.thelancet.com/neurology Vol 17 March 2018