Optimal Timing of Anticoagulant Treatment After

Intracerebral Hemorrhage in Patients With

Atrial Fibrillation
Johanna Pennlert, MD; Rosanna Overholser, PhD; Kjell Asplund, MD, PhD;
Bo Carlberg, MD, PhD; Bart Van Rompaye, PhD; Per-Gunnar Wiklund, MD, PhD;
Marie Eriksson, PhD

Background and PurposeThis study aims to provide observational data on the relationship between the timing of
antithrombotic treatment and the competing risks of severe thrombotic and hemorrhagic events in a cohort of Swedish
patients with atrial fibrillation and intracerebral hemorrhage (ICH).
MethodsPatients with atrial fibrillation and a first-ever ICH were identified in the Swedish Stroke Register, Riksstroke,
2005 to 2012. Riksstroke was linked with other national registers to find information on treatment, comorbidity, and
outcome. The optimal timing of treatment in patients with low and high thromboembolic risk was described through
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017

cumulative incidence functions separately for thrombotic and hemorrhagic events and for the combined end point vascular
death or nonfatal stroke.
ResultsThe study included 2619 ICH survivors with atrial fibrillation with 5759 person-years of follow-up. Anticoagulant
treatment was associated with a reduced risk of vascular death and nonfatal stroke in high-risk patients with no significantly
increased risk of severe hemorrhage. The benefit seemed to be greatest when treatment was started 7 to 8 weeks after ICH.
For high-risk women, the total risk of vascular death or stroke recurrence within 3 years was 17.0% when anticoagulant
treatment was initiated 8 weeks after ICH and 28.6% without any antithrombotic treatment (95% confidence interval for
difference, 1.4%21.8%). For high-risk men, the corresponding risks were 14.3% versus 23.6% (95% confidence interval
for difference, 0.4%18.2%).
ConclusionsThis nationwide observational study suggests that anticoagulant treatment may be initiated 7 to 8 weeks after
ICH in patients with atrial fibrillation to optimize the benefit from treatment and minimize risk.(Stroke. 2017;48:314-320.
DOI: 10.1161/STROKEAHA.116.014643.)
Key Words: anticoagulants atrial fibrillation cerebral hemorrhage ischemia stroke

B oth the prevalence of atrial fibrillation (AF) and oral anti-

coagulant treatment after intracerebral hemorrhage (ICH)
has increased in recent years.1 Still, a large proportion of
these patients are untreated, reflecting the controversy of the
decision.25
There is emerging evidence of the potential benefits of anti-
coagulants in these patients. In a Swedish population-based
study, recurrent ischemic events outnumbered recurrent ICH
among ICH survivors6 and ICH in itself has been identified
as an independent predictor of thromboembolic events among
patients with AF.7
Two recent, Danish observational studies support the reintro-
duction of oral anticoagulants as being associated with a signifi-
cant reduction of all-cause mortality and ischemic stroke rates.8,9
International guidelines highlight the lack of evidence as
to whether and when to resume anticoagulant treatment after
ICH.10,11 The largest retrospective study examining the optimal
time window for initiating treatment included 3 tertiary cen-
ters and 234 patients with warfarin-associated ICH. Fifty-nine
patients resumed anticoagulant treatment, and the study con-
cluded that resumption should be delayed by 10 to 30 weeks to
avoid the early high-risk period for recurrent hemorrhage.12 In
contrast, a systematic review detailing 492 patients suggests
that anticoagulation in high-risk patients may be restarted
3 days from the time of intracerebral bleedings, but authors
emphasize the limitations inherent in the studies analyzed.13
Using the nationwide Swedish Stroke Register,
Riksstroke, we studied the relationship between the timing

Continuing medical education (CME) credit is available for this article. Go to http://cme.ahajournals.org to take the quiz.
Received July 15, 2016; final revision received October 20, 2016; accepted November 11, 2016.
From the Department of Public Health and Clinical Medicine, Medicine (J.P., K.A., B.C., P.-G.W.) and Department of Statistics, Ume School of
Business and Economics (R.O., B.V.R., M.E.), Ume University, Sweden; and Department of Applied Mathematics, Computer Science and Statistics,
Ghent University, Belgium (R.O., B.V.R.).
Presented in part at the 2nd European Stroke Organisation Conference (ESOC 2016), Barcelona, Spain, May 1012, 2016.
The online-only Data Supplement is available with this article at http://stroke.ahajournals.org/lookup/suppl/doi:10.1161/STROKEAHA.
116.014643/-/DC1.
Correspondence to Johanna Pennlert, MD, Department of Public Health and Clinical Medicine, Medicine, Ume University, SE-90185. E-mail johanna.
pennlert@umu.se
2016 American Heart Association, Inc.
Stroke is available at http://stroke.ahajournals.org DOI: 10.1161/STROKEAHA.116.014643

314
 Pennlert et al Optimal Timing of AC in Patients With AF After ICH 315
of antithrombotic treatment and the competing risks of severe
thrombotic events, hemorrhagic events, and death from other
causes in ICH survivors with AF.
Methods
Study Population
All patients with a first-ever ICH (International Classification of
Diseases-Tenth Revision [ICD-10]: I.61) recorded in Riksstroke
between July 1, 2005, and December 31, 2012, with a concomitant
diagnosis of AF and surviving hospital discharge were included.
Patients with traumatic ICH, subdural hematomas, and subarach-
noidal hemorrhages were not included. The diagnosis of AF was
obtained from Riksstroke or found in the Swedish Inpatient registry
(ICD-10: I48) from 1997 to ICH onset.
Data Sources
Patient records were linked from several Swedish nationwide data-
bases so as to describe patient characteristics at the time of ICH, the
prescription of antithrombotic drugs after ICH, and any subsequent
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017
severe clinical events. Riksstroke was established in 1994, and since
1998 all hospitals admitting patients with acute stroke in Sweden
participate. The register has an estimated coverage of >94% of all
patients with acute stroke.14 The Swedish Dispensed Drug Register
carries nationwide information on all dispensed outpatient drug pre-
scriptions collected from all Swedish pharmacies from July 1, 2005.15
The Swedish Inpatient Register, managed by the National Board of
Health and Welfare, has had complete national coverage since 1987,
and >99% of all somatic hospital discharges are registered. Diagnostic
accuracy, measured by positive predictive value, differs between diag-
noses, but is generally 85% to 95%.16 Information on socioeconomic
variables was retrieved from the Longitudinal Integration Database
for Health Insurance and Labour Market Studies,17 managed by
Statistics Sweden. Linkage with the Cause of Death Register,18 man-
aged by the National Board of Health and Welfare, was made to find
information on direct and contributory causes of death.
Approval for this study was obtained from the Ethical Review
Board, Ume, Sweden (Dnr 2014-76-32M), as an extension of the
EqualStroke project (Dnr 2012-321-31M).
Variable Definitions
Outcome Variables
Two different outcome events were defined. First, thrombotic
events were ischemic stroke events (fatal or nonfatal), and all
causes of death directly or indirectly caused by a thrombotic event
(myocardial infarction or systemic arterial thromboembolism). The
second major outcome was hemorrhagic events, defined as either a
recurrent ICH (fatal or nonfatal) or any bleeding event directly or
indirectly causing death. The 2 outcomes were also combined, sim-
ilar to the primary end point of the APACHE-AF study (Apixaban
Versus Antiplatelet Drugs or No Antithrombotic Drugs After
Anticoagulation-Associated Intracerebral Haemorrhage in Patients
With Atrial Fibrillation; vascular death or nonfatal stroke).19 ICD-
10 codes listed for thrombotic and hemorrhagic events are avail-
able in Table I in the online-only Data Supplement.
Time Variables and Censoring
One of the inherent features of the stroke register is that recurrent
events within the first 28 days of onset of a first-ever stroke are not
recorded. Therefore, the starting point for outcome follow-up was set
at day 28 after index ICH. The starting point for follow-up of anti-
thrombotic treatment was time of first dispensed prescription of anti-
thrombotic treatment after discharge, given that the patient was not
on treatment at discharge. Death from any other cause was modeled
as the third possible outcome event. Censoring events during follow-
up was initiation of dual treatment (anticoagulant+antiplatelet drugs),
reaching study-end-date (December 31, 2012) or patients being lost
to follow-up because of emigration.
Anticoagulant and Antiplatelet Treatment
Riksstroke contains information on anticoagulant and antiplatelet
treatment at hospital discharge. The first registered dispensed pre-
scription, if any, from each of the groups of antithrombotic agents
was derived from the Swedish Dispensed Drug Register. ATC code
B01AC was used for antiplatelet drugs. For oral anticoagulant ther-
apy (vitamin K antagonists and direct oral anticoagulants), the fol-
lowing codes were used: B01AA, B01AE, and B01AF. Once having
had a dispensed prescription of either of the antithrombotic agents,
analyses were performed according to the intention-to-treat principle,
unless a patient changed from antiplatelet treatment to anticoagulant
treatment or vice versa (prescribed dual therapy). Once prescribed
dual therapy, the patient was censored. Information on antithrombotic
treatment at the onset of ICH was obtained from Riksstroke or having
had a dispensed prescription 6 months before the index ICH.
Comorbidity
The diagnosis of AF (ICD-10 I48) was based on registry data in
Riksstroke or a diagnosis of AF before the ICH in the Inpatient
Register. Hypertension was defined as being on antihyperten-
sive treatment in Riksstroke or a diagnosis (ICD-10 I10-15) in the
Inpatient Register. An overview of all comorbid factors is present in
Table II in the online-only Data Supplement.
Patient Risk Profiles
Two types of risk patients were defined with given sets of clinically
important patient characteristics. A low-risk patient was 69 years of
age, spent 14 days in hospital after the index ICH, had no previous
risk factors other than AF, and had no previous antithrombotic treat-
ment. To evaluate the risk of ischemic stroke in patients with AF, a
commonly used risk-stratification score is the CHA2DS2-VASc score
(congestive heart failure, hypertension, age [75 years; 2 points], dia-
betes mellitus, stroke/transient ischemic attack [2 points], vascular
disease, age [6574 years], sex [female]).20 By CHA2DS2-VASc, this
patient profile corresponds to 1 point if male and 2 points if female.
A high-risk patient was 80 years of age and spent 28 days in hos-
pital. The patient had a previous ischemic stroke, hypertension, and
diabetes mellitus and was on previous anticoagulant treatment at the
time of ICH (by CHA2DS2-VASc; 6 points if male and 7 points if
female). Baseline CHA2DS2-VASc scores were estimated using the
same methodology as that of a previous study by Pennlert et al.1
Additional patient profiles are presented in the online-only Data
Supplement.
Statistical Methods
A retrospective power calculation based on the 2619 patients
included, of whom 232 (8.9%) received anticoagulants, showed that
the study would be able to detect a difference in cumulative incidence
of total events of 10% versus 17% between treated and nontreated
patients, with 82% power (2-sided test with 5% significance level).
Baseline characteristics are summarized in the Table. To explore
the relationship of anticoagulant and antiplatelet treatment starting
times with the competing risks of thrombotic events, hemorrhagic
events and other causes of death, we focused our analyses on the
estimation of cumulative incidence functions (CIFs). The CIF is the
probability of observing an event before a specified time. CIFs are
defined for thrombotic and hemorrhagic events separately and when
summed give the CIF of the combined outcome vascular death or
nonfatal stroke.
We built a Cox proportional hazard model21 for each event (Table
III in the online-only Data Supplement). This allowed us to adjust
for differences in patient characteristics when computing the cause-
specific hazards. Each model contained 2 time-varying covariates.
For treatments, we used smoothing splines (a nonlinear function
whose shape is determined by the data) of start time of anticoagulant
or antiplatelet treatment during the treatment periods. To reduce the
possibility of overfitting, a linear behavior was used for time periods
with few or no data points (after 38 and 69 weeks, respectively). For
 316StrokeFebruary 2017

Table. Baseline Characteristics of Patients With Atrial The available covariates were baseline characteristics (Table) and a
Fibrillation Who Survived a First Intracerebral Hemorrhage in smoothing spline (with a linear behavior after 26 weeks) of time
Sweden 2005 to 2012 Included in Analysis (n=2619) since discharge. Patients with missing values of level of conscious-
ness were treated as a separate category. Other covariates included
n (%), Mean (SD) for Age <2% missing. After covariates were selected for each of the 3 mod-
els using the same set of patients with information on all covariates
Age, y 78.0 (9.1) (n=2562), the 3 models were refit using the set of patients with infor-
Female sex 1065 (40.7) mation on all covariates selected for at least 1 model (n=2619).The
selection algorithm started with the complete model and eliminated
Level of education the variable that gave the largest decrease in Akaike Information
Primary school 1309 (50.0) Criterion. Selection ended when the removal of any variable resulted
in an Akaike Information Criterion of 2 more than the current Akaike
Secondary school 900 (34.4) Information Criterion.
For a given set of patient characteristics (ie, the high- and low-
University 360 (13.7)
risk patients), the cause-specific hazards were combined to compute
Missing 50 (1.9) the thrombotic event, hemorrhagic event, and the combined CIF.22
Thus, the effect of the treatment starting times could be assessed for
Living alone each event separately and for the combined events. The main results
Yes 1266 (48.3) were stratified by sex and patient risk status. SEs for the CIFs were
computed via a parametric bootstrap.23 Further details are in the
ADL dependency before first ICH online-only Data Supplement. After computing CIFs with and with-
out treatment over a range of start times for a given patient profile, we
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017

Yes 301 (11.5)

identified intervals of starting times where the CIFs were significantly
Level of consciousness on hospital admission different. The optimal time was then chosen as the time of lowest CIF
Alert 1999 (76.3)
with treatment.
In a confirmatory analysis, empirical CIFs were computed for the
Drowsy 495 (18.9) total study population and according to treatment status at the esti-
mated optimal time point (8 weeks after stroke). For simple group
Unconscious 80 (3.1)
comparisons (patients on anticoagulant and antiplatelet treatment
Missing 45 (1.7) versus no treatment at 8 weeks), P values were estimated using the 2
test for categorical variables and t test for age.
Medical history Statistical analysis was performed using R.24
Diabetes mellitus 605 (23.1)
Previous ischemic stroke 640 (24.4) Results
There were 2777 patients in Riksstroke having survived hos-
Venous thromboembolism 140 (5.3)
pital discharge after a first-ever ICH with concomitant AF.
Ischemic heart disease 713 (27.2) We excluded 1 patient because of an obvious recording error.
Hypertension 2180 (83.2) Patients with event times in the 28 days immediately after the
Heart failure 672 (25.7) time of ICH onset (n=103) and 11 patients who were on both
anticoagulant and antiplatelet treatment at discharge (n=11)
Valvular disease 213 (8.1)
were removed. The final study population consisted of 2662
Peripheral arterial disease 156 (6.0) patients, 1568 men and 1094 women, with a mean age of 78
Dementia 153 (5.8) years (Figure1). These 2662 patients were used in the model-
building process. One or more baseline characteristics used
Treatment at ICH onset
in the final model of CIFs were missing for 43 patients, and
AC 1239 (47.3) hence the analysis of CIF included 2619 patients. Baseline
AP 1175 (44.9) characteristics of the ICH survivors included in the final model
Both AC and AP 205 (7.8)
are presented in the Table. Patient characteristics according to
treatment status at 8 weeks after ICH are presented in Table IV
Statins 723 (27.6) in the online-only Data Supplement.
CHA2DS2-VASc The patients constituted 5759 person-years of follow-up
Median (mean) score 4 (4.13) from stroke onset to patients were either censored or expe-
rienced a new event (median follow-up was 1.7 years). Total
Score 04 1595 (60.9)
follow-up time from treatment initiation was 581 person-years
Score 59 1024 (39.1) for anticoagulants, and 3,001 person-years for antiplatelets. Of
AC indicates anticoagulant treatment; ADL, activities of daily living; AP, antiplatelet the 232 patients initiating anticoagulant treatment, 59.5% had
treatment; CHA2DS2-VASc, congestive heart failure, hypertension, age (75 years; a dispensed prescription within the first 3 months after onset
2 points), diabetes mellitus, stroke/transient ischemic attack (2 points), vascular of ICH. Among the 1136 patients who received antiplatelet
disease, age (6574 years), sex (female); and ICH, intracerebral hemorrhage.
therapy, 58.9% claimed a prescription within 3 months.

age, we used a piecewise linear effect with turning points at the quar-
tiles of the observed ages. Other covariates were selected accord-
ing to a backward selection algorithm. Because a covariate related
to one of the outcomes was not necessarily related to another, we
performed model selection for each of the 3 outcomes separately.
Outcome
During follow-up, we observed 379 severe thrombotic events
of which 302 (79.7%) were ischemic strokes. Of 115 severe
hemorrhagic events, 96 (83.5%) were recurrent ICH events.
 Pennlert et al Optimal Timing of AC in Patients With AF After ICH 317
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017
The 28-day case fatality after ischemic stroke was 17.5%
compared with 37.5% after recurrent hemorrhagic stroke
(P<0.001, 2 test). At 3 years, the cumulative incidence of
thrombotic events was 14.5%, and the incidence of severe
hemorrhagic events was 4.4% (Figure2).
Optimal Timing of Treatment
Figures3A (women) and Figure3B (men) show the cumula-
tive incidences (CIFs), adjusted for differences in patient char-
acteristics, of thrombotic, hemorrhagic, and the sum of the 2
events (vascular death and stroke) at 3 years after onset of
ICH in relation to start time of anticoagulant and antiplatelet
treatment. The thick lines represent time periods during which
treatment initiation of anticoagulants (black) and antiplatelet
Figure 2. Empirical cumulative incidence functions (CIF) for the
3 outcome events (thrombotic events, hemorrhagic events, and
death from nonvascular causes). ICH indicates intracerebral
hemorrhage.
Figure 1. Study flow chart. AF indicates atrial
fibrillation; ICH, intracerebral hemorrhage; and
IS, ischemic stroke. *74 switched from antiplate-
lets (AP) to anticoagulants (AC), 22 from AC to
AP, and 2 received AP and AC the same day.
therapy (gray) are significantly different from no treatment.
The risk reduction of thrombotic events in patients treated
with anticoagulants compared with no treatment was statisti-
cally significant in the 4- to 16-week interval for both the low-
and high-risk patients (Figure3, top panel, black thick line).
Initiation of anticoagulants was not associated with any sig-
nificantly increased risk of a hemorrhagic event. However, we
cannot rule out that the initiation of anticoagulation very early
may increase the risk of bleeding, compared with no treatment
(Figure3, mid panel).
In the presented patient profiles (women and men, patients
at low risk and at high risk), there was an early U-shaped
relationship between timing of initiating anticoagulant treat-
ment and the combined end point of vascular death or stroke
(Figure3, bottom panel). The lowest estimated CIFs of vas-
cular death or nonfatal stroke were found when anticoagulant
treatment was started in the 7- to 8-week interval. For high-
risk women, the total risk of vascular death or stroke recur-
rence within 3 years was 17.0% when anticoagulant treatment
was initiated 8 weeks after ICH, when compared with 28.6%
without any antithrombotic treatment (95% confidence inter-
val [CI] for difference, 1.4%21.8%). The corresponding risks
were 14.3% versus 23.6% (95% CI for difference, 0.4%
18.2%) for high-risk men, 8.2% versus 12.6% (95% CI for
difference, 2.1% to 10.8%) for low-risk women, and 7.3%
versus 10.7% (95% CI for difference, 2.7% to 9.4%) for low-
risk men (Figure3).
Changing the covariates, instead investigating patients
with previous anticoagulant treatment and hypertension
and diabetes mellitus, respectively (Figure IV in the online-
only Data Supplement) did not change the association of
a positive effect of anticoagulants on risk of thrombosis,
without an excess risk of hemorrhage. Furthermore, chang-
ing the patient profiles had only a minor effect on the opti-
mal treatment initiation time point although the magnitude
of the effect of treatment varied (description of the sensitiv-
ity analysis is available in the Methods in the online-only
Data Supplement).
When compared with no antithrombotic therapy, antiplate-
let treatment was not associated with a lowered event risk at
 318StrokeFebruary 2017

A
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017

Figure 3. Cumulative incidence functions (CIFs), adjusted for differences in patient characteristics, at three years after onset of intracere-
bral hemorrhage (ICH) vs start time of 2 treatments (anticoagulant [AC]=black line; antiplatelet [AP]=gray line) and no treatment (dashed
line) for both a low-risk and a high-risk female/male profile (A=female profiles, B=male profiles). The event-specific CIFs (thrombotic and
hemorrhagic) sum to the vascular death or stroke CIFs in the bottom row. The thick lines represent time periods during which treatment
initiation of AC (black) and AP (gray) are significantly different from no treatment at the 5% level.

any time of initiating treatment and was associated with an Discussion

increased risk most of the times (Figure3).
The unadjusted cumulative incidences of a thrombotic
event 3 years after stroke was 6.3% in patients who initiated
anticoagulant treatment within 8 weeks after ICH, 18.8% in
patients who initiated antiplatelets within 8 weeks, and 13.8%
in patients with neither anticoagulants nor antiplatelets within
8 weeks. The corresponding incidences were 6.9%, 3.9%, and
4.4% for hemorrhagic events (Figure V in the online-only
Data Supplement). Patients who initiated anticoagulant treat-
ment within 8 weeks after ICH had a reduced rate of throm-
botic events (95% CI for difference, 13.9% to 1.0%), with
no significantly increased rate of hemorrhagic events (95%
CI for difference, 3.7% to 8.7%) as compared with patients
without any antithrombotic treatment within 8 weeks.
The first main finding in our study is that anticoagulant treat-
ment is associated with a significant reduction in 3-year
thrombotic event risk and not associated with a significant
increase in hemorrhagic event risk in patients with ICH and
AF. This is true for men and women with high and low event
risk profiles. In high-risk patients, anticoagulant treatment is
also associated with a reduction of the combined event risk of
vascular death and nonfatal stroke. The second main observa-
tion, of evident clinical significance, is that the optimal time
to start anticoagulant treatment in patients with AF who have
had an ICH seems to be at around 7 to 8 weeks after the bleed-
ing event. Starting sooner than 7 weeks may possibly involve
an increased risk of severe bleeding. Changing patient profile
characteristics did not change the optimal treatment initiation
 Pennlert et al Optimal Timing of AC in Patients With AF After ICH 319
time point although the magnitude of the estimated effect of
treatment varied.
Awaiting results from randomized controlled trials, obser-
vational studies may constitute the best available scientific evi-
dence. The present results are in agreement with those of recent
Danish observational studies,8,9 in that an overall benefit of
anticoagulant treatment after ICH in patients with AF has been
shown. This supports the generalizability of the current findings.
Our nationwide study is, by far, the largest specifically
addressing the problem of when to start or reinitiate antico-
agulant treatment in patients with AF who have had an ICH. In
contrast to previous smaller studies,12,13 the number of patients
in the present study has permitted adjustments for available
confounders with sufficient statistical power. Rather than using
summary scores such as CHA2DS2-VASc and HAS-BLED
(bleeding risk score assigning 1 point for the presence of each
of the following: hypertension [uncontrolled systolic blood
pressure >160 mm Hg], abnormal renal and/or liver function,
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017
previous stroke, bleeding history or predisposition, labile inter-
national normalized ratios, elderly, and concomitant drugs and/
or alcohol excess), individual comorbidities were used in the
statistical models. Because of a modest and not statistically
significant increase in bleeding events in patients with start of
anticoagulant treatment early after ICH, the net benefit for all
events was not statistically significant before 7 weeks. Even if
there are net benefits in subgroups up to 16 weeks, particularly
in high-risk patients, the opportunity for early effective second-
ary prevention should not be missed. It seems therefore that the
optimal timing of initiating anticoagulant therapy in patients
with AF would be at around 7 to 8 weeks after an ICH.
Similar timing patterns were observed in high-risk and
low-risk patients prescribed anticoagulants. In absolute terms,
the time-dependent benefits were much larger in high-risk
patients. Therefore, the timing of onset of anticoagulant treat-
ment seems to be particularly critical in these patients. In low-
risk patients, the difference in risk for recurrent stroke and
vascular death after 3 years was not significant in favor of
starting anticoagulant therapy at 8 weeks. However, the uncer-
tainty of the estimate was large, why these data not could be
used as an argument against starting anticoagulant therapy in
patients with low CHA2DS2-VASc scores.
In the present study, the proportion of patients with AF
surviving ICH who received anticoagulants was low (8.9%).
This seems to reflect the fact that physicians have found the
scientific evidence for anticoagulant treatment after an ICH in
patients with AF to be lacking or insufficient. No randomized
trial has been published, and only recently has support from
observational studies been published showing that anticoagu-
lant treatment reduces all-cause mortality and thromboembolic
events, even after severe hemorrhagic events.8,9,25 A recent
report on ICH survivors with AF also implies that, paradoxi-
cally, the higher the risk of thromboembolic events according
to CHA2DS2-VASc score, the lower the probability to receive
anticoagulants within the first months after ICH.1 Because
patients with higher CHA2DS2-VASc score tend to have an ele-
vated risk of severe bleeding as estimated by the HAS-BLED
score,26 this may reflect clinicians and patients reasonable
tendency to minimize risks in the absence of strong evidence.
The between-hospital variation in the use of anticoagulant
therapy after ischemic stroke in patients with AF is large in
Sweden, ranging from 36% to 100%.14 It therefore seems that
hospital traditions and individual doctors attitudes are major
determinants of the use of anticoagulants in stroke patients with
AF. In the present study, we have taken advantage of this ran-
dom-like heterogeneity in exposure. Yet, the possibility of resid-
ual confounding remains. There was a lower risk of death from
causes other than thrombotic and hemorrhagic events in patients
in whom anticoagulant and antiplatelet treatment was started in
the first months after ICH (unpublished data). This may have
resulted from a selection where patients with an apparent high
risk for nonvascular death were not treated with antithrombotic
agents. The increased thrombotic event risk seen for antiplatelet-
treated patients could be a real finding, but it could also be a sign
of confounding by indication. Although we were able to adjust
for several important confounding factors, we cannot fully adjust
for the doctors view of the patients overall health status.
Our study has several limitations. First, in this register-based
study, we have not had access to brain imaging data to distin-
guish between the different subtypes of ICH (lobar and deep
ICH involve different risks of recurrent bleeding27). The pro-
portion that had lowered consciousness at onset of ICH (as a
proxy for ICH severity) was somewhat lower in patients pre-
scribed anticoagulants than in the other 2 groups (Table IV in
the online-only Data Supplement), this difference was adjusted
for. Second, validation studies have shown that Riksstroke cov-
ers 94% of all hospital admissions for acute stroke14 and that,
conversely, there is an overdiagnosis of acute stroke in routine
hospital practice and in the Swedish Cause of Death register.28
The national drug register is essentially complete.15 Third, the
present study was performed before the large-scale introduc-
tion in routine clinical practice of new oral anticoagulants.
Therefore, it was not possible to analyze separately the optimal
timing of starting new oral anticoagulants after ICH in patients
with AF. Fourth, we had no data on functional outcome after
the recurrent stroke. Thus, case fatality is the only variable
used to describe the severity of the recurrent stroke event, sig-
nificantly higher after recurrent ICH than after recurrent isch-
emic strokes. A final limitation is that we had no information
on antithrombotic therapy given in hospital and that adherence
to anticoagulant or antiplatelet treatment during the follow-up
period was not measured. The analyses were by intention-to-
treat, whereas the on-treatment effects may have been greater.
A randomized controlled trial on anticoagulants in patients
with AF who have had an ICH has recently been initiated.19 It
will provide more definite evidence on the size of beneficial
and adverse effects than can be obtained from observational
studies. It is, however, unlikely that a randomized trial will
provide information on the optimal timing over a wide time
span in the same way an observational study can.
Conclusions
The optimal timing of starting anticoagulant treatment in
patients with AF who have survived an ICH seems to be around
7 to 8 weeks after the hemorrhage. In high-risk patients, anti-
coagulant treatment started in this interval reduces not only
the risk of thrombotic events but also the combined risk of
 320StrokeFebruary 2017
vascular death and nonfatal stroke. If treatment is started in this
interval, there seems to be no excess risk of major bleeding.
Acknowledgments
We thank the members of the Riksstroke Collaboration (http://www.
Riksstroke.org).
Sources of Funding
This study was supported by grants from the Swedish Research
Council for Health, Working Life and Welfare (grant no. 2011-
0657), the Swedish Research Council (20112395), by the foun-
dations for medical research administered by Ume University, by
the Foundation for Stroke Research in Northern Sweden and by the
County Council of Vsterbotten.
Disclosures
Dr Pennlert has served on a scientific advisory board for, and has
received lecture fees from, Bayer. Dr Wiklund has served on a sci-
entific advisory board for Boehringer-Ingelheim. The other authors
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017
report no conflicts.
References
1. Pennlert J, Asplund K, Carlberg B, Wiklund PG, Wisten A, sberg S,
et al. Antithrombotic treatment following intracerebral hemorrhage in
patients with and without atrial fibrillation. Stroke. 2015;46:20942099.
doi: 10.1161/STROKEAHA.115.009087.
2. Hacke W. The dilemma of reinstituting anticoagulation for patients
with cardioembolic sources and intracranial hemorrhage: how
wide is the strait between Skylla and Karybdis? Arch Neurol.
2000;57:16821684.
3. Rabinstein AA, Gupta A. Restarting anticoagulation after intracranial
hemorrhage: a risky decision with no recipe. Neurology. 2014;82:1016
1017. doi: 10.1212/WNL.0000000000000256.
4. Steiner T. Resumption of oral anticoagulation after warfarin-associ-
ated intracerebral hemorrhage: yes. Stroke. 2011;42:36613662. doi:
10.1161/STROKEAHA.111.621797.
5. Schulman S. Resumption of oral anticoagulation after warfarin-asso-
ciated intracerebral hemorrhage: no. Stroke. 2011;42:36633664. doi:
10.1161/STROKEAHA.111.621813.
6. Pennlert J, Eriksson M, Carlberg B, Wiklund PG. Long-term risk
and predictors of recurrent stroke beyond the acute phase. Stroke.
2014;45:18391841. doi: 10.1161/STROKEAHA.114.005060.
7. Friberg L, Rosenqvist M, Lip GY. Evaluation of risk stratification
schemes for ischaemic stroke and bleeding in 182 678 patients with atrial
fibrillation: the Swedish Atrial Fibrillation cohort study. Eur Heart J.
2012;33:15001510. doi: 10.1093/eurheartj/ehr488.
8. Nielsen PB, Larsen TB, Skjth F, Gorst-Rasmussen A, Rasmussen
LH, Lip GY. Restarting anticoagulant treatment after intracranial hem-
orrhage in patients with atrial fibrillation and the impact on recurrent
stroke, mortality, and bleeding: a nationwide cohort study. Circulation.
2015;132:517525. doi: 10.1161/CIRCULATIONAHA.115.015735.
9. Ottosen TP, Grijota M, Hansen ML, Brandes A, Damgaard D, Husted
SE, et al. Use of antithrombotic therapy and long-term clinical out-
come among patients surviving intracerebral hemorrhage. Stroke.
2016;47:18371843. doi: 10.1161/STROKEAHA.116.012945.
10. Steiner T, Al-Shahi Salman R, Beer R, Christensen H, Cordonnier C, Csiba
L, et al; European Stroke Organisation. European Stroke Organisation
(ESO) guidelines for the management of spontaneous intracerebral hem-
orrhage. Int J Stroke. 2014;9:840855. doi: 10.1111/ijs.12309.
11. Hemphill JC 3rd, Greenberg SM, Anderson CS, Becker K, Bendok BR,
Cushman M, et al; American Heart Association Stroke Council; Council
on Cardiovascular and Stroke Nursing; Council on Clinical Cardiology.
Guidelines for the Management of Spontaneous Intracerebral
Hemorrhage: A Guideline for Healthcare Professionals From the
American Heart Association/American Stroke Association. Stroke.
2015;46:20322060. doi: 10.1161/STR.0000000000000069.
12. Majeed A, Kim YK, Roberts RS, Holmstrm M, Schulman S. Optimal
timing of resumption of warfarin after intracranial hemorrhage. Stroke.
2010;41:28602866. doi: 10.1161/STROKEAHA.110.593087.
13. Hawryluk GW, Austin JW, Furlan JC, Lee JB, OKelly C, Fehlings MG.
Management of anticoagulation following central nervous system hem-
orrhage in patients with high thromboembolic risk. J Thromb Haemost.
2010;8:15001508. doi: 10.1111/j.1538-7836.2010.03882.x.
14. Riksstroke. Riksstroke rsrapport 2014 (annual report 2014). http://
www.riksstroke.org/wp-content/uploads/2015/12/Strokerapport_AKUT-
TIA_LR.pdf. Accessed January 19, 2016.
15. Wettermark B, Hammar N, Fored CM, MichaelFored C, Leimanis A,
Otterblad Olausson P, et al. The new Swedish Prescribed Drug Register
opportunities for pharmacoepidemiological research and experience
from the first six months. Pharmacoepidemiol Drug Saf. 2007;16:726
735. doi: 10.1002/pds.1294.
16. Ludvigsson JF, Andersson E, Ekbom A, Feychting M, Kim JL,
Reuterwall C, et al. External review and validation of the Swedish
national inpatient register. BMC Public Health. 2011;11:450. doi:
10.1186/1471-2458-11-450.
17. Statistics Sweden. Longitudinell integrationsdatabas fr hl-
sofrskrings- och arbetsmarknadsstudier (LISA) (longitudinal integra-
tion database for Health Insurance and Labour Market Studies). http://
www.Scb.Se/lisa. Accessed December 03, 2014.
18. National Board of Health and Welfare. Ddsorsaksregistret(Cause of
Death Registry). http://www.socialstyrelsen.se/register/dodsorsaksregis-
tret. Accessed January 19, 2016.
19. van Nieuwenhuizen KM, van der Worp HB, Algra A, Kappelle LJ,
Rinkel GJ, van Gelder IC, et al; APACHE-AF Investigators. Apixaban
versus Antiplatelet drugs or no antithrombotic drugs after anticoagu-
lation-associated intraCerebral HaEmorrhage in patients with Atrial
Fibrillation (APACHE-AF): study protocol for a randomised controlled
trial. Trials. 2015;16:393. doi: 10.1186/s13063-015-0898-4.
20. Lip GY, Nieuwlaat R, Pisters R, Lane DA, Crijns HJ. Refining clinical
risk stratification for predicting stroke and thromboembolism in atrial
fibrillation using a novel risk factor-based approach: the euro heart
survey on atrial fibrillation. Chest. 2010;137:263272. doi: 10.1378/
chest.09-1584.
21. Therneau TM. A Package for Survival Analysis in S. R package version 2.38,
2015. http://CRAN.R-project.org/package=survival. Accessed January 19,
2016.
22. Thomas L, & Reyes E. Tutorial: survival estimation for Cox regression
models with time-varying coeffcients using SAS and R. J Stat Softw.
2014;61:123.
23. Beyersmann J, Latouche A, Buchholz A, Schumacher M. Simulating
competing risks data in survival analysis. Stat Med. 2009;28:956971.
doi: 10.1002/sim.3516.
24. R Core Team. R: A language and environment for Statistical Computing.
Vienna Austria: R Foundation for Statistical Computing; 2014.
25. Staerk L, Lip GY, Olesen JB, Fosbl EL, Pallisgaard JL, Bonde AN,
et al. Stroke and recurrent haemorrhage associated with antithrombotic
treatment after gastrointestinal bleeding in patients with atrial fibrilla-
tion: nationwide cohort study. BMJ. 2015;351:h5876.
26. Apostolakis S, Lane DA, Buller H, Lip GY. Comparison of the CHADS2,
CHA2DS2-VASc and HAS-BLED scores for the prediction of clinically
relevant bleeding in anticoagulated patients with atrial fibrillation: the
AMADEUS trial. Thromb Haemost. 2013;110:10741079. doi: 10.1160/
TH13-07-0552.
27. Viswanathan A, Rakich SM, Engel C, Snider R, Rosand J, Greenberg
SM, et al. Antiplatelet use after intracerebral hemorrhage. Neurology.
2006;66:206209. doi: 10.1212/01.wnl.0000194267.09060.77.
28. Kster M, Asplund K, Johansson , Stegmayr B. Refinement of Swedish
administrative registers to monitor stroke events on the national level.
Neuroepidemiology. 2013;40:240246. doi: 10.1159/000345953.
 Optimal Timing of Anticoagulant Treatment After Intracerebral Hemorrhage in Patients
With Atrial Fibrillation
Johanna Pennlert, Rosanna Overholser, Kjell Asplund, Bo Carlberg, Bart Van Rompaye,
Per-Gunnar Wiklund and Marie Eriksson
Downloaded from http://stroke.ahajournals.org/ by guest on January 26, 2017

Stroke. 2017;48:314-320; originally published online December 20, 2016;

doi: 10.1161/STROKEAHA.116.014643
Stroke is published by the American Heart Association, 7272 Greenville Avenue, Dallas, TX 75231
Copyright 2016 American Heart Association, Inc. All rights reserved.
Print ISSN: 0039-2499. Online ISSN: 1524-4628

The online version of this article, along with updated information and services, is located on the
World Wide Web at:
http://stroke.ahajournals.org/content/48/2/314

Data Supplement (unedited) at:

http://stroke.ahajournals.org/content/suppl/2016/12/20/STROKEAHA.116.014643.DC1.html

Permissions: Requests for permissions to reproduce figures, tables, or portions of articles originally published
in Stroke can be obtained via RightsLink, a service of the Copyright Clearance Center, not the Editorial Office.
Once the online version of the published article for which permission is being requested is located, click
Request Permissions in the middle column of the Web page under Services. Further information about this
process is available in the Permissions and Rights Question and Answer document.

Reprints: Information about reprints can be found online at:

http://www.lww.com/reprints

Subscriptions: Information about subscribing to Stroke is online at:

http://stroke.ahajournals.org//subscriptions/
 ONLINE SUPPLEMENT

Optimal timing of anticoagulant treatment following intracerebral hemorrhage in

patients with atrial fibrillation
Authors: Johanna Pennlert, MD, Rosanna Overholser, PhD, Kjell Asplund, MD, PhD, Bo
Carlberg, MD, PhD, Bart Van Rompaye, PhD, Per-Gunnar Wiklund, MD, PhD, Marie
Eriksson, PhD

Author affiliations:
Johanna Pennlert, Kjell Asplund, Bo Carlberg, Per-Gunnar Wiklund: Department of Public
Health and Clinical Medicine, Medicine, Ume University, Sweden
Rosanna Overholser, Bart Van Rompaye: Department of Applied Mathematics, Computer
Science and Statistics, Ghent University, Belgium and the Department of Statistics, Ume
School of Business and Economics, Ume University, Sweden
Marie Eriksson: Department of Statistics, Ume School of Business and Economics, Ume
University, Sweden

SUPPLEMENTARY MATERIAL
Table of contents Content Page(s)

Supplementary Methods 3-10

Table I List of ICD-10 codes to define outcome variables 3

Table II List of ICD-10 codes and registers used to assess 5

comorbidity at baseline

Statistical Methods 6-10

Table III Cause-specific hazard ratios with 95% confidence 6

intervals for outcome events

Figure I Cause-specific log hazard ratios with 95% confidence 9

intervals (CIs) as a function of start time of treatment
(AC or AP).

Figure II Log hazard ratios with 95% confidence intervals (CIs) 10

for time since hospital discharge.

Figure III Cause-specific log hazard ratios for age. 10

Supplementary Results 11-13

Table IV Baseline characteristics of patients included in 11

analysis (n=2619) and treatment status at 8 weeks

Figure IV Cumulative incidence functions (CIFs), adjusted for 13

differences in patient characteristics, at three years
after onset of ICH vs. start time of treatments in
additional patient risk profiles.

Figure V Empirical cumulative incidence functions (CIF) in 14

patients with AC, AP or no antithrombotic treatment
8 weeks after ICH.

Supplemental References 15

2
Supplementary Methods.
Table I. List of ICD-codes used to define outcome variables.
Severe hemorrhagic outcome events

Recurrent intracerebral hemorrhage I61, from Riksstroke (fatal or non-fatal)

Data below derived from the Swedish Cause of Death register:

Intracranial

Subarachnoidal hemorrhage I60

Subdural hemorrhage I62

Epidural hemorrhage G95.1

GI-tract

Oesophageal hemorrhage I850, I983

Gastro-oesophageal laceration-hemorrhage syndrome K22.6

Ulcer with hemorrhage/Duodenal hemorrhage K25-28 only subcodes 0-2, 4-6

Unspecified GI-hemorrhage K922

Haemorrhojdal bleeding I841, I858, I844

Hemorrhage of anus and rectum K625

Haematemesis K920

Melena K921

Haemoperitoneum K66.1

Acute hemorrhagic gastritis K29.0

Joint

Haemarthrosis M250

UG tract

Hemorrhage of prostate N421

Other uterine/vaginal bleeding N93.8-9

Postmenopausal bleeding N95.0

Haematuria R31, N02

Haematometra N85.7

3
 Nose/throat/respiratory tract

Hemorrhage from throat/airways, unspecified R048, R049

Hemorrhage from throat R04.1

Hemoptysis R04.2

Epistaxis R04.0

Eye

Conjunctival hemorrhage H11.3

Choroidal hemorrhage H31.3

Retinal hemorrhage H35.6

Vitreous hemorrhage H43.1

Other bleedings

Hemopericardium I31.2

Anemia following major bleeding D629

Hemorrhage, not elsewhere classified R58.9

Otorrhagia H92.2

Hematoma of broad ligament N83.7

Postoperative hemorrhage T810

Severe thrombotic events

Ischemic stroke I63, from Riksstroke (fatal or non-fatal)

Data below from the Swedish Cause of Death register:

Other fatal thromboembolic events I21,I22, I23, I74

4


Table II. List of ICD-10 codes and registers used to assess comorbidity at baseline. IPR The Swedish Inpatient
register, RS Riksstroke the Swedish Stroke Register, SDDR the Swedish Dispensed Drug Register.
Baseline covariates ICD-10 code(s) and register origin

Atrial fibrillation I48 RS or IPR

Stroke severity RS - 3 levels

Ischemic Stroke/TIA I63,G45 or RS

Ischemic Stroke I63, IPR

Heart Failure I50, I43, I099, I110, I130, I132, I255, I420, I425, I426, I427, I428, I429, IPR

History of VTE I26, I80 (except 180.0) O87.1, O22.3, O22.5, I81, I82, IPR

Ischemic heart disease I20 (except I20.1), I21 I22 I23 I24 I25.2, IPR

I65, I70-73 (except I71.1, I71.2, I73.0, I73.1, I73.8, I73.8A, I73.8B, I73.8C, I73.8D,
Peripheral arterial disease I73.8W and I73.9A), IPR

I05-08, I09.1, I34-39, Z95.2, Z95.3, Z95.4, I342, I050, Q232, procedure codes FG, FJE,
Valvular diease FJF, FK, FM, IPR

Hypertension I10-15 IPR or RS

Diabetes Mellitus E10-14, IPR or RS

Dementia F00-F03, IPR

Hyperlipidemia E78, IPR or treated with statins < 6 months prior to ICH, found in the SDDR

Smoking RS

ADL dependency RS

Living alone RS

5
Supplementary Statistical methods
Cause-specific hazards

We first built a Cox PH1 model of the following form for each of the three causes separately, each time treating
the other events as censorings:

a piecewise linear function of age, with turning points at the quartiles,

a smoothing spline of AC start time, with an indicator of treatment as a time varying covariate, in three
steps: 1 month, 3 months or 6 months after onset of ICH, to reduce the possibility of overfitting, the shape
was constrained to be linear after 272 days after the treatment start time for AC (90th percentile),
a smoothing spline of AP start time, with an indicator of treatment as a time varying covariate, in three
steps: 1 month, 3 months or 6 months after onset of ICH, to reduce the possibility for overfitting, the shape
was constrained to be linear after 482 days after the treatment start time for AP (90th percentile),
a smoothing spline of time since hospital discharge, with linear behavior after 186 days after hospital
discharge,
a dummy coding of each of the 16 categorical variables listed under "Patient Characteristics".

Our model selection procedure started from the complete model, and used backwards selection as follows:

1. First, we removed covariates which lead to a decrease in AIC, while protecting the treatment smoothing
splines as the covariates of interest, the smoothing spline for the second time-scale, and the commonly
accepted risk factors of age, sex and consciousness at admission.
2. Secondly, we assessed the need for the second time-scale, time since hospital discharge.
3. Next, we reduced the complexity of the treatment effects: removing the steps (1, 3 or 6 months after onset
of ICH) in the indicator of treatment for AP and AC if the reduced model had a lower AIC.
4. We further reduced the covariate effects using AIC, this time also allowing sex and consciousness to be
excluded, but still protecting age, and a single smoothing spline for each treatment during the treatment
period.
5. Finally, we reassessed the need for a second-time scale.

Smoothing parameters for each smoothing spline were chosen by AIC. Models were selected on the subset of
patients with complete covariate information and re-fit using data from all patients with covariate information on
the chosen covariates.

The cause-specific hazard ratios for the covariates without splines from these three models are shown in table III.
The smoothing splines of start time of AP and AC treatments and the second time scale are shown in figures I
and II, along with pointwise 95% confidence intervals. The effect of age, which was modeled as a piecewise
linear function, is shown in figure III.

Table III: Cause-specific hazard ratios with 95% confidence intervals for severe thrombotic events, hemorrhagic
events and for other deaths. Separate Cox proportional hazard models were built for three events with a
piecewise linear function of age and a smoothing spline of start time for each of the two treatments. See figures
I, II and III for a visual representation of smoothing splines of treatment start times, time since discharge and of
the age effects. The table lists variables that entered into at least one of the three models: blanks in the table
mean that the variable in the corresponding row did not come into the model of the corresponding column. The
following variables were not significant in any model: education, living alone, venous thromboembolism,
valvular disease, ischemic heart disease, statin treatment before stroke, or hospital. Data are not given on
variables eliminated in the backward selection processes.

Thrombotic events Hemorrhagic events Other Deaths

95% 95% 95% 95% 95% 95%
HR LCL UCL HR LCL UCL HR LCL UCL
Sex: Female 1.31 1.06 1.61 -- -- -- -- -- --
Age (per 1 yr) 1.05 1.01 1.09 1.05 0.98 1.11 1.06 103 110
> 73 yrs 0.98 0.90 106 0.95 0.83 109 103 096 110
> 80 yrs 0.97 0.87 1.09 1.05 0.84 1.29 0.99 0.92 1.08
> 85 yrs 1.02 0.90 1.15 0.96 0.76 1.20 0.97 0.90 1.04

6
Baseline AC 1.57 1.22 2.02 -- -- -- -- -- --
Baseline AP 1.55 1.21 1.99 -- -- -- -- -- --
ADL before ICH -- -- -- -- -- -- 2.01 1.67 2.43
Dementia -- -- -- -- -- -- 1.43 1.12 1.82
Diabetes 1.25 0.99 1.57 -- -- -- -- -- --
Heart failure -- -- -- -- -- -- 1.46 1.25 1.70
Hypertension -- -- -- -- -- -- 0.82 0.69 0.98
PAD -- -- -- -- -- -- 1.39 1.05 1.85
Previous IS 1.31 1.05 1.64 -- -- -- 1.26 1.07 1.48
Status: Alert (ref)
Drowsy -- -- -- -- -- -- 1.46 1.23 1.73
Missing -- -- -- -- -- -- 0.99 0.58 1.68
Unconscious -- -- -- -- -- -- 2.38 1.65 3.45

Cumulative incidence functions

The cause-specific CIF was computed from the cause-specific instantaneous hazard functions and the overall
survival function2. More specifically, let ! (| ()) be the cause-specific instantaneous hazard of cause at
time given the covariate trajectory (), which may include information on both time-independent baseline
characteristics and time-dependent characteristics, up to time . Let (| ()) be the probability of event-free
survival up to time given the covariate trajectory (). Then, the cumulative incidence function for cause
given a covariate trajectory is:
!
CIF! (| ()) = !
(| ())! (| ()) .

Furthermore, the all-cause CIF is the sum of the individual cause-specific CIFs.

We followed a Cox proportional hazards (PH) model and assumed that each instantaneous hazard function was
of the form

! (| ()) = !,! ()exp( ()! ),

where !,! (), the baseline hazard, is assumed to be independent of the covariate trajectories, and exp( ()! )
describes the effect of the covariate trajectory on the hazard at time .

The model parameter ! for a given cause was estimated by fitting a Cox PH model for ! (| ()) on the data
from = 2662 subjects, with covariates from the model building procedure described earlier and with all events
but event type censored. Let ! denote this estimate. We used a Breslow type estimator for the cumulative
hazard function for cause , (| ()),

! ! !"#(! (!)!! )!!!,! (!)

! (| ()) = !!! ! ,
! !!,! (!)!"#(! (!)!! )

where !,! () is a indicator for event at time for the th subject and !,! () is an indicator for subject being
at risk of event at time . Given ! (| ()) for each cause, the overall survival funtion was estimated as:

(| ()) = exp( ! ! (|

())).

Since the baseline instantaneous hazard !,! () is not specified in a Cox PH model, we chose to approximate
each ! (| ()) with a piecewise constant function and thus arrived at the estimate
!
CIF! (| () =
!!! (! | (! ))[! (! | (! )) ! (!!! | (!!! ))]/(! !!! ),

where ! , ! , . . . ! is a partition of (0, ).

This calculation required specification of a covariate trajectory: we considered male and female versions of the
"high risk" and "low risk" profiles described in the manuscript. In order to explore the effect of varying treatment

7
type and start time, we computed the cause-specific and all-cause CIFs for a range of start-times (28 to 365 days
after onset of ICH, in increments of 10) and for each treatment type (AP, AC, or neither).

Confidence intervals for the cumulative incidence functions

We used a parametric bootstrap3 to obtain confidence intervals for the difference in CIFs for patients with and
without treatment (AC or AP). A bootstrap sample dataset was created from the original set of patient
characteristics as described below.

For each patient in the dataset:

1. an event time was generated from the estimated all-cause survival function using the inverse probability
transform,
2. the type of event was decided using a multinomial experiment with probabilities of each event type equal
to the proportions of cause-specific hazards at the event time from step 1,
3. a censoring time was generated from the Kaplan Meier estimate of survival from the observed censor
times,
4. the minimum of the times in 1 and 3 was selected and the associated event type chosen.

Once a bootstrap sample dataset was created, the three Cox PH models for thrombotic events, hemorrhagic
events and deaths from other causes were refit, using the bootstrapped outcomes but the same covariates as
selected for the original dataset. Next, the cause-specific CIFs were estimated as for the original dataset.

The sampling distributions of the cause-specific CIFs were estimated from the above bootstrap scheme and
examined via QQ-plots. Since these distributions appeared to be normal, we computed 95% confidence intervals
for the estimate of the differences in CIF with and without treatment (AC or AP) at a given start time as:

estimate +/- 1.96 standard deviation (bootstrap estimates).

Treatment start times where a significance difference was found, i.e. regions where the CI of the difference did
not include 0) are indicated by thick lines in plots of cause-specific and all-cause CIFs for the male and female
versions of our high and low risk patient profiles (manuscript figure 3).

Sensitivity analysis of minimizing AC start time

We define the optimal time to start AC treatment for a given patient profile as the AC treatment start time that
minimizes the sum of the CIFs of thrombotic and hemorrhagic events at 3 years after ICH: this time can be seen
for our profiles of high and low risk patients by locating the time on the horizontal axis that corresponds to the
minimum value of the CIF of vascular death or stroke 3 years after onset of ICH on the AC treatment curve
(manuscript Figure 3). To see how this optimal time changed across patient profiles, we computed the relevant
CIF curve for a series of patient profiles, starting from a profile with no risk factors, and adding the covariates
shown in supplement table III one at a time until all risk factors were included. Risk factors were added in order
of the magnitude of the associated cause-specific hazard ratio; four different ages were considered, one in each
quartile of the observed ages, and two hospital discharge times. The resulting optimal start times were all found
to be within the range of 7-8 weeks after onset of ICH. Because our model will suggest similar optimal times for
profiles with risk factors that correspond to similar sums of log cause specific hazard ratios for a given event
type in supplement table III (for example, the optimal start time of a profile with a single risk factor of baseline
AC will be similar to a profile with a single risk factor of baseline AP), we expect this finding to be relatively
robust. Supplemental Figure IV demonstrates the CIFs for two additional patient profiles.

Software

We used the statistical software R, version 3.2.1, for all calculations. The package "survival" was used to fit Cox
PH models with time-varying covariates and smoothing splines1.The calculations for the CIFs were programmed
using the formulas mentioned before, which relied on the results of Cox PH models.

8
Additional figures of cause-specific instantaneous hazard functions

Thrombotic Hemorrhagic Other deaths

log (instan. hazard ratio) with pointwise 95% CIs

2.5

0.0

AC
2.5

5.0

1.0

0.5

0.0

AP
0.5

1.0

0 4 8 12 16 20 24 28 32 36 40 44 48 52 0 4 8 12 16 20 24 28 32 36 40 44 48 52 0 4 8 12 16 20 24 28 32 36 40 44 48 52

Weeks since onset of ICH

Figure I: Cause-specific log hazard ratios with 95% confidence intervals (CIs) as a function of start time of
treatment (AC or AP).
Interpretation of figure I
Top left: the instantaneous hazard of a thrombotic event is significantly lower compared to no treatment when
AC treatment was started within 12 weeks of the onset of ICH, holding all other covariates in the model
constant; no difference was observed after 12 weeks. Bottom left: the instantaneous hazard of a thrombotic event
was significantly higher for all patients who started AP in the year following ICH, compared to patients who did
not start a treatment. Top center: no significant difference in instantaneous hazard of a hemorrhagic event was
observed between patients who received AC or no treatment in the first year after ICH. Bottom center: the
instantaneous hazard of a hemorrhagic event was observed to be significantly higher for patients who received
AP in the first year after ICH compared with patients who received no treatment. Top right: no significant
difference was found in instantaneous hazard of death from another cause in patients who received AC or no
treatment. Bottom right: patients who received AP around 22 weeks after ICH had a significantly lower
instantaneous hazard of death from another cause compared to those patients who received no treatment; no
difference was detected when AP treatment was started at other times during the first year after ICH.

9
 log (instan. hazard ratio) with pointwise 95% CIs Other deaths

0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 68 72 76
Weeks since hospital discharge

Figure II: Log hazard ratios with 95% confidence intervals (CIs) for time since hospital discharge.
Interpretation of Figure II
While the instantaneous hazards of hemorrhagic or thrombotic events appear to be primarily driven by time since
onset of ICH, part of the hazard of death from other causes is driven by the time since hospital discharge. The
hazard generally decreases as time since discharge increases.

Thrombotic Hemorrhagic Other deaths

6
log (instan. hazard ratio)

50 60 70 80 90 50 60 70 80 90 50 60 70 80 90
Age (years)

Figure III: Cause-specific log hazard ratios for age.

Interpretation of Figure III
Age was included in the models as a piece-wise linear function with turning points at the observed quartiles. All
three risks increase with age for younger subjects; the risk of death from other causes increases for all age
groups. The observed ages ranged from 42 to 98.

10
Supplementary results

Table IV: Baseline characteristics of patients included in analysis (n=2619) and treatment status at 8 weeks
(trtm treatment, AC anticoagulant, AP antiplatelet, SD standard deviation, ADL activities of daily
living, IS ischemic stroke, VTE venous thromboembolism). P-vaules for group comparisons between AC-
and AP-treated vs non-treated patients at 8 weeks (2-tests for categorical variables, t-test for age).

No trtm at 8 AC at 8 weeks AP at 8 weeks

weeks (n=97) (n=489)
(n=2033) p-value p-value

Mean age (SD) 77.86(9.18) 76.62(9.011) 0.194 78.9(8.5) 0.018

Female Sex 817(40.2) 39(40.2) 0.997 209(42.7) 0.302

Level of 0.837 0.348

Education

Primary 1012(49.8) 45(46.4) 252(51.5)

Secondary 692(34.0) 35(36.1) 173(35.4)

University 288(14.2) 14(14.4) 58(11.9)

Missing 41(2.0) 3(3.1) 6(1.2)

Living Alone

Yes 972(47.8) 47(48.5) 0.902 247(50.5) 0.283

Missing 4(0.2) 0(0) 4(0.8)

ADL
dependency

Yes 238(11.7) 5(5.2) 0.047 58(11.9) 0.924

Level of 0.057* 0.002

consciousness*

Alert 1514(74.5) 81(83.5) 404(82.6)

Drowsy 412(20.3) 14(14.4) 69(14.1)

Unconscious 68(3.3) 1(1.0) 11(2.2)

Missing 39(1.9) 1(1.0) 5(1)

Diabetes 465(22.9) 36(37.1) 0.001 104(21.3) 0.446

Mellitus

Previous IS 466(22.9) 30(30.9) 0.068 144(29.4) 0.002

VTE 109(5.4) 10(10.3) 0.038 21(4.3) 0.338

Ischaemic heart 535(26.3) 29(29.9) 0.435 149(30.5) 0.064

disease

11
Hypertension 1676(82.4) 85(87.6) 0.187 419(85.7) 0.086

Heart Failure 489(24.1) 30(30.9) 0.123 153(31.3) 0.001

Valvular 155(7.6) 19(19.6) 0.000 39(8.0) 0.794

disease

Peripheral 117(5.8) 7(7.2) 0.548 32(6.5) 0.506

arterial disease

Dementia 12(6.0) 0(0) 0.013 31(6.3) 0.778

AC at baseline 916(45.1) 77(79.4) 0.000 246(50.3) 0.037

AP at baseline 884(43.5) 20(20.6) 0.000 271(55.4) 0.000

Both AC and 147(7.2) 8(8.2) 0.706 50(10.2) 0.027

AP at baseline

Statins 527(25.9) 34(35.1) 0.046 162(33.1) 0.001

*For level of consciousness, the variables were grouped into alert, drowsy/unconscious when comparing
AC treated patients to non-treated.

12
 A
Hypertension, previous AC Diabetes, previous AC
0.4

Thrombotic
0.3
0.2
0.1
CIF at 3 years after ICH

0.0
0.4

Hemorrhagic
0.3
0.2
0.1
0.0
0.4

Vasc. death
0.3

or stroke
0.2
0.1
0.0
0 4 8 12 16 20 24 28 32 36 40 44 48 0 4 8 12 16 20 24 28 32 36 40 44 48
Rx start time (weeks since ICH)

B
Hypertension, previous AC Diabetes, previous AC
0.4

Thrombotic
0.3
0.2
0.1
CIF at 3 years after ICH

0.0
0.4

Hemorrhagic
0.3
0.2
0.1
0.0
0.4
Vasc. death
0.3
or stroke

0.2
0.1
0.0
0 4 8 12 16 20 24 28 32 36 40 44 48 0 4 8 12 16 20 24 28 32 36 40 44 48

Rx after ICH AP AC none

Figure IV: Cumulative incidence functions (CIFs), adjusted for differences in patient characteristics, at three
years after onset of ICH vs. start time of two treatments (AC=black line, AP=grey line) and no treatment (dashed
line) for a 78-year old with hypertension and previous AC and a 78-year old with diabetes and previous AC,
female/male patient profile (A=female profiles, B=male profiles). The event specific CIFs (thrombotic and
hemorrhagic) sum to the vascular death or stroke CIFs in the bottom row. The thick lines represent time periods
during which treatment initiation of AC (black) and AP (grey) are significantly different from no treatment at the
5% level.

13
 Rx at 8 weeks after onset of ICH AC AP none

Vasc. death or stroke

0.5

0.4

0.3

0.2

0.1

0.0

Thrombotic
0.5
Empirical CIF

0.4

0.3

0.2

0.1

0.0

Hemorrhagic
0.5

0.4

0.3

0.2

0.1

0.0
0 1 2 3 4 5 6 7
Years after onset of ICH

Thrombotic 0 0 0 141 47 2 205 61 2 238 78 4 259 86 4 273 90 4 280 92 4 283 92 4

Hemorrhagic 0 0 0 50 13 3 66 16 4 77 17 5 85 17 5 91 18 5 91 18 6 91 18 6

Other deaths 0 0 0 309 59 6 443 89 10 528 111 10 573 130 10 611 139 12 623 140 16 630 145 17

At risk 2033 489 97 1291 317 69 904 225 42 591 148 29 372 85 21 218 44 14 113 24 3 33 4 1

Cumulative counts and patients at risk by year, colored by Rx at 8 weeks

Figure V: Empirical cumulative incidence functions (CIF) in patients with AC, AP or no antithrombotic
treatment 8 weeks after ICH.

14
Supplementary references

1. Therneau T. A Package for Survival Analysis in S. version 2.38, 2015. http://CRAN.R-

project.org/package=survival.
2. Thomas L, Reyes, E. Tutorial: Survival estimation for Cox regression models with time-
varying coeffcients using SAS and R. J Stat Softw. 2014;61(Cs1):1-23.
3. Beyersmann J, Latouche A, Buchholz A, Schumacher M. Simulating competing risks data in
survival analysis. Stat. Med. 2009;28:956-971

15