Keywords
non-valvular atrial fibrillation - stroke risk - bleeding risk
Introduction
Atrial fibrillation (AF) is the most common cardiac arrhythmia seen in clinical practice,
occurring in up to 6.1 million people in the United States and accounting for approximately
one-third of hospitalizations for cardiac rhythm disturbances.[1]
[2]
[3] Further, AF is associated with significant morbidity and mortality, including increased
risk of embolic stroke, heart failure and cognitive impairment; reduced quality of
life; and higher overall mortality.[4]
[5]
[6]
Optimal clinical management of AF is critical to reducing this associated morbidity
and mortality, and includes prevention of AF-related thromboembolic events in at-risk
patients. Vitamin K antagonists or direct oral anticoagulants have been shown to reduce
thromboembolic events, but long-term use of these medications puts certain patients
at higher risk for serious bleeding events. As such, accurate risk stratification
for both thromboembolic and bleeding risk is paramount in identifying patients for
whom anti-thrombotic therapy would achieve maximum treatment benefit with the lowest
risk of complications.
Unfortunately, it is challenging to estimate the trade-off between stroke risk and
risk of bleeding complications from long-term anticoagulation therapy because many
risk factors for stroke are also associated with increased risk of bleeding. There
are several available risk stratification tools used to determine thromboembolic and
bleeding risk that incorporate diagnostic imaging as well as patient factors such
as age, sex and history of heart disease to aid in clinical decision-making around
treatment strategies for AF. Of the many available risk stratification tools, the
2014 American Heart Association/American College of Cardiology/Heart Rhythm Society
(AHA/ACC/HRS) guideline for patients with AF recommends the use of the CHA2DS2-VASc score to estimate the stroke risk and the HAS-BLED score for bleeding risk.[7]
[8]
[9] However, these risk scores have been previously categorized as poor to moderate
predictors of risk, and are just two of many different published and validated methods
for assessing stroke and bleeding risk in patients with AF. Because patients, providers
and policymakers have numerous decision tools that could inform treatment decisions
and policy recommendations, there is a need for a compilation and analysis of the
currently available data. This systematic review was commissioned by the Patient-Centered
Outcomes Research Institute (PCORI) to update a 2013 Agency for Healthcare Research
and Quality (AHRQ) review,[10] with a focus on evaluating the comparative diagnostic accuracy and impact on clinical
decision-making of available clinical and imaging tools and associated risk factors
for predicting thromboembolic and bleeding risk in U.S. patients with AF. Our findings
related to stroke prevention treatments are discussed in a companion paper.
Methods
Methods for this updated comparative effectiveness review (CER) follow the AHRQ's
Methods Guide for Effectiveness and Comparative Effectiveness Reviews (hereafter referred to as the Methods Guide)[11] and Methods Guide for Medical Test Reviews (hereafter referred to as the Medical Test Guide).[12] This article is part of the larger updated review; complete details of our methods,
including exact search strings, and full results and conclusions can be found in the
full report, available at www.effectivehealthcare.ahrq.gov.
Defining the Key Questions
PCORI convened two multi-stakeholder virtual workshops in December 2016 and January
2017 to (1) gather input from end users and clinical, content and methodological experts
on scoping for the updated review; (2) prioritize the key questions; (3) discuss changes
in the evidence base since the 2013 review; and (4) explore emerging issues in AF.
The protocol for this systematic review was informed by discussion at the January
2017 workshop and builds upon the original report. The final protocol for this review
is posted on the Effective Health Care (EHC) website (www.effectivehealthcare.ahrq.gov) and registered at PROSPERO (CRD42017069999).
In this article, we summarize the evidence and findings related to two key questions
(KQs): (1) In patients with non-valvular AF, what are the comparative diagnostic accuracy
and impact on clinical decision-making (diagnostic thinking, therapeutic and patient
outcome efficacy) of available clinical and imaging tools and associated risk factors
for predicting thromboembolic risk? and (2) In patients with non-valvular AF, what
are the comparative diagnostic accuracy and impact on clinical decision-making (diagnostic
thinking, therapeutic and patient outcome efficacy) of clinical tools and associated
risk factors for predicting bleeding events?
Data Sources and Study Selection
In consultation with an expert medical librarian, we searched PubMed, Embase and the
Cochrane Database of Systematic Reviews for relevant literature published from 1 August
2011 to 14 February 2018 (exact search strings are given in [Supplementary Table S1], available in the online version). We supplemented electronic searches with a manual
search of citations from a set of systematic review articles. Our findings were combined
with those from the 2013 review, and so the literature summarized here reflects evidence
back through 1 January 2000.[10] Due to updates in inclusion criteria, any studies excluded from the original review
were also re-reviewed for eligibility. We used search criteria to identify relevant
on-going clinical trials through ClinicalTrials.gov as well as citations to guide
the conclusions ([Supplementary Table S1], available in the online version).
Our pre-specified inclusion and exclusion criteria are given in [Supplementary Table S2] (available in the online version). We included English-language studies of adults
with non-valvular AF (including atrial flutter) that reported the efficacy of clinical
or imaging tools, or patient risk factors, on predicting thromboembolic and/or bleeding
outcomes. Clinical or imaging tools considered for predicting thromboembolic events
were CHADS2 score, CHA2DS2-VASc score, Framingham risk score, age, biomarkers, and clinical history (ABC) stroke
score, transthoracic and transoesophageal echocardiography, computed tomography scans
and cardiac magnetic resonance imagings (MRIs). Clinical or imaging tools considered
for predicting bleeding events were the HAS-BLED score, HEMORR2HAGES score, Anticoagulation and Risk Factors in Atrial Fibrillation (ATRIA) score,
Bleeding Risk Index (BRI) and ABC bleeding risk score. Thromboembolic outcomes included
cerebrovascular infarction, transient ischaemic attack and systemic embolism (excluding
pulmonary embolism and deep vein thrombosis). Bleeding outcomes included haemorrhagic
stroke, intra-cranial haemorrhage (ICH) and major and minor bleeds. We excluded studies
that evaluated patients exclusively from Asia, Africa or the Middle East. We also
sought to identify studies which used the same patients and linked these as companion
papers to an individual study.
Data Extraction and Quality Assessment of Individual Studies
Pairs of investigators screened all citations and abstracts for eligibility, and those
considered relevant by either investigator advanced to full-text review. Paired investigators
then reviewed all full-text articles and resolved disagreements through discussion
or adjudication by a third investigator. Paired investigators independently abstracted
data and assessed study quality. Disagreements were resolved by consensus or arbitration
by a third investigator. Articles that represented evidence from the same overall
study were linked to avoid duplication of patient cohorts.
We assessed methodological quality, or risk of bias, for each individual study using
tools specific to the study's characteristics. For studies assessing diagnostic accuracy,
we used the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool.[13] Our outcome-specific quality assessment classified study outcomes as containing
low, medium or high risk of bias as defined by QUADAS-2.
Data Synthesis and Analysis
We summarized key features of the included studies for each KQ, including information
on study design; patient characteristics; clinical settings; diagnostic tools; and
intermediate, final and adverse event outcomes. We ordered our findings by diagnostic
comparison, and then within these comparisons by outcome, with long-term final outcomes
emphasized.
Grouping interventions by prediction tool, we determined the feasibility of completing
a quantitative synthesis (i.e. meta-analysis) based on the volume of relevant literature
(at least three appropriate studies), conceptual homogeneity of the studies in terms
of study population and outcomes and completeness of the reporting of results. When
at least three comparable studies reported the same outcome, we used the R statistical
package (version 3.1.2) (The R Foundation) with the ‘metafor’ meta-analysis library
(version 1.9–7) to synthesize available c-statistics, which quantify the discrimination ability of the studied tools, for each
appropriate thromboembolic or bleeding risk prediction tool. We used the random-effects
DerSimonian and Laird estimator[14] to generate summary values. In addition, we used the Knapp–Hartung approach to adjust
the standard errors of the estimated coefficients. Since the diagnostic tools considered
are not binary, it was not possible to consider summary receiver operating characteristic
curves. When possible, the c-statistics were pooled by considering their estimated values (point estimates) and
confidence intervals (CIs), and the ‘generic point estimates’ effect specification
option in the Comprehensive Meta-Analysis software. For a clinical prediction rule,
we assumed that a c-statistic of < 0.6 had no clinical value, 0.6 to 0.7 had limited value, 0.7 to 0.8
had modest value and > 0.8 had discrimination adequate for genuine clinical utility.[15]
Strength of Evidence
We assigned strength of evidence scores for each diagnostic tool using the approach
described in the AHRQ's Methods Guide.[11]
[16] We assessed five domains: study limitations; consistency; directness; precision;
and reporting bias, which includes publication bias, outcome reporting and analysis
reporting bias. These domains were considered qualitatively, and a summary rating
of high, moderate or low strength of evidence was assigned for each outcome after
independent assessment and discussion by two reviewers. In cases where ratings were
impossible or imprudent to make, a grade of ‘insufficient’ was assigned.
Role of the Funding Source
This topic was nominated and funded by PCORI for systematic review by an Evidence-based
Practice Center in partnership with AHRQ. A representative from AHRQ served as a Contracting
Officer's Representative (COR) and provided technical assistance during the conduct
of the full evidence report. The AHRQ COR and PCORI program officers provided comments
on draft versions of the protocol and full evidence report. PCORI and AHRQ did not
directly participate in the literature search; determination of study eligibility
criteria; data analysis or interpretation; or preparation, review or approval of the
manuscript for publication.
Results
We screened 11,274 publications and found 45 articles (25 studies) for KQ1 and 34
articles (18 studies) for KQ2 that investigated our included tools for determining
stroke or bleeding risk in patients with non-valvular AF and that met the other inclusion
criteria. We combined these newly identified studies with those included in the 2013
review, yielding a total of 83 articles (61 studies) for KQ1 and 57 articles (38 studies)
for KQ2 included in this updated review ([Fig. 1]). Complete results of the review, including long-term stroke and bleeding risk summaries,
are in the full report.
Fig. 1 Literature flow diagram. KQ, key question.
Predicting Thromboembolic Risk in Patients with AF
We considered findings from the 61 studies reporting the predictive value of the CHADS2, CHA2DS2-VASc, Framingham and ABC stroke clinical tools for thromboembolic risk ([Table 1]). Twenty-nine studies directly compared the predictive ability for thromboembolic
events of the CHADS2 risk score with other risk scores,[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45] 24 compared CHA2DS2-VASc,[18]
[19]
[20]
[21]
[23]
[24]
[26]
[37]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54] 6 compared Framingham[18]
[24]
[33]
[34]
[37]
[55] and 4 compared ABC stroke.[54]
[56]
[57]
[58] c-Statistics for predicting thromboembolic risk, when available, are reported in [Supplementary Table S3] (available in the online version). Sufficient data existed to permit meta-analysis
of studies evaluating c-statistics for the CHADS2 score using a continuous score ([Fig. 2A]) and categorical score ([Fig. 2B]), the CHA2DS2-VASc continuous score ([Fig. 2C]) and categorical score ([Fig. 2D]), the Framingham categorical score ([Fig. 2E]) and the ABC stroke categorical score ([Fig. 2F]). For both the continuous and categorical CHADS2 scores (continuous: 14 studies with 489,335 patients; categorical: 16 studies, 548,464
patients; [Table 2]), there was moderate strength of evidence that the scores provide limited prediction
of stroke events (continuous: c-statistic of 0.69; 95% CI, 0.66–0.73; categorical: c-statistic of 0.66; 95% CI, 0.63–0.69). There was also moderate strength of evidence
(16 studies, 511,481 patients) that the continuous CHA2DS2-VASc score provides limited prediction of stroke events (c-statistic of 0.66; 95% CI, 0.63–0.69). For the categorical CHA2DS2-VASc score (13 studies, 496,683 patients), there was low strength evidence of its
ability to predict stroke risk (c-statistic of 0.64; 95% CI, 0.58–0.70). Based on a meta-analysis of 6 studies (282,572
patients), we found moderate strength of evidence that the categorical Framingham
score provides limited prediction of stroke events (c-statistic of 0.63; 95% CI, 0.62–0.65). For the categorical ABC score (4 studies,
25,614 patients), we found a moderate strength of evidence of limited prediction of
stroke events (c-statistic of 0.67; 95% CI, 0.63–0.71) ([Table 2]).
Table 1
Description and interpretation of included risk scores
|
Thromboembolic risk score
|
Reference
|
Risk factors included
|
Interpretation
|
|
CHADS2
|
Gage et al, 2001[35]
|
Congestive heart failure, hypertension, age ≥75, diabetes mellitus, prior stroke/transient
ischaemic attack [2 points]
|
Low (0), moderate (1–2), high (3–6)
|
|
CHA2DS2-VASc
|
Lip et al, 2010[37]
|
Congestive heart failure/left ventricular ejection fraction ≤ 40%, hypertension, age
≥75 [2 points], diabetes mellitus, prior stroke/transient ischaemic attack/thromboembolism
[2 points], vascular disease, age 65–74, sex category female
|
Low (0), moderate (1), high (2–9)
|
|
Framingham
|
|
Advancing age, female sex, increasing systolic blood pressure, prior stroke or transient
ischaemic attack and diabetes
|
|
|
ABC
|
Hijazi et al, 2016[93]
|
Age, biomarkers (cTnI-hs and NT-proBNP), and clinical history (prior stroke/TIA)
|
Low < 1%, moderate 1–2%, high > 2%
|
|
Bleeding risk score
|
Reference
|
Risk factors included
|
Interpretation
|
|
ABC
|
Hijazi et al, 2016[93]
|
Age, biomarkers [GDF-15, cTnT-hs, and haemoglobin], and clinical history [previous
bleeding]
|
Low < 1%, medium 1–2%, high > 2%
|
|
ATRIA
|
Fang et al, 2011[76]
|
Anaemia, renal disease (CrCl < 30) (3 points each); age ≥75 (2 points); any prior
bleeding, hypertension (1 point each)
|
Low (0–3), moderate (4), high (5–10)
|
|
BRI
|
Beyth et al, 1998[109]
|
Age ≥65, GI bleed in past 2 wk, previous stroke, co-morbidities (recent MI, haematocrit < 30%,diabetes,
creatinine > 1.5), with 1 point for presence of each condition and 0 if absent
|
Low (0), moderate (1–2), high (3–4)
|
|
HAS-BLED
|
Pisters et al, 2010[9]
|
Hypertension, abnormal renal (CrCl < 50) or liver function (1 point each); stroke,
bleeding history or predisposition, labile INR (TTR < 60%), age > 65, drugs of interest/alcohol
(1 point each)
|
Low (0), moderate (1–2), high (≥3)
|
|
HEMORR2HAGES
|
Gage et al, 2006[79]
|
Liver/renal disease, ethanol abuse, malignancy, age > 75, low platelet count or function,
re-bleeding risk, uncontrolled hypertension, anaemia, genetic factors (CYP2C9), risk
of fall or stroke (1 point for each risk factor present with 2 points for previous
bleed)
|
Low (0–1), moderate (2–3), high (≥4)
|
Abbreviations: ABC, age, biomarkers, clinical history; ATRIA, Anticoagulation and
Risk Factors in Atrial Fibrillation; BRI, Bleeding Risk Index; CrCl, creatinine clearance;
cTnT-hs, high-sensitivity cardiac troponin T; GDF, growth differentiation factor-15;
GI, gastrointestinal; HAS-BLED, Hypertension, Abnormal renal/liver function, Stroke,
Bleeding history or predisposition, Labile international normalized ratio, Elderly
(> 65 years), Drugs/alcohol concomitantly; HEMORR2HAGES, Hepatic or renal disease, Ethanol abuse, Malignancy, Older (age >75 years),
Reduced platelet count or function, Re-bleeding risk (2 points), Hypertension (uncontrolled),
Anaemia, Genetic factors, Excessive fall risk, Stroke; INR, international normalized
ratio; MI, myocardial infarction; TTR, time in therapeutic range.
Table 2
Strength of evidence domains for prediction of thromboembolic risk
|
Outcome
|
Number of studies (subjects)
|
Risk of bias
|
Consistency
|
Directness
|
Precision
|
SOE and effect
(95% CI)
|
|
CHADS2 (Categorical)
|
16 (548,464)
|
Observational/ Moderate
|
Inconsistent
|
Direct
|
Precise
|
SOE = Moderate
Limited risk prediction ability (c-statistics 0.66, 95% CI, 0.63–0.69)
|
|
CHADS2 (Continuous)
|
14 (489,335)
|
Observational/ Moderate
|
Inconsistent
|
Direct
|
Precise
|
SOE = Moderate
Limited risk prediction ability (c-statistic = 0.69; 95% CI, 0.66–0.73)
|
|
CHA2DS2-VASc (Categorical)
|
13 (496,683)
|
Observational/ Moderate
|
Inconsistent
|
Direct
|
Imprecise
|
SOE = Low
Limited risk prediction ability (c-statistic = 0.64; 95% CI, 0.58–0.70)
|
|
CHA2DS2-VASc (Continuous)
|
16 (511,481)
|
Observational/ Moderate
|
Inconsistent
|
Direct
|
Precise
|
SOE = Moderate
Limited risk prediction ability (c-statistic = 0.66; 95% CI, 0.63–0.69)
|
|
Framingham (Categorical)
|
6 (282,572)
|
Observational/ Moderate
|
Consistent
|
Direct
|
Precise
|
SOE = Moderate
Limited risk prediction ability (c-statistic = 0.63; 95% CI, 0.62–0.65)
|
|
Framingham (Continuous)
|
4 (274,538)
|
Observational/ Moderate
|
Consistent
|
Direct
|
Imprecise
|
SOE = Low
Limited risk prediction ability (c-statistic ranges between 0.64 and 0.69 across studies)
|
|
ABC (Categorical)
|
4 (25,614)
|
Observational/ Moderate
|
Consistent
|
Direct
|
Imprecise
|
SOE = Moderate
Limited risk prediction ability (c-statistic = 0.67; 95% CI, 0.63–0.71)
|
|
Imaging risk tools
|
7 (4,962)
|
Observational/ Moderate
|
Inconsistent
|
Direct
|
Imprecise
|
SOE = Insufficient
|
Abbreviations: CHA2DS2-VASc, Congestive heart failure/left ventricular ejection fraction ≤ 40%, Hypertension,
Age ≥ 75 (2 points), Diabetes mellitus, prior Stroke/transient ischaemic attack/thromboembolism
(2 points), Vascular disease, Age 65–74, Sex category female; CHADS2,Congestive heart failure, Hypertension, Age ≥ 75, Diabetes mellitus, prior Stroke/transient
ischaemic attack (2 points); CI, confidence interval; SOE, strength of evidence.
Fig. 2 (A–E) Summary estimate of c-statistics for prediction ability of clinical tools for thromboembolic risk (A) CHADS2 continuous score. (B) CHADS2 categorical score. (C) CHA2DS2-VASc continuous score. (D) CHA2DS2-VASc categorical score. (E) Framingham categorical score. (F) ABC stroke categorical score.
Seven imaging studies examined specific anatomical findings and their association
with stroke risk in patients with AF.[59]
[60]
[61]
[62]
[63]
[64]
[65] Imaging studies included MRI, magnetic resonance angiography quantification of left
atrial appendage dimensions, transoesophageal echocardiography and transthoracic echocardiography.
There was insufficient evidence for the relationship between findings on echocardiography
(transthoracic) and subsequent stroke based on 7 studies (4 low risk of bias, 3 medium
risk of bias; 4,962 patients) that reported discrepant results.
We found 20 studies that evaluated either the predictive role of international normalized
ratio (INR), pattern of AF, renal impairment or other risk factors.[31]
[43]
[48]
[57]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75] There was insufficient evidence, however, for further meta-analysis of the results.
These abstracted data are in the full report.
Predicting Bleeding Risk in Patients with AF
Of the 38 studies which explored bleeding risk in patients with AF, 26 studies evaluated
various risk scores (BRI, HEMORR2HAGES, HAS-BLED, ATRIA, ABC) for estimating the outcome of major bleeding risk in
patients with AF, including patients on warfarin, aspirin and no anti-thrombotic therapy.[9]
[18]
[21]
[22]
[46]
[54]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
[90]
[91]
[92]
[93]
[94]
[95]
[96]
[97] Thirteen studies (10 low risk of bias, 2 medium risk of bias, 1 high risk of bias;
351,985 patients) compared different risk scores (BRI, HEMORR2HAGES, HAS-BLED, ATRIA, ABC) in predicting major bleeding events in AF patients on
warfarin. These studies differed markedly in population, major bleeding rates and
statistics reported for evaluating risk prediction scores for major bleeding events.
Assessment of major bleeding events based on individual risk factors was reported
by 17 studies ([Supplementary Table S4], available in the online version). Eight of these (7 low risk of bias, 1 medium
risk of bias; 322,010 patients) evaluated the risk of major bleeding in patients with
chronic kidney disease (CKD). All studies demonstrated increased risk of bleeding
in patients with CKD (moderate strength of evidence). Other risk factors abstracted
included the impact of INR, age, prior stroke, presence of heart disease, diabetes
mellitus, sex, cancer, race/ethnicity and cognitive impairment; however, the evidence
was insufficient to support findings (results in full report).
Most available studies for KQ2 included ICH within the outcome ‘major bleeding’, but
three studies presented this outcome separately. One of these studies evaluated both
HAS-BLED and HEMORR2HAGES,[18] another study evaluated both HAS-BLED and ATRIA[97] and a third study evaluated the INR.[66] The single included study comparing HAS-BLED and HEMORR2HAGES did not show a statistically significant difference between the risk scores
in prediction abilities for ICH in any patient population. Better understanding of
ICH risk prediction will be particularly important, because this represents the most
devastating variety of major bleeding event that patients on anticoagulation suffer.
The comparative risk discrimination abilities of each clinical tool was evaluated,
when data were available, for (1) major bleeding risk in AF patients on warfarin,
(2) AF patients on aspirin alone, (3) AF patients not on therapy and (4) ICH risk
in AF patients on warfarin (see [Supplementary Table S5] for c-statistics, available in the online version). For AF patients on warfarin, evidence
favoured HAS-BLED based on two studies demonstrating that it has significantly higher
prediction (by c-statistic) for major bleeding events than other scores among patients on warfarin,
but the majority of studies showed no statistically significant differences in prediction
abilities, reducing the strength of evidence (moderate; [Table 3]). For AF patients on aspirin alone, three studies (2 low risk of bias, 1 medium
risk of bias; 177,538 patients) comparing different combinations of bleeding risk
scores (BRI, HEMORR2HAGES and HAS-BLED) in predicting major bleeding events showed no statistically significant
differences (low strength of evidence). Among AF patients not on therapy, six studies
(4 low risk of bias, 2 medium risk of bias; 310,607 patients) comparing different
combinations of bleeding risk scores (BRI, HEMORR2HAGES, HAS-BLED and ATRIA) in predicting major bleeding events showed no statistically
significant differences (low strength of evidence). Evaluating ICH in AF patients
on warfarin, one study (low risk of bias; 48,599 patients) compared HEMORR2HAGES and HAS-BLED in predicting ICH. This study showed no statistically significant
difference in prediction abilities between the two scores (low strength of evidence).
Table 3
Strength of evidence domains for prediction of bleeding risk[a]
|
Outcome
|
Number of studies (subjects)
|
Risk of bias
|
Consistency
|
Directness
|
Precision
|
SOE and effect
(95% CI)
|
|
Summary c-statistic (patients on warfarin)
|
|
BRI
|
4 (11,939)
|
Observational/Moderate
|
Consistent
|
Direct
|
Precise
|
SOE = Moderate
Limited risk discrimination ability (c-statistic ranging from 0.56 to 0.65)
|
|
HEMORR2HAGES
|
10 (115,348)
|
Observational/Moderate
|
Consistent
|
Direct
|
Imprecise
|
SOE = Moderate
Limited risk discrimination ability (c-statistic ranging from 0.53 to 0.78)
|
|
HAS-BLED
|
11 (194,839)
|
Observational/Moderate
|
Consistent
|
Direct
|
Imprecise
|
SOE = Moderate
Modest risk discrimination ability (c-statistic ranging from 0.50 to 0.80)
|
|
ATRIA
|
7 (76,163)
|
Observational/Moderate
|
Inconsistent
|
Direct
|
Imprecise
|
SOE = Insufficient
|
|
ABC
|
1 (22,998)
|
Observational/Moderate
|
NA
|
Direct
|
Precise
|
SOE = Low
Limited risk discrimination (c-statistic of 0.65 in validation study)
|
|
Comparative risk discrimination abilities
|
|
Major bleeding events among patients with AF on warfarin
|
13 (351,985)
|
Observational/Moderate
|
Consistent
|
Direct
|
Imprecise
|
SOE = Moderate
Favours HAS-BLED
|
|
Intra-cranial haemorrhage among patients with AF on warfarin
|
2 (71,597)
|
Observational/Moderate
|
NA
|
Direct
|
Precise
|
SOE = Low
No evidence of a difference
|
|
Major bleeding events among patients with AF on aspirin alone
|
3 (177,538)
|
Observational/Moderate
|
Inconsistent
|
Direct
|
Imprecise
|
SOE = Low
No evidence of a difference
|
|
Major bleeding events among patients with AF not on anti-thrombotic therapy
|
6 (310,607)
|
Observational/Moderate
|
Consistent
|
Direct
|
Imprecise
|
SOE = Low
No evidence of a difference
|
Abbreviations: ABC, age, biomarkers, clinical history; AF, atrial fibrillation; ATRIA,
Anticoagulation and Risk Factors in Atrial Fibrillation; BRI, Bleeding Risk Index;
CHA2DS2-VASc, Congestive heart failure/left ventricular ejection fraction ≤ 40%, Hypertension,
Age ≥ 75 (2 points), Diabetes mellitus, prior Stroke/transient ischaemic attack/thromboembolism
(2 points), Vascular disease, Age 65–74, Sex category female; CHADS2, Congestive heart failure, Hypertension, Age ≥ 75, Diabetes mellitus, prior Stroke/transient
ischaemic attack (2 points); CI, confidence interval; HAS-BLED, Hypertension, Abnormal
renal/liver function, Stroke, Bleeding history or predisposition, Labile international
normalized ratio, Elderly (> 65 years), Drugs/alcohol concomitantly; HEMORR2HAGES, Hepatic or renal disease, Ethanol abuse, Malignancy, Older (age > 75 years),
Reduced platelet count or function, Re-bleeding risk (2 points), Hypertension (uncontrolled),
Anaemia, Genetic factors, Excessive fall risk, Stroke; INR, international normalized
ratio; KQ, key question; NA, not applicable; SOE, strength of evidence.
a
c-Statistics given are for categorical risk scores unless otherwise noted.
Discussion
Our review included studies comparing the diagnostic accuracy and impact on clinical
decision-making of available clinical tools, imaging tools and associated risk factors
for predicting thromboembolic and bleeding risk in patients with AF. For predicting
thromboembolic risk, the CHADS2, CHA2DS2-VASc and ABC scores appeared similar and had the best predictive abilities given
the available evidence, but this advantage was not substantial on an absolute basis.
Imaging risk tools, however, found conflicting results when the presence of a left
atrial thrombus was assessed, and there was insufficient evidence to support conclusions
regarding the predictive ability of the presence of a left atrial thrombus. Among
the tools for predicting risk of major bleeding and ICH, there was a suggestion that
HAS-BLED is the best score for predicting major bleeds in patients on warfarin, although
it only has modest prediction abilities. However, the majority of studies for other
patient scenarios showed no statistically significant differences in predictive accuracy
among tools.
Findings in Relation to What is Already Known
Findings in Relation to What is Already Known
ESC guidelines recommend using the CHA2DS2VASc score, and AHA guidelines recommend using the CHADS2 or CHA2DS2-VASc to categorize thromboembolic risk when making treatment decisions in patients
with AF.[98] Additionally, recent American College of Clinical Pharmacy (ACCP), Australian and
New Zealand (ANZ), and Asia Pacific Heart Rhythm Society (APHRS) guidelines endorse
using the CHA2DS2-VASc score (excluding sex in the calculation under ACCP and ANZ guidelines) to identify
low-risk patients that can be excluded from anticoagulation.[99]
[100]
[101] In the current CER, we found that of the available risk scores, the CHADS2 and CHA2DS2VASc scores are the most commonly studied and that the CHADS2, CHA2DS2-VASc and ABC risk scores appeared to be similar and to have the highest predictive
ability for stroke events. While some studies have explored the inclusion of biomarkers
in stroke risk scores such as the ABC stroke risk score, and preliminary evidence
supports the ABC score being comparable to CHADS2 and CHA2DS2-VASc, the experience with ABC is limited and more data are needed on the contribution
of these and other biomarkers to the overall risk assessment. Further, few comparisons
of the ABC score in predicting thromboembolic risk have been completed in ‘real-world’
populations, which may better clarify its predictive ability.[102]
In predicting bleeding risk, our review found limited evidence favouring the HAS-BLED
risk score based on two studies demonstrating that it has a significantly higher predictive
ability for major bleeding events than other scores among patients on warfarin. The
majority of studies, however, showed no statistically significant differences in prediction,
which reduced the strength of evidence. Recent evidence suggests that inclusion of
time to therapeutic range (TTR), included in the HAS-BLED score, might enhance the
predictive ability of other bleeding scores.[54]
[87] Bleeding risk scores are not included in the most recent AHA/ACC guideline recommendations
on AF, and they are generally not used to decide whether to prescribe an oral anticoagulant
to individual patients. However, bleeding risk scores may inform shared decision-making
discussions of the risks of stroke and bleeding incorporating patients' values and
preferences. As more data on stroke and bleeding risk scores emerge, it is possible
that improvement in the tools and methods for risk stratification of both stroke and
bleeding will be important to better individualize treatment using different oral
anticoagulants in patients with AF.
Limitations of the Evidence Base and the Comparative Effectiveness Review Process
Limitations of the Evidence Base and the Comparative Effectiveness Review Process
Comparisons across studies were difficult due to varying categorical arrangements
of stroke risk scores, inter-study differences in approach to calculating some of
the bleeding risk scores, limited comparison of bleeding risk scores across populations,
heterogeneous patient populations and the variability in treating patients with anti-platelets
and oral anticoagulants. It is known that risk scores correlate to differing event
rates based on patient setting and treatment, such as whether they are in a clinical
trial or in the outpatient setting, which further added to between study event rate
discrepancies.[103] Additionally, there was inconsistency among individual studies in reporting measures
of calibration, strength of association and diagnostic accuracy. While the nature
of a meta-analysis precludes the ability to directly account for individual study-level
bias, we were able to carefully assess for risk of bias, consistency, directness,
precision and strength of evidence as outlined by best practice guidelines in systematic
review methodology.
Further, our conclusions may be limited by the limitations in the development and
validation of risk scores. Specifically, although many of the studies use clinical
data sources to derive or validate these risk scores, some studies relied on billing
data and institutional electronic medical records to identify patients with AF and
co-morbidity information, which could under-estimate stroke risk due to lack of clinical
adjudication of events. Likewise, lack of validated results or common event definitions
for the endpoints of thromboembolism and bleeding could have under-estimated the performance
of these risk scores. Additionally, lack of standard definitions for co-morbidities
such as heart failure, diabetes mellitus and hypertension could also lead to discrepancies
across studies validating the various risk scores. Moreover, our review included both
ambulatory and hospitalized patients, which inherently introduces bias in comparing
studies and results in heterogeneity with regards to stability of covariates, concomitant
medications, stroke inducing procedures, etc.
Our review methods also had limitations. Our study was limited to English-language
publications and excluded studies conducted exclusively in Asia, Africa or the Middle
East. We also limited our analysis to studies published since 2000 as the recent literature
was considered the most relevant to today's clinical and policy uncertainties. Lastly,
we were unable to include systematic review of all available clinical risk score tools
for stroke and bleeding risk. We are aware of other tools, such as QStroke and ORBIT
scores, but our scope was focused on the scores used most frequently in clinical settings
and prioritized through the stakeholder panel and topic refinement process with PCORI.
Research Recommendations
In our analyses, we have identified several areas for recommended future research.
Given the aforementioned limitations of the currently available studies, further studies
are needed that: (1) utilize complete data; (2) use validated clinical outcomes; and
(3) compare all available risk scores using consistent and appropriate statistical
evaluations.
Despite the availability and validation of numerous tools for both stroke and bleeding
risk assessment in patients with non-valvular AF, meaningful comparisons of the tools
could not be performed in this CER. Although the 2014 AHA/ACC guideline recommends
using the CHA2DS2-VASc score for stroke risk stratification and that all patients with a CHA2DS2-VASc score of ≥ 2 be considered for oral anticoagulant therapy, the guideline acknowledged
the limitation of current risk tools, including the CHA2DS2-VASc score, to identify patients at high risk for thromboembolic risk. As a response
to this poor predictive ability in high-risk patients, recently published ACCP, ANZ
and APHRS guidelines suggest using the CHA2DS2-VASc score to identify low-risk patients in the initial step of determining whether
anti-thrombotic therapy should be offered.[99]
[100]
[101] Whether biomarkers such as brain natriuretic peptide, C-reactive protein or troponin
can enhance the CHA2DS2-VASc score and as a result be incorporated in guideline recommendations remains to
be seen.
Also, the current ACC/AHA guidelines[7] do not recommend use of bleeding risk scores, but rather focusing on modifiable
bleeding risks. Our results found moderate strength of evidence for modest risk discrimination
of the HAS-BLED score; how this modestly predictive score could potentially be utilized
in clinical treatment decisions has yet to be investigated. Preliminary data in non-clinical
trial populations show that biomarkers may not enhance risk scores' predictive ability
of bleeding risk and further research is needed to conclusively determine whether
biomarkers (e.g. brain natriuretic peptide, C-reactive protein or troponin) can enhance
these scores.[102]
[104]
With the growing prevalence of digitized medical records, there is an opportunity
to continue to evaluate and modify risk prediction tools to improve their accuracy
in predicting stroke and bleeding risk, particularly with newer anticoagulants diffusing
into clinical practice. These records might also facilitate research investigating
risk as a non-static variable, observing changes in risk factors as predictive for
stroke or bleeding events.[105]
[106] Also, newer clinical markers (e.g. MRI to assess scar), co-morbidities (i.e. renal
failure, etc.) and biomarkers should be tested and validated with or alongside current
risk tools to improve their prediction of both stroke and bleeding risks. Additionally,
more specific guidelines on how to use risk scores and apply necessary therapies,
possibly in the form of physician decision-support tools, will be important for clinical
decision-making. Efforts to create computer-based clinical decision-making supporting
tools are on-going and may represent a way to better integrate clinical risk tools
into practice.[107]
[108] Preliminary evidence of such decision support systems is discussed within the full
AHRQ report.
In addition, although we are able to identify patients at risk for stroke, many of
these patients are also at a high risk for bleeding. Thus, there is a need for a score
that could be used for decision-making about anti-thrombotic therapy in AF patients
taking into account both thromboembolic and bleeding risks. Scores that identify only
patients at risk for stroke or only those at risk for bleeding are not so helpful
since the clinical factors in these scores are usually similar and treatments which
reduce one or the other risk may increase the other for the same patient. Another
challenge is that both stroke events and bleeding events are on a spectrum of severity
and therefore predicting overall stroke might not align with outcomes that matter
most to patients. For example, some strokes may have symptoms lasting < 24 hours with
complete resolution, whereas others can cause death. It may be good for future risk
tools to account for differences in severity of outcomes. Another research need specific
to bleeding risk is a prospective comparison of the standard deviation of transformed
INR (SDTINR) and TTR to establish which variable has better predictive accuracy for major bleeding
including ICH.
Additionally, even assuming an optimal risk prediction score can be identified, further
work is needed to clarify how scores should be used prospectively in clinical practice.
Clinical risk scores must take into account the balance between simplicity and practicality
versus accurate prediction, especially in a high-capacity clinical environment. While
clinical risk scores are necessarily reductionist and cannot feasibly consider all
patient parameters, our results here show moderate predictive ability of risk scores
that can be calculated relatively easily from patient history and demographics. Future
research might explore this trade-off between ease of implementation and increasing
the predictive value of clinical risk scores with more difficult-to-obtain parameters
such as biomarkers.
Conclusion
Overall, we found that CHADS2, CHA2DS2-VASc and ABC stroke scores have the best prediction for stroke events in patients
with AF among the risk scores we reviewed, whereas HAS-BLED provides the best prediction
for bleeding risk. Imaging tools require further evidence in regard to their appropriate
use in clinical decision-making. Additionally, simple clinical decision tools are
needed that incorporate both stroke risk and bleeding risk to assist providers treating
patients with AF. Additional work will be required to develop risk tools for patients
to discriminate those individuals with AF where the bleeding risk may be high enough
to warrant more intensive follow-up and monitoring. These tools could be embedded
into electronic medical record systems for point-of-care decision-making, developed
into applications for smartphones and tablets or be delivered via web-based interfaces.
Additional evidence of the use of these stroke and bleeding risk scores (and clinical
decision tools which balance these risks) among patients on therapy is also required.
What is known about this topic?
What does this paper adds?
-
CHADS2, CHA2DS2-VASc, and ABC risk scores have the best evidence to support prediction of stroke
events.
-
HAS-BLED has the best evidence to support prediction of bleeding risk.
-
Imaging tools for stroke prediction require further evidence.
