A comparison of risk stratification schemes for stroke in 79 884 atrial fibrillation patients in general practice
Gregory Y. H. Lip, University of Birmingham Centre for Cardiovascular Sciences, City Hospital, Birmingham, UK.
Tel.: +44 121 50 75080; fax: +44 121 554 4083.
Summary. Background: Anticoagulation management of patients with atrial fibrillation (AF) should be tailored individually on the basis of ischemic stroke risk. The objective of this study was to compare the predictive ability of 15 published stratification schemes for stroke risk in actual clinical practice in the UK. Methods: AF patients aged ≥ 18 years in the General Practice Research Database, which contains computerized medical records, were included. The c-statistic was estimated to determine the predictive ability for stroke for each scheme. Outcomes included stroke, hospitalizations for stroke, and death resulting from stroke (as recorded on death certificates). Results: The study cohort included 79 844 AF patients followed for an average of 4 years (average of 2.4 years up to the start of warfarin therapy). All risk schemes had modest discriminatory ability in AF patients, with c-statistics for predicting events ranging from 0.55 to 0.69 for strokes recorded by the general practitioner or in hospital, from 0.56 to 0.69 for stroke hospitalizations, and from 0.56 to 0.78 for death resulting from stroke as reported on death certificates. The proportion of patients assigned to individual risk categories varied widely across the schemes, with the proportion categorized as moderate risk ranging from 12.7% (CHA2DS2-VASc) to 61.5% (modified CHADS2). Low-risk subjects were truly low risk (with annual stroke events < 0.5%) with the modified CHADS2, National Institute for Health and Clinical Excellence and CHA2DS2-VASc schemes. Conclusion: Current published risk schemes have modest predictive value for stroke. A new scheme (CHA2DS2-VASc) may discriminate those at truly low risk and minimize classification of subjects as intermediate/moderate risk. This approach would simplify our approach to stroke risk stratification and improve decision-making for thromboprophylaxis in patients with AF.
Atrial fibrillation (AF) is the most common disorder of the cardiac rhythm. This condition carries an increased risk of arterial thromboembolism and ischemic stroke resulting from embolization of thrombi that form within the left atrium of the heart. Recent practice guidelines for the management of patients with AF recommend that the choice of anticoagulation treatment should be tailored individually on the basis of age, comorbidities, and contraindications [1–3]. Guidelines recommend that patients at higher risk of stroke should receive dose-adjusted warfarin, and those at lower risk should be treated with aspirin [1–3]. For those who fall into the intermediate/moderate-risk category, current guidelines recommend ‘either aspirin or warfarin’ [1–4], which may cause uncertainty for clinicians regarding the best treatment option, especially if a large proportion of patients are categorized as moderate risk, with some clinicians even using the guidelines as an excuse not to prescribe warfarin, as aspirin is ‘allowed’ by the guidelines .
A large number of risk stratification schemes have been developed to quantify the risk of stroke in AF [2–12]. One widely used scheme is the CHADS2 score (an acronym for Congestive heart failure, Hypertension, Age > 75 years, Diabetes mellitus and prior Stroke or transient ischemic attack), which estimates the long-term probability of stroke on the basis of risk scores . There have been a large number of studies proposing risk stratification schemes for the prediction of stroke in AF patients, and several studies have compared and tested these risk stratification schemes [13–17]; however, few have used data from a large, contemporary cohort representing ‘real-world’ clinical practice. New schemes, such as the CHA2DS2-VASc score , also require formal testing in several independent cohorts.
The objective of this study was to evaluate and compare the value of contemporary stroke risk stratification schemes in predicting stroke in AF patients in ‘real-world’ routine contemporary UK clinical practice.
Information for the study was obtained from the General Practice Research Database (GPRD), which comprises the computerized medical records of all general practitioners (GPs) in the UK. GPs play a key role in the UK healthcare system, as they are responsible for primary healthcare and specialist referrals. The data recorded in the GPRD include demographic information, prescription details, clinical events, preventive care provided, specialist referrals, hospital admissions, and major outcomes. Specialists and hospitals in the UK are required to inform the GP about the patient’s medical care, and summary information from specialist care is typically entered into the GPRD by the GP. The GPRD currently includes data on over 10 million patients . GPRD patients in English practices are now linked individually and anonymously to the national registry of hospital admission (Hospital Episode Statistics [HES]) and to the death certificates (as collected by the Office of National Statistics). For each hospitalized patient, the hospital charts are reviewed, and dates of admission and discharge and main diagnoses are extracted, coded by coding staff, and collated nationally. The death certificates list the date and causes of death. HES data were available from 2001 to 2007, and death certificate data from 2001 to 2008 (for about 200 practices and 40% of the study population). A previous GPRD study reported a high level of validity in the recording of AF by GPs . The GPs confirmed the accuracy of the GPRD record of AF in 95.9% of a sample of 1606 AF cases .
The study population included patients aged ≥ 18 years with a documented record of AF. Patients with rheumatic valve disease were excluded. Patients were followed from the first record of AF after 1 January 1990 (i.e. index date) up to the latest data collection at the time of study (December 2008). The study population included incident and prevalent AF patients. Medical data in the GPRD are coded by the contributing GP during routine medical practice, using READ terminology; unfortunately, there is no READ term for chronic AF, and there are only two terms for paroxysmal AF. Most terms in the GPRD do not specify the type of AF.
The total period of follow-up was from the index date up to the date of censoring (i.e. latest GPRD data collection, patient’s transfer out of the practice, or patient’s death, whichever date came first). In some of the analyses, the follow-up was also censored at the start of warfarin treatment or first International Normalized Ratio (INR) measurement. The outcomes of interest were stroke as recorded in the GPRD, hospitalization for stroke as recorded in HES (ICD-10 codes I61, I62, I63, and I64), and mortality resulting from stroke as recorded on death certificates (ICD-10 codes I61, I62, I63, and I64). Cases of stroke as identified in HES were classified into ischemic and hemorrhagic strokes; this classification was not feasible within the GPRD, as most stroke cases are recorded using non-specific READ codes .
Risk stratification schemes
Appendix 1 shows the risk factors used in the 15 schemes and the classification within each scheme into low, moderate and high risk for stroke. The following risk factors are not recorded in the GPRD, and were not considered in the estimation of risk: left ventricular fractional shortening < 25% on echocardiography , enlarged cardiothoracic ratio on chest roentgenogram , visible ischemic lesion on computed tomography scan , and impaired left ventricular function on echocardiography [2,3,21]. Also, blood pressure was not available for all AF patients, as GPs do not routinely screen all patients for these risk factors (for clinical reasons or because of the patient not visiting the practice) [6,7,9,10].
Predictive value of risk schemes
Logistic regression was used to evaluate the predictive value of the various risk schemes in discriminating between patients who developed the outcome of interest and those who did not. The c-statistic, which represents the area under the receiver operating characteristic curve, was estimated to determine the predictive ability for stroke of each scheme. A c-statistic of 1.0 indicates perfect ability to predict the risk of stroke, and a score of 0.5 indicates no predictive value. Each scheme was divided into low, moderate and high predicted risk categories, except for CHADS2, modified CHADS2, Framingham, and CHA2DS2-VASc, where actual scores were also used to calculate the c-statistics. With the use of non-parametric bias-corrected bootstrapping, 95% confidence intervals (CIs) were estimated [22–24]. Because patients may have different durations of follow-up, the analyses were weighted by the duration of follow-up . In this analysis, AF patients were censored at the start of warfarin treatment or at the first INR measurement, whichever came first. Aspirin use was not considered in our analysis, as no significant effects on stroke prevention in AF were observed with aspirin (in contrast to warfarin) in a recent GPRD study .
Individual risk of stroke
The risk scheme with the largest predictive value (i.e. c-statistic) were evaluated for the extent of misclassification (i.e. the proportion of patients who are classified into the high-risk group but who have a stroke incidence below a threshold, and the proportion of patients who are classified into moderate/low-risk groups but who have an incidence above a threshold). The long-term cumulative incidence for stroke was estimated individually for each patient on the basis of age, sex, and clinical risk factors for stroke (as available in the GPRD), using Cox proportional hazards regression models. In addition to the risk factors measured in any of the different risk schemes, the following risk factors for stroke were included into the Cox regression models: smoking history, alcohol consumption history, body mass index, LDL levels (if recorded), C-reactive protein (if recorded), and socio-economic status at the postcode of the practice location. Age was included, using 10-year age categories. Risk factors were measured in a time-dependent manner, and could vary over the follow-up period. For example, the follow-up period of a patient who developed incident heart failure after the index date was classified into a period of time without heart failure (i.e. prior to the record of heart failure) and a period of time with heart failure (i.e. after the record). For the LDL, C-reactive protein, and systolic blood pressure, the total follow-up period was also divided into periods of time with and without measurements. The period of time with a measurement was taken as the time period from the date of the measurement up to 12 months after or up to the next measurement, whichever date came first. The follow-up period was also divided into time with and without warfarin use. The time on warfarin was taken as the time from the first warfarin prescription or INR measurement up to 6 months after the last warfarin prescription or INR measurement.
Cox proportional hazards models were used to estimate the long-term risk of stroke (during the time without warfarin use). For each set of patient characteristics, the Cox model allows calculation of an individual’s probability of stroke (i.e. survivor function). The first regression model included age, sex, and all risk factors, and forward selection was then used for the risk factors with a statistically significant association with stroke (using a significance level of 0.05). Possible statistical interactions between the remaining risk factors and age and sex were then investigated, and interaction terms were included if sufficiently strong (i.e. if statistically significant after the Bonferroni adjustment of significance level). Only the interaction between age and history of stroke/transient ischemic attack (TIA) was strong enough to merit inclusion in the final Cox model. The β-coefficients of the final Cox model (the exponentials of which constitute the relative risks [RRs]) were converted into integer risk scores. Because of the time-dependent exposure variables, the risk score of a patient was averaged over the total follow-up period. The 5-year risk of stroke was then estimated using these scores. This score represents the probability of stroke, conditional on patient survival. We compared the observed 5-year probability of stroke (based on the Kaplan–Meier estimate) with the probability estimated by the Cox model. This was done by dividing the study population into 10 groups, based on the estimated probability of stroke. The observed and estimated probabilities for stroke were then compared.
The study population included 79 844 AF patients. As shown in Table 1, the mean age of AF patients was 73.3 years (standard deviation 12.5 years) and 49.7% were women. The median duration of follow-up was 2.9 years. A high CHADS2 score (≥ 3) was present in 20.1%. AF patients commonly had a medical history of congestive heart failure, diabetes mellitus, hypertension, stroke, and TIA (Table 1).
Table 1. Baseline characteristics of the 79 884 atrial fibrillation (AF) patients
| Percentiles 5, 95||0.2, 11.0|
|Follow-up (years)† (up to start of warfarin treatment)|
| Percentiles 5, 95||0.1, 9.1|
| Mean (SD)||73.3 (12.5)|
| 18–39||1523 (1.9)|
| 40–64||15 051 (18.9)|
| 65–79||35 807 (44.8)|
| ≥ 80||27 463 (34.4)|
|Sex, n (%)|
| Men||40 140 (50.3)|
| Women||39 704 (49.7)|
|Smoking, n (%)|
| Non-smoker||36 282 (45.4)|
| Ex-smoker||22 397 (28.1)|
| Smoker||9954 (12.5)|
| Not recorded||11 211 (14.0)|
|Body mass index, n (%)|
| < 20||4515 (5.7)|
| 20–25.9||27 313 (34.2)|
| ≥ 26||33 817 (42.4)|
| Not recorded||14 199 (17.8)|
|CHADS2 score, n (%)|
| 0||18 168 (22.8)|
| 1 or 2||45 645 (57.2)|
| ≥ 3||16 031 (20.1)|
|Medical history, n (%)|
| Congestive heart failure||23 307 (29.2)|
| Diabetes mellitus||13 269 (16.6)|
| Hypertension||40 111 (50.2)|
| Stroke or TIA||14 406 (18.0)|
|Anticoagulation history, n (%)|
| Warfarin prescription or INR measurement||16 060 (20.1)|
All risk schemes had modest discriminatory ability (Table 2), with c-statistics ranging from 0.55 to 0.69 for strokes recorded by the GP or in hospital, from 0.56 to 0.69 for stroke hospitalizations, and from 0.56 to 0.78 for death resulting from stroke as reported on the death certificates (Table 2). Higher c-statistics were found with CHADS2 (0.66 [95% CI 0.64–0.68] for strokes recorded by the GP or in hospital), modified CHADS2 (0.69 [95% CI 0.67–0.71]), Framingham (0.65 [95% CI 0.63–0.68]) and CHA2DS2-VASc (0.67 [95% CI 0.65–0.69]) than with the other schemes. The results for the c-statistics were comparable when follow-up time during or after warfarin treatment was included (data not shown).
Table 2. Discriminatory ability of different risk schemes in the prediction of stroke with patients censored at the start of warfarin treatment or first International Normalized Ratio (INR) measurement. For each risk scheme, c-statistics (95%CI) are shown
|AFI 1994||0.60 (0.58–0.61)||0.59 (0.58–0.60)||0.61 (0.58–0.62)||0.61 (0.57–0.63)||0.61 (0.58–0.63)|
|AFI 1998||0.61 (0.60–0.62)||0.59 (0.58–0.60)||0.62 (0.61–0.63)||0.61 (0.60–0.63)||0.63 (0.61–0.64)|
|ACCP 2001||0.62 (0.60–0.63)||0.60 (0.59–0.60)||0.62 (0.61–0.63)||0.62 (0.60–0.63)||0.63 (0.62–0.64)|
|ACCP 2004||0.61 (0.60–0.62)||0.59 (0.58–0.60)||0.62 (0.61–0.63)||0.62 (0.60–0.63)||0.62 (0.61–0.63)|
|NICE 2006 (guideline 36)||0.64 (0.62–0.65)||0.61 (0.60–0.62)||0.64 (0.63–0.66)||0.64 (0.62–0.66)||0.68 (0.66–0.69)|
|ACC/AHA/ESC 2006||0.64 (0.62–0.66)||0.60 (0.59–0.62)||0.65 (0.63–0.67)||0.64 (0.62–0.67)||0.68 (0.65–0.70)|
|ACCP 2008||0.64 (0.62–0.65)||0.60 (0.59–0.62)||0.65 (0.62–0.67)||0.64 (0.62–0.67)||0.68 (0.65–0.70)|
|CHA2DS2-VASc (three categories)||0.60 (0.59–0.61)||0.59 (0.58–0.60)||0.60 (0.59–0.61)||0.60 (0.59–0.61)||0.61 (0.61–0.62)|
|CHA2DS2-VASc (risk score*)||0.67 (0.65–0.69)||0.64 (0.62–0.65)||0.68 (0.66–0.70)||0.67 (0.64–0.69)||0.74 (0.71–0.76)|
|CHADS2 2001 (3 categories)||0.65 (0.63–0.67)||0.61 (0.60–0.62)||0.66 (0.64–0.68)||0.65 (0.62–0.67)||0.70 (0.68–0.73)|
|CHADS2 2001 (risk score*)||0.66 (0.64–0.68)||0.61 (0.60–0.63)||0.66 (0.64–0.69)||0.66 (0.63–0.69)||0.72 (0.69–0.74)|
|Modified CHADS2 2008 (three categories)||0.63 (0.61–0.65)||0.59 (0.57–0.60)||0.63 (0.61–0.66)||0.60 (0.57–0.64)||0.71 (0.68–0.74)|
|Modified CHADS2 2008 (risk score*)||0.69 (0.67–0.71)||0.64 (0.63–0.65)||0.69 (0.67–0.71)||0.66 (0.63–0.69)||0.78 (0.75–0.80)|
|SPAF 1995||0.63 (0.61–0.65)||0.61 (0.59–0.62)||0.64 (0.61–0.66)||0.62 (0.59–0.65)||0.68 (0.66–0.71)|
|Hart 1999||0.62 (0.60–0.64)||0.60 (0.59–0.61)||0.62 (0.60–0.65)||0.61 (0.58–0.64)||0.67 (0.65–0.70)|
|van Walraven 2002||0.55 (0.54–0.58)||0.55 (0.54–0.57)||0.56 (0.54–0.58)||0.57 (0.54–0.60)||0.56 (0.53–0.59)|
|van Latum 1995||0.57 (0.55–0.59)||0.56 (0.55–0.57)||0.57 (0.54–0.60)||0.57 (0.54–0.61)||0.60 (0.57–0.62)|
|Framingham 2003 (three categories)||0.62 (0.60–0.64)||0.61 (0.60–0.62)||0.64 (0.61–0.66)||0.61 (0.58–0.64)||0.71 (0.68–0.73)|
|Framingham 2003 (risk score)||0.65 (0.63–0.68)||0.62 (0.61–0.64)||0.67 (0.64–0.69)||0.64 (0.60–0.67)||0.75 (0.72–0.77)|
Table 3 shows the results of the forward regression model listing the variables that were statistically significantly associated with stroke (as recorded by GPs or in the hospitalization registry). Age was found to be a strong predictor of stroke risk, with the adjusted RR being 2.22 (95% CI 1.78–2.76) for those aged ≥ 80 years as compared with those aged 60–69 years. The RR of stroke was 2.86 (95% CI 2.53–3.22) in AF patients with a history of stroke/TIA. A history of heart failure and diabetes mellitus were also significant risk factors for stroke (Table 3). Although not recorded in all patients, systolic blood pressure and C-reactive protein levels were also associated with the stroke risk (RR of 2.74 [95% CI 1.21–6.19] with blood pressure of 140–159 mmHg; RR of 2.11 [95% CI 1.09–4.09] with C-reactive protein ≥ 50 mg L−1 in AF patients). Ischemic heart disease was not associated with stroke in AF patients.
Table 3. Relative risk (RR) of stroke (as recorded by general practitioners or in the hospitalization registry) with different risk factors*
| < 50||5||0.14 (0.06–0.34)|
| 50–59||23||0.44 (0.28–0.69)|
| 70–79||289||1.42 (1.12–1.78)|
| ≥ 80||820||2.22 (1.78–2.76)|
| Male||477||0.95 (0.84–1.06)|
|History of TIA/stroke|
| Yes||419||2.86 (2.53–3.22)|
| Yes||364||1.26 (1.11–1.42)|
|Systolic blood pressure (mmHg)|
| < 120||7||Reference|
| 120–139||30||1.98 (0.87–4.52)|
| 140–159||33||2.74 (1.21–6.19)|
| 160–179||9||1.49 (0.55–4.00)|
| ≥ 180||4||4.28 (1.25–14.64)|
| Not recorded||1150||1.54 (0.73–3.23)|
| Yes||199||1.33 (1.14–1.55)|
|C-reactive protein (mg L−1)|
| < 4.0||21||Reference|
| 4.0–7.9||29||1.33 (0.76–2.33)|
| 8.0–17.9||25||1.27 (0.71–2.27)|
| 18.0–49.9||15||1.28 (0.66–2.48)|
| ≥ 50.0||15||2.11 (1.09–4.09)|
| Not recorded||1128||1.01 (0.65–1.55)|
|Body mass index (kg m−2)|
| < 20||110||1.31 (1.06–1.61)|
| ≥ 26||418||0.91 (0.80–1.04)|
| Not recorded||236||1.44 (1.23–1.69)|
|Heart valve disorder|
| Yes||16||1.65 (1.01–2.71)|
Risk categorization and the distribution of individualized incidence of stroke with various schemes are presented in Table 4. The proportion of patients categorized as ‘high risk’ varied substantially, from 10.3% (Framingham) to 78.9% (CHA2DS2-VASc). The proportion categorized as ‘moderate risk’ also varied from 12.7% (CHA2DS2-VASc) to 61.5% (modified CHADS2). The frequency of 5-year incidence of stroke (estimated for each patient individually) by risk category of each scheme is shown in Table 4. The modified CHADS2 and CHA2DS2-VASc schemes had no or very few (0.4%) patients with a 5-year incidence of stroke ≥ 10% in their respective ‘low-risk’ categories. It is of note that 24.4% of the patients classified as ‘low risk’ with the Framingham risk scheme had an estimated 5-year incidence of stroke ≥ 10.0%. There was large heterogeneity in 5-year incidence of stroke.
Table 4. Distribution of atrial fibrillation (AF) patients* and stroke cases (as recorded by general practitioner or in hospital registry) in low-risk, moderate-risk and high-risk groups of different risk schemes and individualized 5-year incidence of stroke in AF patients (based on a Cox regression model including the risk factors listed in Table 3)
| Low||1793 (8.6)||0.5||1323 (73.8)||462 (25.8)||4 (0.2)||4 (0.2)||0 (0)|
| Moderate||2649 (12.7)||1.1||1105 (41.7)||1382 (52.2)||121 (4.6)||28 (1.1)||13 (0.5)|
| High||16 395 (78.7)||4.6||425 (2.6)||4873 (29.7)||5104 (31.1)||3194 (19.5)||2799 (17.1)|
| Low||5072 (24.3)||1.0||2137 (42.1)||2504 (49.4)||332 (6.5)||69 (1.4)||30 (0.6)|
| Moderate||11 962 (57.4)||3.7||703 (5.9)||4006 (33.5)||4271 (35.7)||2260 (18.9)||722 (6.0)|
| High||3803 (18.3)||8.3||13 (0.3)||207 (5.4)||626 (16.5)||897 (23.6)||2060 (54.2)|
|Modified CHADS2 (Reitbrock et al.)|
| Low||327 (1.6)||0||327 (100.0)||0 (0)||0 (0)||0 (0)||0 (0)|
| Moderate||12 820 (61.5)||2.0||2517 (19.6)||6628 (51.7)||2716 (21.2)||725 (5.7)||234 (1.8)|
| High||7690 (36.9)||6.7||9 (0.1)||89 (1.2)||2513 (32.7)||2501 (32.5)||2578 (33.5)|
|Framingham (Wang 2003)|
| Low||8836 (42.4)||1.8||2776 (31.4)||3903 (44.2)||1577 (17.8)||460 (5.2)||120 (1.4)|
| Moderate||9860 (47.3)||4.3||76 (0.8)||2759 (28.0)||3396 (34.4)||2320 (23.5)||1309 (13.3)|
| High||2141 (10.3)||9.5||1 (0.05)||55 (2.6)||256 (12.0)||446 (20.8)||1383 (64.6)|
With a study population of 79 844 AF patients, this represents the largest cohort validating the published stroke risk stratification schemes for AF. We found that all published risk schemes have modest predictive value for stroke. Nonetheless, three schemes – the modified CHADS2 (by Rietbrock et al. ), National Institute for Health and Clinical Excellence (NICE) , and CHA2DS2-VASc (Birmingham 2009)  – all identified ‘low-risk’ patients who were truly low risk, with stroke events being < 0.5% per year.
We confirm observations from previous validation studies [13–17] that stroke risk schemes have only modest predictive value for stroke (i.e. ‘high-risk’ subjects), with c-statistics of approximately 0.6. In the present analysis, higher c-statistics were found with CHADS2 (0.66), modified CHADS2 (0.69), Framingham (0.65) and CHA2DS2-VASc (0.67) than with the other schemes. We also observe the suboptimal application of thromboprophylaxis in UK clinical practice , necessitating greater emphasis on management guidelines .
Although it would be difficult to make comparisons of c-statistics between different published studies [12–17], given the diverse study populations and validation cohorts, the present analysis represents the largest validation cohort in a large contemporary ‘real-world’ UK GP dataset, and does not have the limitations of clinical trial cohorts  or previously anticoagulated patients . Indeed, our data confirm and also substantially extend the ‘real-world cohort’ information provided from the ATRIA study  and the Euro Heart survey . For example, the paper by Fang et al.  did not include contemporary stroke risk stratification schemes, such as ACCP8 and NICE.
As expected, our dataset showed that prior stroke, age, a history of heart failure, systolic blood pressure and diabetes mellitus were risk factors for stroke in our population. The systematic review of stroke risk factors by the Stroke Risk in AF Working Group  reported that prior stroke/TIA, age, diabetes mellitus and female gender were risk factors for stroke – whereas coronary artery disease and heart failure were not consistent risk factors. The systematic review as part of the UK NICE guidelines found that history of stroke or TIA, increasing age, hypertension and structural heart disease (left ventricular dysfunction or hypertrophy) were good predictors of stroke risk in AF patients, whereas diabetes mellitus, gender and other patient characteristics were less consistent risk factors . These differences may reflect the populations studied, given that the Stroke Risk in AF Working Group  largely based their analysis on trial cohorts, and uncertainty remains over some definitions of risk factors. For example, a ‘history of heart failure’ may not necessarily mean systolic dysfunction, which is more consistently related to increased stroke risk [21,28]. Also, female gender is increasingly recognized as a risk factor for stroke, in anticoagulated and non-anticoagulated populations . Age is clearly an increasing risk factor for stroke, but rather than it being a categorized variable, stroke risk increases from age 65 years upwards, and risk schemes need to consider that an AF patient aged 65–74 years would be at increased stroke risk, with the risk being even higher for those aged ≥ 75 years . Well-controlled hypertension may also represent less of a risk than uncontrolled hypertension . In the present study, ischemic heart disease was not associated with stroke in AF patients, although prior myocardial infarction [32,33], peripheral artery disease  and complex aortic plaque  are all strong predictors of stroke and thromboembolism in AF patients. Ultimately, the management of stroke risk in AF demands a multifactorial holistic approach, with all cardiovascular risk factors being adequately identified and treated .
Consistent with the Stroke Risk in AF Working Group findings , the proportion of patients categorized as high risk varied substantially in the present study. Of more concern is that the proportion categorized as moderate risk also varied substantially – with current guidelines, the moderate-risk category can be treated with warfarin or aspirin. As previously mentioned, this may cause clinical uncertainty if a large proportion of AF subjects are classified as moderate risk, and clinicians may use the guidelines as an excuse not to prescribe warfarin. In the present analysis, the scheme that placed the largest proportion into the moderate-risk category was the modified CHADS2 scheme by Rietbrock et al. .
In the present analysis, those categorized as low risk were truly low risk, with stroke events being < 0.5% per year with the modified CHADS2, NICE 2006, and CHA2DS2-VASc schemes. The modified CHADS2 and CHA2DS2-VASc schemes had a 5-year incidence of stroke of ≥ 10% in their respective low-risk categories of zero and 0.4%. It is of note that > 24% of the patients classified as low risk with the Framingham risk scheme had an estimated 5-year incidence of stroke ≥ 10.0%. Also, many AF patients who were placed into the moderate-risk category were found to have a substantial risk of stroke, depending on the schemes tested.
Aspirin may have limited benefits for stroke prevention in AF, even among low-risk patients . The putative benefits for aspirin in AF are largely driven by one single positive trial, the 1st Stroke Prevention in AF trial (SPAF-1) . The latter trial claimed a 42% RR reduction (RRR) for stroke with aspirin 325 mg daily as compared with placebo, despite major internal inconsistencies within the trial for the aspirin benefit between the warfarin-eligible (RRR 94%) and warfarin-ineligible (RRR 8%, P = not significant) arms; also, aspirin had less effect in those aged > 75 years in SPAF-1, and nor did it prevent severe or recurrent strokes . Recent analyses [38,39] have shown that even AF patients with a CHADS2 score of 1 do benefit from oral anticoagulation over aspirin. A recent overview by Lip and Halperin  has appealed for a paradigm shift, whereby simple stroke risk schemes should be better at identification of truly low-risk subjects for who no antithrombotic therapy can be recommended. All other AF subjects (i.e. those who are not truly low risk) can be prescribed oral anticoagulants, especially given the existence of new agents, such as dabigatran , that overcome the limitations of warfarin.
We stress that the CHA2DS2VASc score was not meant to replace CHADS2 but to complement it. Thus, where the patient has a CHADS2 score ≥ 2, the decision is clearly to anticoagulate . Where patients have a CHADS2 score of 0–1, or where a more comprehensive stroke risk assessment is needed, additional stroke risk modifiers need to be considered, and the CHA2DS2VASc score may help with this. Thus, truly low-risk subjects (i.e. CHA2DS2-VASC score of 0) do not need antithrombotic therapy, whereas all others (CHA2DS2VASc score ≥ 1) can be managed with oral anticoagulation [40,42]. Such a simplified yes/no approach would help in our management of such patients.
This study is limited by its dependence on information recorded in the GPRD dataset, and residual confounding cannot be excluded. Also, risk factors are dynamic in nature, and given the type of population (i.e. AF patients with multiple comorbidities, etc.), the risk of stroke does not remain static over time, and our analysis based on baseline risk factors would not necessarily account for emergent risk factors over the follow-up period. In our analysis, we calculated c-statistics on the baseline characteristics, as it is far more complex to estimate c-statistics in time-dependent models, and a clinician may assess stroke risk only periodically. Also, many risk factors for stroke are also risk factors for bleeding, and ultimately a balance between stroke and bleeding risk assessment is needed. However, current AF guidelines do not recommend a formal bleeding risk assessment score [1–3], given the lack of a simple, user-friendly score to aid everyday clinical decision-making. We also did not consider the impact of radiofrequency or surgical ablation in our analysis, as these interventions have not been established as altering stroke risk, especially in the presence of risk factors. Another limitation was that blood pressure and laboratory values (C-reactive protein) were not measured consistently for all patients, as these are only measured if the patient visits the practice and there is a clinical reason to do so. Finally, we did not distinguish between paroxysmal, persistent and permanent AF, owing to coding limitations, but stroke risk is managed in the presence of risk factors, irrespective of clinical subtype.
In conclusion, the published risk schemes have modest predictive value for stroke. A new scheme (CHA2DS2-VASc) may discriminate those at truly low risk, and minimize classification of subjects as intermediate/moderate risk. This would simplify our approach to stroke risk stratification and improve decision-making for thromboprophylaxis in patients with AF.
The views expressed in this article are those of the authors, and do not reflect the official policy or position of the Medicines and Healthcare products Regulatory Agency, UK. T. van Staa had full access to all of the data in the study, and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Disclosure of Conflict of Interests
The GPRD is owned by the UK Department of Health, and operates within the Medicines and Healthcare Products Regulatory Agency (MHRA). The GPRD is funded by the MHRA, the Medical Research Council, various universities, contract research organizations, and pharmaceutical companies. The department of Pharmacoepidemiology & Pharmacotherapy, Utrecht Institute for Pharmaceutical Sciences, has received unrestricted funding for pharmacoepidemiological research from GlaxoSmithKline, Novo Nordisk, the private–public-funded Top Institute Pharma (http://www.tipharma.nl; includes co-funding from universities, government, and industry), the Dutch Medicines Evaluation Board, and the Dutch Ministry of Health.
The authors state that they have no conflict of interest.
Table Appendix1. Risk factors used in the classification into low-risk, moderate-risk and high-risk categories with each of the different schemes
|AFI 1994 ||Age < 65 years + no risk factors||Age 65–75 years + no risk factors||Hypertension (any age)Diabetes (any age)History of stroke/TIA (any age)Age ≥ 75 years|
|SPAF 1995 ||No risk factors||Hypertension||Prior ischemiaWomen > 75 yearsRecent congestive heart failure or left ventricular fractional shortening < 25% on echocardiographySystolic blood pressure > 160 mmHg|
|Van Latum 1995 ||No risk factors||1–2 risk factors||≥ 3 risk factors (risk factors: history of stroke/TIA, ischemic heart disease, enlarged cardiothoracic ratio on chest roentgenogram, systolic blood pressure > 160 mmHg, AF > 1 year, visible ischemic lesion on CT scan)|
|AFI 1998 ||Age < 65 years + no risk factors||Age 65–75 years + no risk factors||Moderate/severe left ventricular dysfunction (echocardiography)Age < 65 years + ≥ 1 risk factorsAge 65–75 years + ≥ 1 risk factorsAge ≥ 75 years (risk factors: history of stroke/TIA, hypertension, diabetes)|
|Hart 1998 [w5]||None||Hypertension + age < 75 yearsDiabetes||History of stroke/TIAWomen aged > 75 yearsMen aged > 75 years + hypertensionSystolic blood pressure > 160 mmHg|
|CHADS2 2001* ||Score 0||Score 1 or 2||Score 3–6|
|ACCP 2001 ||None||Age 65–75 yearsDiabetesCoronary artery disease||Age > 75 yearsHistory of ischemic stroke/TIASystemic embolismHypertensionPoor left ventricular systolic functionRheumatic valve diseaseProsthetic valve disease|
|Framingham† ||Score 0–7||Score 8–15||Score 16–31|
|van Walraven ||No risk factors||Moderate or high: (any of the below)History of stroke/TIATreated hypertension or systolic blood pressure > 140 mmHgPrevious MI/anginaDiabetes|
|ACCP 2004 ||None||Age 65–75 years + no risk factors||Age > 75 yearsHistory of ischemic stroke/TIASystemic embolismPoor left ventricular function or heart failureHypertensionDiabetes|
|NICE 2006 ||Age < 65 years + no risk factors||Age ≥ 65 years + no risk factorsAge < 75 years + hypertensionAge < 75 years + diabetesAge< 75 years + vascular disease||History of ischemic stroke/TIAAge ≥ 75 years + hypertension,Age ≥ 75 years + diabetesAge ≥ 75 years + vascular diseaseValve disease, heart failure or impaired left ventricular function (echocardiography)|
|ACC/AHA/ESC 2006 ||None||Age ≥ 75 yearsHypertensionHeart failureDiabetesLeft ventricular function 35% or less||History of stroke/TIA> 1 moderate risk factorMitral stenosisProsthetic heart valve|
|Modified CHADS2‡ ||Score 0||Score 1–5||Score 6–14|
|ACCP 2008 ||None||One of the following: age ≥ 75 years, hypertension, poor left ventricular function or heart failure, diabetes||One of the following: history of ischemic stroke/TIA systemic embolismTwo or more of the following: age ≥ 75 years, hypertension, poor left ventricular function or heart failure, diabetes|
|CHA2DS2-VASc (Birmingham 2009) [17,42]||None||One of the following: heart failure or moderate/severe left ventricular dysfunction (< 40%), hypertension, diabetes, age 65–74 years, women, vascular disease||One of the following:age ≥ 75 years, history of ischemic stroke/TIA, systemic embolism, mitral stenosis, prosthetic heart valveTwo or more of the following: heart failure or moderate/severe left ventricular dysfunction (< 40%), hypertension, diabetes, age 65–74 years, women, vascular disease|