[Skip to Navigation]
Sign In
Table 1. Comparison of National Registry of Atrial Fibrillation (NRAF) Participants and Clinical Trial Participants*
Image description not available.
Table 2. Risk of Stroke in National Registry of Atrial Fibrillation (NRAF) Participants, Stratified by CHADS2 Score*
Image description not available.
Table 3. Two Existing Risk-Classification Schemes and Their Stroke Rates*
Image description not available.
1.
Kopecky SL, Gersh BJ, McGoon MD.  et al.  Lone atrial fibrillation in elderly persons: a marker for cardiovascular risk.  Arch Intern Med.1999;159:1118-1122.Google Scholar
2.
Atrial Fibrillation Investigators.  Risk factors for stroke and efficacy of antithrombotic therapy in atrial fibrillation.  Arch Intern Med.1994;154:1449-1157.Google Scholar
3.
Stroke Prevention Atrial Fibrillation III Writing Committee.  Patients with nonvalvular atrial fibrillation at low risk of stroke during treatment with aspirin: Stroke Prevention in Atrial Fibrillation III Study.  JAMA.1998;279:1273-1277.Google Scholar
4.
Stroke Prevention in Atrial Fibrillation Investigators.  Predictors of thromboembolism in atrial fibrillation, I: clinical features of patients at risk.  Ann Intern Med.1992;116:1-5.Google Scholar
5.
van Latum JC, Koudstaal PJ, Venables GS, van Gijn J, Kappelle LJ, Algra A.for the European Atrial Fibrillation Trial (EAFT) Study Group.  Predictors of major vascular events in patients with a transient ischemic attack or minor ischemic stroke and with nonrheumatic atrial fibrillation.  Stroke.1995;26:801-806.Google Scholar
6.
Hart RG, Benavente O, McBride R, Pearce LA. Antithrombotic therapy to prevent stroke in patients with atrial fibrillation: a meta-analysis.  Ann Intern Med.1999;131:492-501.Google Scholar
7.
The European Atrial Fibrillation Trial Study Group.  Secondary prevention in non-rheumatic atrial fibrillation after transient ischaemic attack or minor stroke.  Lancet.1993;342:1255-1262.Google Scholar
8.
Laupacis A, Boysen G, Connolly S.  et al.  The efficacy of aspirin in patients with atrial fibrillation: analysis of pooled data from 3 randomized trials.  Arch Intern Med.1997;157:1237-1240.Google Scholar
9.
Gage BF, Cardinalli AB, Albers GW, Owens DK. Cost-effectiveness of warfarin and aspirin for prophylaxis of stroke in patients with nonvalvular atrial fibrillation.  JAMA.1995;274:1839-1845.Google Scholar
10.
Laupacis A, Albers G, Dalen J, Dunn MI, Jacobson AK, Singer DE. Antithrombotic therapy in atrial fibrillation.  Chest.1998;114:579S-589S.Google Scholar
11.
Go AS, Hylek EM, Phillips KA.  et al.  Implications of stroke risk criteria on the anticoagulation decision in nonvalvular atrial fibrillation: the anticoagulation and risk factors in the Atrial Fibrillation (ATRIA) study.  Circulation.2000;102:11-13.Google Scholar
12.
Hart RG. Warfarin in atrial fibrillation: underused in the elderly, often inappropriately used in the young [editorial].  Heart.1999;82:539-540.Google Scholar
13.
Pearce LA, Hart RG, Halperin JL. Assessment of three schemes for stratifying stroke risk in patients with nonvalvular atrial fibrillation.  Am J Med.2000;109:45-51.Google Scholar
14.
Atrial Fibrillation Investigators.  Echocardiographic predictors of stroke in patients with atrial fibrillation: a prospective study of 1066 patients from 3 clinical trials.  Arch Intern Med.1998;158:1316-1320.Google Scholar
15.
Stroke Prevention in Atrial Fibrillation Investigators.  Risk factors for thromboembolism during aspirin therapy in patients with atrial fibrillation: The Stroke Prevention in Atrial Fibrillation study.  J Stroke Cerebrovasc Dis.1995;5:147-157.Google Scholar
16.
Sudlow M, Thomson R, Thwaites B, Rodgers H, Kenny RA. Prevalence of atrial fibrillation and eligibility for anticoagulants in the community.  Lancet.1998;352:1167-1171.Google Scholar
17.
Harrell Jr FE, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors.  Stat Med.1996;15:361-387.Google Scholar
18.
Hart RG, Pearce LA, McBride R, Rothbart RM, Asinger RW.for The Stroke Prevention in Atrial Fibrillation (SPAF) Investigators.  Factors associated with ischemic stroke during aspirin therapy in atrial fibrillation: analysis of 2012 participants in the SPAF I-III clinical trials.  Stroke.1999;30:1223-1229.Google Scholar
19.
Feinberg WM, Kronmal RA, Newman AB.  et al.  Stroke risk in an elderly population with atrial fibrillation.  J Gen Intern Med.1999;14:56-59.Google Scholar
20.
D'Agostino RB, Wolf PA, Belanger AJ, Kannel WB. Stroke risk profile: adjustment for antihypertensive medication: the Framingham Study.  Stroke.1994;25:40-43.Google Scholar
21.
Flegel KM, Hanley J. Risk factors for stroke and other embolic events in patients with non-rheumatic atrial fibrillation.  Stroke.1989;20:1000-1004.Google Scholar
22.
Yuan Z, Bowlin S, Einstadter D, Cebul R, Conners A, Rimm A. Atrial fibrillation as a risk factor for stroke: a retrospective cohort study of hospitalized Medicare beneficiaries.  Am J Public Health.1998;88:395-400.Google Scholar
23.
Stroke Prevention in Atrial Fibrillation Investigators.  Predictors of thromboembolism in atrial fibrillation, II: Echocardiographic features of patients at risk.  Ann Intern Med.1992;116:6-12.Google Scholar
24.
Moulton AW, Singer DE, Hass JS. Risk factors for stroke in patients with nonrheumatic atrial fibrillation: a case-control study.  Am J Med.1991;91:156-161.Google Scholar
25.
Jencks SF, Wilensky GR. The Health Care Quality Improvement Initiative: a new approach to quality assurance in Medicare.  JAMA.1992;268:900-903.Google Scholar
26.
Benesch C, Witter DMJ, Wilder AL, Duncan PW, Samsa GP, Matchar DB. Inaccuracy of the International Classification of Diseases (ICD-9-CM) in identifying the diagnosis of ischemic cerebrovascular disease.  Neurology.1997;49:660-664.Google Scholar
27.
Lee E. Statistical Methods for Survival Data Analysis2nd ed. New York, NY: John Wiley & Sons; 1992.
28.
Allison PD. Survival Analysis Using the SAS System: A Practical GuideCary, NC: SAS Institute Inc; 1995.
29.
Rothman KJ, Greenland S. Modern Epidemiology2nd ed. Philadelphia, Pa: Lippincott-Raven Publishers; 1998.
30.
Mooney CZ, Duval RD. Bootstrapping: A Nonparametric Approach to Statistical InferenceNewbury Park, Calif: Sage Publications Inc; 1993.
31.
Stroke Prevention in Atrial Fibrillation Investigators.  Adjusted-dose warfarin versus low-intensity, fixed-dose warfarin plus aspirin for high-risk patients with atrial fibrillation: Stroke Prevention in Atrial Fibrillation III randomised clinical trial.  Lancet.1996;348:633-638.Google Scholar
32.
Hellemons BS, Langenberg M, Lodder J.  et al.  Primary prevention of arterial thromboembolism in non-rheumatic atrial fibrillation in primary care: randomised controlled trial comparing two intensities of coumarin with aspirin.  BMJ.1999;319:958-964.Google Scholar
33.
Kearon C, Hirsh J. Management of anticoagulation before and after elective surgery.  N Engl J Med.1997;336:1506-1511.Google Scholar
34.
Gage BF, Cardinalli AB, Owens DK. The effect of stroke and stroke prophylaxis with aspirin or warfarin on quality of life.  Arch Intern Med.1996;156:1829-1836.Google Scholar
35.
Howitt A, Armstrong D. Implementing evidence based medicine in general practice: audit and qualitative study of antithrombotic treatment for atrial fibrillation.  BMJ.1999;318:1324-1327.Google Scholar
36.
Gage BF, Cardinalli AB, Owens DK. Cost-effectiveness of preference-based antithrombotic therapy for patients with nonvalvular atrial fibrillation.  Stroke.1998;29:1083-1091.Google Scholar
37.
Man-Son-Hing M, Laupacis A, O'Connor AM.  et al.  A patient decision aid regarding antithrombotic therapy for stroke prevention in atrial fibrillation: a randomized controlled trial.  JAMA.1999;282:737-743.Google Scholar
38.
Foulkes MA, Sacco RL, Mohr JP, Hier DB, Price TR, Wolf PA. Parametric modeling of stroke recurrence.  Neuroepidemiology.1994;13:19-27.Google Scholar
Original Contribution
June 13, 2001

Validation of Clinical Classification Schemes for Predicting Stroke: Results From the National Registry of Atrial Fibrillation

Author Affiliations

Author Affiliations: Divisions of General Medical Sciences (Drs Gage, Waterman, and Shannon) and Cardiology (Dr Rich), Washington University School of Medicine, St Louis, Mo; Innovative Emergency Management Inc, Baton Rouge, La (Dr Boechler); and Center for Outcomes Research and Evaluation, Yale–New Haven Hospital, New Haven, Conn (Dr Radford).

JAMA. 2001;285(22):2864-2870. doi:10.1001/jama.285.22.2864
Abstract

Context Patients who have atrial fibrillation (AF) have an increased risk of stroke, but their absolute rate of stroke depends on age and comorbid conditions.

Objective To assess the predictive value of classification schemes that estimate stroke risk in patients with AF.

Design, Setting, and Patients Two existing classification schemes were combined into a new stroke-risk scheme, the CHADS2 index, and all 3 classification schemes were validated. The CHADS2 was formed by assigning 1 point each for the presence of congestive heart failure, hypertension, age 75 years or older, and diabetes mellitus and by assigning 2 points for history of stroke or transient ischemic attack. Data from peer review organizations representing 7 states were used to assemble a National Registry of AF (NRAF) consisting of 1733 Medicare beneficiaries aged 65 to 95 years who had nonrheumatic AF and were not prescribed warfarin at hospital discharge.

Main Outcome Measure Hospitalization for ischemic stroke, determined by Medicare claims data.

Results During 2121 patient-years of follow-up, 94 patients were readmitted to the hospital for ischemic stroke (stroke rate, 4.4 per 100 patient-years). As indicated by a c statistic greater than 0.5, the 2 existing classification schemes predicted stroke better than chance: c of 0.68 (95% confidence interval [CI], 0.65-0.71) for the scheme developed by the Atrial Fibrillation Investigators (AFI) and c of 0.74 (95% CI, 0.71-0.76) for the Stroke Prevention in Atrial Fibrillation (SPAF) III scheme. However, with a c statistic of 0.82 (95% CI, 0.80-0.84), the CHADS2 index was the most accurate predictor of stroke. The stroke rate per 100 patient-years without antithrombotic therapy increased by a factor of 1.5 (95% CI, 1.3-1.7) for each 1-point increase in the CHADS2 score: 1.9 (95% CI, 1.2-3.0) for a score of 0; 2.8 (95% CI, 2.0-3.8) for 1; 4.0 (95% CI, 3.1-5.1) for 2; 5.9 (95% CI, 4.6-7.3) for 3; 8.5 (95% CI, 6.3-11.1) for 4; 12.5 (95% CI, 8.2-17.5) for 5; and 18.2 (95% CI, 10.5-27.4) for 6.

Conclusion The 2 existing classification schemes and especially a new stroke risk index, CHADS2, can quantify risk of stroke for patients who have AF and may aid in selection of antithrombotic therapy.

The atrial fibrillation (AF) population is heterogeneous in terms of ischemic stroke risk. Subpopulations have annual stroke rates that range from less than 2% to more than 10%.1-5 Because the relative risk reductions from warfarin sodium (62%) and aspirin (22%) therapy are consistent across these subpopulations,2,6-8 the absolute benefit of antithrombotic therapy depends on the underlying risk of stroke. Although there has been agreement that warfarin therapy is favored when the risk of stroke is high and that aspirin is favored when the risk of stroke is low,9,10 there has been little agreement about how to predict the risk of stroke.11-13 Thus, an accurate, objective scheme to estimate the risk of stroke in the AF population would allow physicians and patients to choose antithrombotic therapy more judiciously.

The Atrial Fibrillation Investigators (AFI) pooled data from several trials to form a unified stroke classification scheme. Among trial participants who did not receive antithrombotic therapy, these researchers found that the risk of stroke increased by a factor of 1.4 per decade of age and by 3 clinical risk factors: hypertension, prior cerebral ischemia (either stroke or transient ischemic attack [TIA]), and diabetes mellitus (DM).2,8 There were 5.9 to 10.4 strokes per 100 patient-years among participants randomized to no antithrombotic therapy who had at least 1 of the 3 clinical risk factors.2,8 In contrast to these high-risk participants, Medicare-aged participants without any of these risk factors were at moderate risk of stroke, averaging 2.7 to 4.3 strokes per 100 patient-years. Participants younger than age 65 years who had none of the 3 risk factors were at low risk for stroke, averaging 1.0 to 1.8 strokes per 100 patient-years, and if they also lacked 2 equivocal stroke risk factors—coronary artery disease and congestive heart failure (CHF)—they had 0.0 to 1.6 strokes per 100 patient-years.2,8,14

The Stroke Prevention and Atrial Fibrillation (SPAF) investigators reported their classification scheme from SPAF participants who were treated with aspirin therapy. Based on data from their first 2 trials, the SPAF investigators identified 4 independent risk factors for stroke: blood pressure higher than 160 mm Hg, prior cerebral ischemia, recent heart failure (ie, active within the past 100 days) or documented by echocardiography, or the combination of 75 years or older and being female.15 In SPAF III, participants with hypertension lacking these risk factors had an annual stroke rate of 3.2 to 3.6 per 100 patient-years of aspirin therapy,3,13 and the rate was only 1.1 in those who also lacked hypertension.

The promulgation of 2 stroke-risk classification schemes (AFI and SPAF), each with cautions about how equivocal risk factors influenced the risk of stroke in low-risk participants, complicates the estimation of stroke risk. First, the 2 schemes conflict: many patients classified as low risk by one scheme are classified as moderate or high risk by the other.11-13,16 Second, the classification schemes are sometimes ambiguous. For example, into which SPAF risk group should one classify a patient who initially presented with a systolic blood pressure higher than 160 mm Hg but whose blood pressure is controlled on follow-up evaluation? Third, the development of both classification schemes was data driven, and therefore the schemes could have captured apparent risk factors that represented idiosyncrasies in the data set rather than true associations.17 Fourth, because the original schemes were based on trial participants whose average age was 69 years, their performance in older and frailer populations is not well characterized.12 Given these limitations, we decided to validate the 2 existing classification schemes and their variations10,18,19 in an independent sample.

Our goal was to find a convenient and accurate classification scheme to estimate stroke risk in a national registry of Medicare-aged patients who have nonrheumatic AF and were not prescribed warfarin at the time of hospital discharge.

Methods
Formation of CHADS

We amalgamated the 2 classification schemes to form a new stroke-risk index, CHADS2. We then assessed the predictive accuracy of the AFI, SPAF, and CHADS2 schemes using data from a registry of Medicare beneficiaries who had AF. To create the CHADS2 index, we included independent risk factors that were identified in either the AFI or SPAF schemes: prior cerebral ischemia, history of hypertension, DM, CHF, and age 75 years or older. We included a history of hypertension, rather than having blood pressure higher than 160 mm Hg, because even well-controlled hypertension is an independent risk factor for stroke.20 We included age 75 years or older rather than the combination of age 75 years or older plus female sex, because there is an age-related increase in stroke in both women and men.2,5,21,22 We included recent CHF exacerbation, rather than any CHF, because the former is similar to the CHF definition used in the SPAF scheme.

To create CHADS2, we assigned 2 points to a history of prior cerebral ischemia and 1 point for the presence of other risk factors because a history of prior cerebral ischemia increases the relative risk (RR) of subsequent stroke commensurate to 2 other risk factors combined.2,4,5,23,24 We calculated CHADS2 by adding 1 point each for any of the following—recent CHF, hypertension, age 75 years or older, and DM—and 2 points for a history of stroke or TIA. Thus, CHADS2 is an acronym for the risk factors and their scoring. For example, an 82-year-old (+1) patient who had hypertension (+1) and a prior stroke (+2) would have a CHADS2 score of 4.

State-Specific Cohorts

The NRAF data set contained anonymous patient records gathered by 5 quality improvement/peer review organizations (QIO/PROs) that serve 7 states (California, Connecticut, Louisiana, Maine, Missouri, New Hampshire, and Vermont). These QIO/PROs had assembled state-specific cohorts of patients with AF for quality improvement projects under the Health Care Quality Improvement Initiative of the Health Care Financing Administration.25 Using Medicare Part A claims records (MEDPAR), the QIO/PRO analysts used the appropriate International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) code 427.31 in either a principal or secondary diagnosis to identify Medicare beneficiaries who may have had AF.

Through record review, including electrocardiographic and physician documentation, QIO/PRO reviewers confirmed the presence of chronic or recurrent AF during the index hospitalization: Medicare beneficiaries who had acute AF and beneficiaries who died during hospitalization were excluded. During their chart abstractions, QIO/PRO reviewers documented stroke risk factors, other comorbid conditions, and the antithrombotic therapy prescribed at hospital discharge. No additional charts were abstracted to create the NRAF dataset. Abstractors used standardized abstraction forms and statewide sampling techniques that had been adapted by each QIO/PRO participant, but QIO/PRO protocols were sufficiently similiar to allow the statewide data sets to be combined. The 2 QIO/PRO reviewers who calculated the inter-rater reliability for the chart abstraction found agreement between abstractors of more than 90%.

To obtain outcomes, each QIO/PRO linked chart abstractions from the index hospitalizations to MEDPAR. The QIO/PRO analyst obtained the dates of death from a separate source, the denominator file of living Medicare beneficiaries. After linking a maximum of 3 years of follow-up data and removing all patient and provider identifiers, the QIO/PRO analysts sent the unidentified records to Washington University for inclusion into the NRAF data set. The study was approved by the human subjects committee at Washington University Medical Center and by the participating PRO/QIOs.

Formation of the NRAF Data Set

We used the QIO/PRO records to develop an NRAF data set of Medicare beneficiaries who had documented chronic or recurrent nonrheumatic AF. With few exceptions, we obtained the potential stroke-risk factors from the chart reviews. One exception is that we defined recent CHF as an index hospitalization that carried the principal diagnosis of CHF based on ICD-9-CM codes 398.91, 402.01, 402.11, 402.91, 428.x, or 518.4. The other exceptions were that 1 PRO/QIO did not document DM in Medicare beneficiaries and 2 did not exclude the presence of rheumatic heart disease. For these missing fields, we imputed the relevant history from the appropriate ICD-9-CM codes (250.x for DM and 393.x-398.x for rheumatic heart disease) and then excluded patients who had rheumatic heart disease.

From the original NRAF data set of 3932 Medicare beneficiaries who had documented AF, we excluded 2199 beneficiaries from the original data set for the following reasons: 229 for mitral stenosis or rheumatic heart disease; 555 for recent surgery or trauma; 81 for transfer to another acute care facility; 65 whose ages were younger than 65 or older than 95 years; and 1269 for being discharged with a prescription for warfarin. Thus, the study cohort included 1733 patients, aged 65 to 95 years, who had nonrheumatic AF and who were not prescribed warfarin therapy at the time of hospital discharge.

Outcomes Assessment

The study outcome was hospitalization for ischemic stroke as determined by Medicare claims. To identify stroke from the MEDPAR data, we used the following ICD-9-CM codes in the primary position: 434 (occlusion of cerebral arteries), 435 (transient cerebral ischemia), and 436 (acute, but ill-defined, cerebrovascular disease). We did not use the ICD-9-CM code of 433 (occlusion and stenosis of precerebral arteries) because that code is used for asymptomatic carotid artery disease.26 We had a minimum of 365 days of follow-up claims for all Medicare beneficiaries, and we censored beneficiaries at the time of nonstroke death or at a maximum of 1000 days after the index hospitalization. For beneficiaries who experienced multiple strokes, we excluded events and patient-days of follow-up that occurred after the initial stroke.

Statistical Analyses

To calculate the stroke rate as a function of CHADS2, we used an exponential survival model.27 We used the survival model to measure how the hazard rate for stroke was affected by each 1-point increase in CHADS2 and by prescription of aspirin. We also used the model to predict the annual rate of stroke as a function of CHADS2 and of aspirin use. We confirmed the appropriateness of using an exponential survival model graphically (by plotting the negative of the logarithm of the survival curve vs time).28 We performed our survival analyses in SAS (Version 6.12; SAS Institute Inc; Cary, NC) using the LIFEREG and LIFETEST procedures.28 We calculated the RR reduction from aspirin therapy as 1 minus the relative hazard of prescribing aspirin (as obtained from the exponential model).

To determine the predictive validity of each of the 3 classification schemes, we performed additional time-to-event analyses. We censored deaths that were not accompanied by a hospitalization for stroke and then calculated the stroke rate based on patient-years of follow-up data29 for each risk group identified by each classification scheme. We calculated the 95% confidence interval (CI) of these rates using the binomial approximation. We quantified the predictive validity of the classification schemes by using the c statistic17 to test the hypothesis that these classification schemes performed significantly better than chance (indicated by a c statistic of 0.5).

To determine the predictive accuracy of the classification schemes, we used the bootstrap analysis to generate 95% CIs using the percentile method.30 We calculated the c statistic on a random sample of all patients 1000 times and then noted the 2.5% and 97.5% percentiles. We declared the classification schemes statistically significantly different if these CIs did not overlap.

Results
Baseline Characteristics of the NRAF Cohort

Compared with participants in the AF trials, the 1733 patients in the NRAF cohort were more likely to be women and elderly and more often had stroke risk factors: history of CHF (56%), CHF as the reason for the index hospitalization (14%); history of hypertension (56%); DM (23%); and history of cerebral ischemia (25%) (Table 1). The mean CHADS2 score was 2.1 for the 1204 members of the NRAF cohort who were not prescribed any antithrombotic therapy and 2.3 for the 529 members who were prescribed aspirin.

Stroke Rate in the NRAF Cohort

The 1733 patients were followed up for a mean (median) of 1.2 (1.0) years. During the 2121 patient-years of follow-up, 94 NRAF patients were readmitted for an ischemic event (rate, 4.4 per 100 patient-years), 71 patients were admitted for a stroke as indicated by ICD-9-CM codes 434 or 436, and 23 patients were admitted for transient cerebral ischemia as indicated by ICD-9-CM code 435. We refer to all of these events as stroke for simplicity and because 8 of the 23 patients had a subsequent hospitalization with ICD-9-CM code 434 or 436. Of the 94 patients admitted for a stroke, 25 (27%) died within 30 days of the hospital admission.

The stroke rate was lowest among the 120 patients in the NRAF cohort who had a CHADS2 score of 0, a crude stroke rate of 1.2, and an adjusted rate of 1.9 per 100 patient-years without antithrombotic therapy (Table 2). The stroke rate increased by a factor of 1.5 (95% CI, 1.3-1.7) for each 1-point increase in the CHADS2 score (P<.001). Aspirin was associated with a hazard rate of 0.80 (95% CI, 0.5-1.3), corresponding to a nonsignificant 20% RR reduction in the rate of stroke (P = .27).

The AFI and the SPAF classification schemes also identified patients at low risk for stroke. The 303 patients (17%) in the NRAF cohort identified as low risk according to the SPAF scheme had 1.5 strokes per 100 patient-years of follow-up (Table 3), which is similar to the published rate of thromboembolism for this population, 1.1 per 100 patient-years of aspirin therapy.3 The 490 cohort members (27%) classified as moderate risk, according to the AFI scheme, averaged 2.2 strokes per 100 patient-years of follow-up (Table 3), which is similar to the published stroke rate of 2.7 to 4.3 for this population.2,8 When we excluded all cohort members aged 75 years or older from the AFI moderate-risk cohort, the stroke rate was 1.1 per 100 patient-years of follow-up, but only 130 patients (8%) of the NRAF population were in this cohort.

Accuracy of the Stroke Classification Schemes

The AFI scheme had a c statistic of 0.68 (95% CI, 0.65-0.71); the SPAF scheme, 0.74 (95% CI, 0.71-0.76); and CHADS2 0.82 (95% CI, 0.80-0.84). Variations of the classification schemes did not improve their predictive accuracy. When we included all patients aged 75 years or older as high risk in the AFI scheme, its corresponding c statistic was 0.49. A variation of the SPAF scheme that included patients with DM in the moderate risk group and did not include CHF as risk factor18 had a c statistic of 0.72, not significantly different from the original SPAF scheme. A variation of CHADS2 that counted 1 point for the presence of any CHF had a c statistic of 0.82, identical to CHADS2 that included CHF only if it were the principal diagnosis coded on admission.

In post hoc analyses, we found that CHADS2 was a more accurate predictor of stroke both in cohort members who did (n = 529) and who did not (n = 1204) receive aspirin. For example, in NRAF cohort members who were not prescribed aspirin the c statistics were 0.71 for the AFI scheme, 0.76 for the SPAF scheme, and 0.84 for CHADS2 scheme.

We also performed post hoc analyses to determine whether the greater predictive accuracy of CHADS2 was due to its greater number of risk strata. We collapsed the 7 CHADS2 strata (Table 2) into 3 strata: low risk (CHADS2 0 or 1), moderate risk (CHADS2 2 or 3), or high risk (CHADS2 4, 5, or 6) ; CHADS2 with 3 strata had a c statistic of 0.78, which was 0.04 less than the value obtained from the complete CHADS2. We also collapsed CHADS2 into 2 strata by combining scores 0 with 1 and then scores 2 through 6; CHADS2 with 2 strata had a c statistic of 0.71.

Comment

This study validated 2 existing stroke-risk classification schemes and a combination of these schemes, CHADS2, in Medicare beneficiaries with nonrheumatic AF who had been followed up from 365 to 1000 days after an index hospitalization. The AFI and SPAF schemes successfully identified cohorts with stroke rates of 1.5 to 2.2 per 100 patient-years, whereas CHADS2 identified a low-risk cohort with an adjusted stroke rate of 1.5 per 100 patient-years without antithrombotic therapy. Overall, CHADS2 had greater predictive accuracy than did either AFI or SPAF schemes.

Other studies support our finding that Medicare-aged patients with AF at low risk for stroke can be identified prospectively. Feinberg et al19 observed a stroke rate of 1.7 per 100 patient-years without warfarin therapy in 66 patients classified as low risk according to the SPAF definition and observed a stroke rate of 2.4 per 100 patient-years without warfarin therapy in 47 patients classified as moderate risk according to the AFI scheme. Hellemons et al32 found an annual rate of stroke of approximately 2.1 per 100 patient-years of aspirin therapy in an AF trial that excluded patients who had a prior stroke or TIA or age greater than 77 years. During long-term follow-up of 55 elderly patients with lone AF, Kopecky et al1 found a stroke rate of 0.9 per 100 patient-years, despite a low use of warfarin. These studies support the premise that by applying a classification scheme, clinicians can identify AF patients who are at low risk for stroke even without warfarin therapy.

Because the net benefit of antithrombotic therapy correlates with the underlying risk of stroke, CHADS2 may be helpful in several clinical settings. For example, in identifying low-risk patients, a CHADS2 score of 0 defines an AF population who should be offered the option of aspirin therapy. In addition, CHADS2 could aid in decision making for patients with AF and who are undergoing surgical or dental procedure because perioperative management depends on their risk of hemorrhage from the procedure compared with the underlying thrombotic risk.33 Also, in patients for whom taking warfarin would be burdensome,34,35 CHADS2 could facilitate risk stratification and selection of antithrombotic therapy based on a patient-specific risk of stroke.36,37

Our study has several strengths. First, we used chart review, rather than ICD-9-CM claims, to document the presence of AF and to identify the stroke risk factors. These chart reviews also identified patients who were discharged from the hospital and received aspirin, enabling adjustment for the prescription of aspirin in our calculations of the CHADS2-specific stroke rates. Because of the number of strokes (94), we were able to calculate stroke rates with precision (Table 2). The NRAF cohort included Medicare beneficiaries from 7 states that represented all geographic regions of the United States. Because we formulated CHADS2 based on previous studies, rather than on the NRAF data set, our study validates CHADS2. In addition, because we validated CHADS2 in Medicare beneficiaries who were recently hospitalized, rather than in healthier trial participants, CHADS2 should be generalizable to frail or elderly patients who have nonrheumatic AF.

The CHADS2 scheme with either definition of CHF that we tested in CHADS2—any CHF as identified on chart review and CHF identified as the principal diagnosis by ICD-9-CM code—had a c statistic of 0.82. We used the later definition in CHADS2 because it was closer to the definition of CHF that was an independent predictor of stroke in other studies: CHF that caused symptoms within the past 100 days was an independent predictor of stroke in SPAF,4,15 and moderate or severe left-ventricular systolic dysfunction on echocardiography was an independent predictor of stroke in both SPAF and AFI.14,15,23 We did not have access to echocardiographic results, which would have allowed us to assess how they could have improved the predictive accuracy of CHADS2 and SPAF.

Our study had several important limitations. First, NRAF cohort members were older and sicker than participants in clinical trials (Table 1), and the AFI and SPAF schemes may have performed better in a healthier population that included patients younger than 65 years. Because CHADS2 was based on the AFI and SPAF schemes, it too might have performed better in a younger or healthier population. Second, we used a single chart review to assess most of the stroke risk factors and had no way of capturing new risk factors if they developed. Third, we studied patients who had been hospitalized and who were not prescribed warfarin from our analyses. Future analyses are needed to evaluate CHADS2 and the other classification schemes in other populations. Fourth, we used Medicare claims to ascertain ischemic events and have no way to verify these events. Because Medicare claims cannot be expected to capture all strokes, our estimates of the CHADS2-specific stroke rates may be biased downward. Comparison of stroke rates that we observed with published rates (Table 3) suggests that this ascertainment bias is modest. In contrast, the CHADS2-specific stroke rates may significantly underestimate the stroke rate in patients who have high-risk conditions that we did not consider, such as mitral stenosis, cardiac thrombus, mechanical heart valve, recent anterior myocardial infarction, or high-grade carotid artery stenosis. Likewise, CHADS2 may underestimate stroke risk in patients with hypertension that exceeds 160 mm Hg5,15 or in patients with a prior stroke or TIA that occurred in the previous 3 months.7,38

Although the 20% risk reduction for aspirin effectiveness in preventing stroke was not statistically significant in this study, it does have clinical significance when combined with other research. Our finding is consistent with a recent meta-analysis that estimated a 22% risk reduction for aspirin therapy6 but differs from a subgroup analysis that found no effectiveness of aspirin in AF populations aged 75 years or older.8 Thus, although our nonrandomized study could not determine the effectiveness of aspirin, our results suggest that aspirin therapy should be prescribed for elderly patients with AF who are not suitable candidates for warfarin.

In summary CHADS2 is an easy-to-use classification scheme that estimates the risk of stroke in elderly patients with AF. Physicians and patients could use CHADS2 to make decisions about antithrombotic therapy based on patient-specific risk of stroke.

References
1.
Kopecky SL, Gersh BJ, McGoon MD.  et al.  Lone atrial fibrillation in elderly persons: a marker for cardiovascular risk.  Arch Intern Med.1999;159:1118-1122.Google Scholar
2.
Atrial Fibrillation Investigators.  Risk factors for stroke and efficacy of antithrombotic therapy in atrial fibrillation.  Arch Intern Med.1994;154:1449-1157.Google Scholar
3.
Stroke Prevention Atrial Fibrillation III Writing Committee.  Patients with nonvalvular atrial fibrillation at low risk of stroke during treatment with aspirin: Stroke Prevention in Atrial Fibrillation III Study.  JAMA.1998;279:1273-1277.Google Scholar
4.
Stroke Prevention in Atrial Fibrillation Investigators.  Predictors of thromboembolism in atrial fibrillation, I: clinical features of patients at risk.  Ann Intern Med.1992;116:1-5.Google Scholar
5.
van Latum JC, Koudstaal PJ, Venables GS, van Gijn J, Kappelle LJ, Algra A.for the European Atrial Fibrillation Trial (EAFT) Study Group.  Predictors of major vascular events in patients with a transient ischemic attack or minor ischemic stroke and with nonrheumatic atrial fibrillation.  Stroke.1995;26:801-806.Google Scholar
6.
Hart RG, Benavente O, McBride R, Pearce LA. Antithrombotic therapy to prevent stroke in patients with atrial fibrillation: a meta-analysis.  Ann Intern Med.1999;131:492-501.Google Scholar
7.
The European Atrial Fibrillation Trial Study Group.  Secondary prevention in non-rheumatic atrial fibrillation after transient ischaemic attack or minor stroke.  Lancet.1993;342:1255-1262.Google Scholar
8.
Laupacis A, Boysen G, Connolly S.  et al.  The efficacy of aspirin in patients with atrial fibrillation: analysis of pooled data from 3 randomized trials.  Arch Intern Med.1997;157:1237-1240.Google Scholar
9.
Gage BF, Cardinalli AB, Albers GW, Owens DK. Cost-effectiveness of warfarin and aspirin for prophylaxis of stroke in patients with nonvalvular atrial fibrillation.  JAMA.1995;274:1839-1845.Google Scholar
10.
Laupacis A, Albers G, Dalen J, Dunn MI, Jacobson AK, Singer DE. Antithrombotic therapy in atrial fibrillation.  Chest.1998;114:579S-589S.Google Scholar
11.
Go AS, Hylek EM, Phillips KA.  et al.  Implications of stroke risk criteria on the anticoagulation decision in nonvalvular atrial fibrillation: the anticoagulation and risk factors in the Atrial Fibrillation (ATRIA) study.  Circulation.2000;102:11-13.Google Scholar
12.
Hart RG. Warfarin in atrial fibrillation: underused in the elderly, often inappropriately used in the young [editorial].  Heart.1999;82:539-540.Google Scholar
13.
Pearce LA, Hart RG, Halperin JL. Assessment of three schemes for stratifying stroke risk in patients with nonvalvular atrial fibrillation.  Am J Med.2000;109:45-51.Google Scholar
14.
Atrial Fibrillation Investigators.  Echocardiographic predictors of stroke in patients with atrial fibrillation: a prospective study of 1066 patients from 3 clinical trials.  Arch Intern Med.1998;158:1316-1320.Google Scholar
15.
Stroke Prevention in Atrial Fibrillation Investigators.  Risk factors for thromboembolism during aspirin therapy in patients with atrial fibrillation: The Stroke Prevention in Atrial Fibrillation study.  J Stroke Cerebrovasc Dis.1995;5:147-157.Google Scholar
16.
Sudlow M, Thomson R, Thwaites B, Rodgers H, Kenny RA. Prevalence of atrial fibrillation and eligibility for anticoagulants in the community.  Lancet.1998;352:1167-1171.Google Scholar
17.
Harrell Jr FE, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors.  Stat Med.1996;15:361-387.Google Scholar
18.
Hart RG, Pearce LA, McBride R, Rothbart RM, Asinger RW.for The Stroke Prevention in Atrial Fibrillation (SPAF) Investigators.  Factors associated with ischemic stroke during aspirin therapy in atrial fibrillation: analysis of 2012 participants in the SPAF I-III clinical trials.  Stroke.1999;30:1223-1229.Google Scholar
19.
Feinberg WM, Kronmal RA, Newman AB.  et al.  Stroke risk in an elderly population with atrial fibrillation.  J Gen Intern Med.1999;14:56-59.Google Scholar
20.
D'Agostino RB, Wolf PA, Belanger AJ, Kannel WB. Stroke risk profile: adjustment for antihypertensive medication: the Framingham Study.  Stroke.1994;25:40-43.Google Scholar
21.
Flegel KM, Hanley J. Risk factors for stroke and other embolic events in patients with non-rheumatic atrial fibrillation.  Stroke.1989;20:1000-1004.Google Scholar
22.
Yuan Z, Bowlin S, Einstadter D, Cebul R, Conners A, Rimm A. Atrial fibrillation as a risk factor for stroke: a retrospective cohort study of hospitalized Medicare beneficiaries.  Am J Public Health.1998;88:395-400.Google Scholar
23.
Stroke Prevention in Atrial Fibrillation Investigators.  Predictors of thromboembolism in atrial fibrillation, II: Echocardiographic features of patients at risk.  Ann Intern Med.1992;116:6-12.Google Scholar
24.
Moulton AW, Singer DE, Hass JS. Risk factors for stroke in patients with nonrheumatic atrial fibrillation: a case-control study.  Am J Med.1991;91:156-161.Google Scholar
25.
Jencks SF, Wilensky GR. The Health Care Quality Improvement Initiative: a new approach to quality assurance in Medicare.  JAMA.1992;268:900-903.Google Scholar
26.
Benesch C, Witter DMJ, Wilder AL, Duncan PW, Samsa GP, Matchar DB. Inaccuracy of the International Classification of Diseases (ICD-9-CM) in identifying the diagnosis of ischemic cerebrovascular disease.  Neurology.1997;49:660-664.Google Scholar
27.
Lee E. Statistical Methods for Survival Data Analysis2nd ed. New York, NY: John Wiley & Sons; 1992.
28.
Allison PD. Survival Analysis Using the SAS System: A Practical GuideCary, NC: SAS Institute Inc; 1995.
29.
Rothman KJ, Greenland S. Modern Epidemiology2nd ed. Philadelphia, Pa: Lippincott-Raven Publishers; 1998.
30.
Mooney CZ, Duval RD. Bootstrapping: A Nonparametric Approach to Statistical InferenceNewbury Park, Calif: Sage Publications Inc; 1993.
31.
Stroke Prevention in Atrial Fibrillation Investigators.  Adjusted-dose warfarin versus low-intensity, fixed-dose warfarin plus aspirin for high-risk patients with atrial fibrillation: Stroke Prevention in Atrial Fibrillation III randomised clinical trial.  Lancet.1996;348:633-638.Google Scholar
32.
Hellemons BS, Langenberg M, Lodder J.  et al.  Primary prevention of arterial thromboembolism in non-rheumatic atrial fibrillation in primary care: randomised controlled trial comparing two intensities of coumarin with aspirin.  BMJ.1999;319:958-964.Google Scholar
33.
Kearon C, Hirsh J. Management of anticoagulation before and after elective surgery.  N Engl J Med.1997;336:1506-1511.Google Scholar
34.
Gage BF, Cardinalli AB, Owens DK. The effect of stroke and stroke prophylaxis with aspirin or warfarin on quality of life.  Arch Intern Med.1996;156:1829-1836.Google Scholar
35.
Howitt A, Armstrong D. Implementing evidence based medicine in general practice: audit and qualitative study of antithrombotic treatment for atrial fibrillation.  BMJ.1999;318:1324-1327.Google Scholar
36.
Gage BF, Cardinalli AB, Owens DK. Cost-effectiveness of preference-based antithrombotic therapy for patients with nonvalvular atrial fibrillation.  Stroke.1998;29:1083-1091.Google Scholar
37.
Man-Son-Hing M, Laupacis A, O'Connor AM.  et al.  A patient decision aid regarding antithrombotic therapy for stroke prevention in atrial fibrillation: a randomized controlled trial.  JAMA.1999;282:737-743.Google Scholar
38.
Foulkes MA, Sacco RL, Mohr JP, Hier DB, Price TR, Wolf PA. Parametric modeling of stroke recurrence.  Neuroepidemiology.1994;13:19-27.Google Scholar
×