Assessing absolute stroke risk in patients with atrial fibrillation using a risk factor-based approach

Abstract

Aim

To assess the risk of stroke and thromboembolism in patients with atrial fibrillation (AF) based on risk factor combinations of the CHA₂DS₂-VASc score.

Methods and results

Using nationwide Danish registries, patients with AF were included from 1997 to 2015 in this retrospective observational study. A multiple logistic regression, including interactions of history of stroke with age at AF, calendar year of AF, and the CHA₂DS₂-VASc score risk factors (congestive heart failure, hypertension, diabetes, vascular disease, and female sex) were used to predict the personalized risks of stroke within 1 year. A total of 147 842 patients with AF were included in the study cohort (median age 76 years, range 20–100 years, 51% females). Within the first year, 6% of the cohort were diagnosed with stroke. The predicted personalized 1-year absolute risk of stroke varied widely within each CHA₂DS₂-VASc score. To estimate the personalized risk of stroke an online calculator was created, the Calculator of Absolute Stroke Risk (CARS), which allows calculation of all the possible combinations of the CHA₂DS₂-VASc score (https://hjerteforeningen.shinyapps.io/riskvisrr/).

Conclusion

Calculation of the individual risk using a risk factor-based approach as opposed to using average risk for a particular CHA₂DS₂-VASc score can improve risk estimates. Furthermore, CARS can assist in the communication of the stroke risk for a more evidence-based shared decision-making of whether to initiate oral anticoagulation therapy.

Introduction

Atrial fibrillation (AF) is a major risk factor for stroke¹ and the need for anticoagulation is a key consideration in the management of patients with AF. To identify patients at either low or high risk of stroke, guidelines recommend risk stratification by the CHA₂DS₂-VASc score,^2–4 which incorporates risk factors for stroke (congestive heart failure, hypertension, age 65–74 years or ≥75 years, diabetes mellitus, previous stroke, vascular disease, and female sex). Anticoagulation therapy is recommended for patients with a CHA₂DS₂-VASc score of ≥2 in males or ≥3 in females and may be considered in those with one non-sex CHA₂DS₂-VASc risk factor. The CHA₂DS₂-VASc score is a simplification of risk, designed for easy clinical application at a time when advanced calculators were not ubiquitously available, and includes only the more common and validated risk factors.⁵ Notably, its risk factor components do not all carry equal weight, the combinations of risk factors for different patients are highly variable, and the CHA₂DS₂-VASc score has little precision.^6,7 The low precision might originate in an oversimplification inherent to the model. The broad age categories do not take into account age as a substantial risk factor for stroke, and the significance of age depends heavily on a prior stroke.⁸ Furthermore, the risk of stroke in females has been debated.^9–12

To provide more precise guidance for anticoagulation therapy in individual patients, we examined the 1-year absolute risk of stroke and thromboembolism associated with the risk factors in the CHA₂DS₂-VASc score and their combinations in patients with AF without prevalent anticoagulation therapy. Furthermore, based on the risk prediction model, an online calculator was constructed, the Calculator of Absolute Stroke Risk (CARS), which is available online to present the personalized absolute risk of stroke.

Methods

Data registers

All residents in Denmark are provided a unique identification number at the date of birth or immigration, which facilitates cross-linkage of individual data across all Danish administrative registers.¹³ The Danish National Patient Register holds information on all discharges from hospitals in Denmark and the treating physician has assigned every discharge one primary and in some cases, secondary diagnoses according to the International Classification of Diseases (ICD) of the 8th revision (ICD-8) until 1994 and the 10th revision (ICD-10) from 1994.¹⁴ Operation classification codes have been assigned according to the Nordic Medical Statistics Committees Classification of Surgical Procedures. The Danish register of Medicinal Product Statistics (the national prescription register) holds records on drug prescriptions dispensed from Danish pharmacies.¹⁵ Each dispensed prescription is registered according to the Anatomical Therapeutic Chemical Classification system (ATC). Vital status and cause of death can be identified from the Civil Registration System and the Danish Register of Causes of Death.¹⁶ See Supplementary material online, eTable 1 for ICD and ATC codes used in this study.

Population

All patients admitted with first-time AF or atrial flutter (AF) between 1 January 1997 and 31 December 2015 were included in the study. The diagnosis of AF has been validated in the Danish National Patient Register with a positive predictive value of 92%.¹⁷

A 7-day blanking period between the diagnosis of AF and the index date of follow-up (AF diagnosis date plus 7 days) was introduced to exclude patients from the cohort if they commenced anticoagulation immediately after AF diagnosis and patients of if they suffered and event or died shortly after diagnosis. Additional exclusion criteria were age below 30 or above 100 years or residence outside Denmark. Valvular AF was excluded and defined as no previous diagnoses of rheumatic or prosthetic valve operations as previously done.⁶

Outcome

The outcome was admission to hospital or death from stroke within 1 year after the index date. Stroke was defined as cerebral infarction, stroke not specified as haemorrhage or infarction, or transient ischaemic attack.^18,19

The risk factors of the CHA₂DS₂-VASc score

Risk factors were defined from a combination of hospital records and redeemed prescriptions. Prior stroke was defined as the outcome stroke and arterial thromboembolism. Heart failure was defined as present with either a diagnosis or a redeemed prescription of loop diuretics, or both. Hypertension was redemption of at least two antihypertensive drugs concomitantly,⁶ or a diagnosis of hypertension. Diabetes mellitus was defined as redeeming antidiabetic drugs. Vascular disease was defined as having at least one diagnosis of peripheral artery disease, prior myocardial infarction, or aortic plaque.

Statistical analyses

Patient characteristics at index AF were presented as medians with interquartile ranges (IQRs), and frequencies with percentages as appropriate. The time origin for all survival analyses started at index AF. Patients were followed for 1 year until stroke (event of interest), or death without stroke (competing risk), whichever came first. Emigration was minimal and neglected so that time-to-event data were uncensored after 1 year. Multiple logistic regression was used to predict the personalized risks of stroke within 1 year. The model included interactions of history of stroke (yes/no) with age at AF, calendar year of AF, and the CHA₂DS₂-VASc score risk factors (sex, heart failure, hypertension, diabetes, and vascular disease). The continuous variables, patient age and calendar year, were included with non-linear effects by restricted cubic splines.²⁰ Presented are personalized 1-year risk predictions with 95% confidence intervals (CIs). Based on the model, we also predicted the 1-year risks of all combinations of the risk factors and for the calendar year 2015. These are available at our online calculator CARS: https://hjerteforeningen.shinyapps.io/riskvisrr/.

To evaluate and compare our model in terms of the predictive performance in new patients, the logistic regression model was re-estimated by including only the data of patients diagnosed with AF before 1 January 2015 (CARS-2014). The predictive performance was then estimated in patients diagnosed with AF in year 2015. The predictive performance was assessed overall with the Brier score and the area under the receiver operating characteristic curve (AUC)²¹ both evaluated at the 1-year prediction horizon. Calibration plots were obtained by using a running average²² and used to assess the model’s predicted risks with observed 1-year risks in patients diagnosed with AF in 2015. We also fitted a second logistic regression model which included the CHA₂DS₂-VASc score (Scores 1–9, as eight dummy variables) as the only risk factor information also in data of patients diagnosed before 1 January 2015. The significance of the difference in AUC between our CARS-2014 model and the CHA₂DS₂-VASc score model was evaluated using Delong–Delong tests²³ using data of patients diagnosed in 2015. Further analyses were performed using a wider definition of the endpoint thromboembolism, which was defined as stroke and systemic arterial embolism or myocardial infarction, and a stricter definition where stroke was defined only as cerebral infarction. The level of statistical significance was set at 5%. Statistical calculations were performed using SAS version 9.4 (SAS Institute, Cary, NC, USA), and R version 3.5.0.²⁴

Ethics

Anonymized register studies do not require prior approval from the ethics committees in Denmark. The Danish Data Protection Agency approved this study (j.nr. 2007-58-0015, local j.nr. GEH-2014-013, I-Suite no. 02731).

Results

Study cohort

Among 294 224 patients with an initial diagnosis of AF, use of anticoagulation therapy before the index date excluded 110 008 (37%) patients (Figure 1). A cohort of 147 842 patients were included in the study (median age 76 years, 51% females). Table 1 shows the baseline characteristics of the study cohort by their CHA₂DS₂-VASc scores from 0 to 3 and above stratified by sex (see Supplementary material online, eTable 2 for the study cohort). One percent of the cohort had a prior stroke or thromboembolism. Women with a prior stroke or thromboembolism were older than the men and had a higher percentage of diabetes and heart failure (Supplementary material online, eTable 3). During the 1-year follow-up, 17% of the cohort initiated oral anticoagulation therapy.

Risk of stroke

Figure 2 shows the variance of the 1-year risk of stroke by the CHA₂DS₂-VASc score from 0 to 9 in our model. A total of 8377 (6%) patients had a stroke within 1 year. The median age of patients at each strata ranged from 54 years (IQR 45–60) at CHA₂DS₂-VASc point 0 to 81 years (IQR 75–87) at CHA₂DS₂-VASc ≥3. The estimated absolute 1-year risk varied widely for each CHA₂DS₂-VASc score point, especially for CHA₂DS₂-VASc scores of 2 and above.

The absolute risk of stroke with CHA₂DS₂-VASc scores of 0 (perceived low risk) and 1 (intermediate risk) in men overlapped with the risk of a CHA₂DS₂-VASc score ranging from 2 (the minimum score required for recommendation of anticoagulation therapy) to 5.

The AUC for the CARS-2014 model of 1-year risk of primary event stroke was 78.7 (95% CI 76.3–81.2). The logistic regression model with CHA₂DS₂-VASc scores had a significantly lower 1-year AUC [73.3 (95% CI 70.8–75.8), difference in AUC 5.4% (95% CI 3.8–7.0), Delong–Delong test P-value <0.001] (Supplementary material online, eFigure 1). Also, the Brier score of the CHA₂DS₂-VASc model was significantly higher than the Brier score of our model [delta Brier: 0.2% (0.1–0.3%), P < 0.001]. The calibration plot of the CARS-2014 model is shown in Supplementary material online, eFigure 2.

To illustrate the impact of age as a risk factor and the variation of the effect of the different risk factors, Figure 3 illustrates the estimated absolute 1-year risk of stroke with age as a continuous variable. Figure 3A provides the risk of stroke for men, when no risk factors other than age were present and shows that the risk increased steadily with age. When an additional risk factor, hypertension was present in addition to age, the risk increased, mostly for men more than 65 years of age (Figure 3B). Having a history of previous stroke as the only additional risk factor gave the highest stroke risk regardless of age (Figure 3C), and being younger with a previous stroke approximately equalled the risk of stroke having three risk factors, i.e. heart failure, hypertension, and diabetes in an older patient (Figure 3D).

Figure 4 shows the 1-year absolute stroke risk for all combinations of the risk factors resulting in a CHA₂DS₂-VASc of 0, 1, and 2 in specific age groups. Other than prior stroke, age was the most important determinant of the estimated stroke risk. For a CHA₂DS₂-VASc score of 1, the absolute stroke risk for the different age groups varied from 0.4% to 1.8%, and the absolute risk ranged from 0.5% to 14.1% for the CHA₂DS₂-VASc score of 2. Depending on the additional risk factors besides sex, women with a CHA₂DS₂-VASc score of 2 had a similar or lower stroke and thromboembolism risk compared to men with a score of 2 (excluding prior stroke). To estimate the personalized risk of stroke, CARS is available online, where all the possible combinations of the CHA₂DS₂-VASc score can be calculated (https://hjerteforeningen.shinyapps.io/riskvisrr/).

Sensitivity analysis

Employing the definitions of outcome used in the sensitivity analyses, 7395 (5%) of the patients had a stroke excluding transient ischaemic attacks, and 3262 (2%) had a cerebral infarction (specific stroke definition). When comparing the results for the outcome stroke vs. thromboembolism (stroke and thromboembolism), the results were similar (Supplementary material online, eFigure 3). Excluding transient ischaemic attacks from the outcome definition of stroke and thromboembolism did not change the results significantly (Supplementary material online, eFigure 4). For the specific outcome of stroke, a lower risk was found compared to the main outcome of stroke definition in the CHA₂DS₂-VASc scores 1 and above (Supplementary material online, eFigure 5).

A sensitivity analysis for heart failure, where the definition only included a diagnosis was done, and no significant differences were found in the outcome of stroke (Supplementary material online, eFigure 6).

Discussion

In this nationwide retrospective cohort study of patients with an initial diagnosis of non-valvular AF and without prior anticoagulation treatment, we found that the CHA₂DS₂-VASc score was less reliably associated with the 1-year risk of stroke and thromboembolism compared to a flexible risk factor-based model. Furthermore, the most clinically important CHA₂DS₂-VASc categories of 0, 1, and 2 encompassed patients with very different risks of stroke. Finally, increasing age was highly correlated with increased risk of stroke and thromboembolism, and the risk factor that correlated with the most significant increase in stroke risk was the history of a prior stroke.

The CHA₂DS₂-VASc score evolved from the Birmingham risk algorithm, was validated in a nationwide cohort,⁶ and has been shown to identify those at lower risk better than the previous CHADS₂ score.²⁵ Debate has focused on value of various the components of the CHA₂DS₂-VASc score, but the score was designed to be reductionist in order to facilitate practical decision-making. As with prior studies,¹² we did not find a substantial increase in risk among women compared to men with no risk factors and in the low age range. Indeed, female sex would be more appropriately regarded as a risk modifier. Vascular disease as an independent risk factor has also been debated, but more recent studies show that vascular disease independently increases stroke in AF.²⁶

Given that clinical risk scores only identify high-risk patients with modest accuracy, and that the CHA₂DS₂-VASc score is most useful for identification of low-risk patients, recent guidelines have focused on initially identifying low-risk patients who do not need any antithrombotic therapy.⁴ Following this step, patients with one or more additional stroke risk factors can be considered for stroke prevention.

We found a large group of patients who, according to their CHA₂DS₂-VASc score, would be considered high risk of a stroke or thromboembolism but were not taking anticoagulants. Despite the focus on identifying higher-risk AF patients who should be prescribed thromboprophylaxis, anticoagulants are often not prescribed or therapy not adhered to, even among patients without any contraindications.^27,28 However, information of why these patients did not receive oral anticoagulation was unavailable. While we cannot ascertain to which extent anticoagulants were prescribed but not redeemed, studies have shown a reluctance to prescribing anticoagulants among physicians which might relate to a lack of trust in the risk stratification.

Existing risk models of stroke in patients with AF stratify patients into low-, intermediate-, and high-risk groups, but these are artificial categories given that risk is a continuum and also varies over time with increasing age and incident comorbidities.²⁹ For optimal risk stratification, the users of simple clinical risk scores need to appreciate the limitations of such a reductionist approach to decision-making. In order to ensure accuracy in the development of risk classification models, complexity cannot be avoided but requires balance against simple, practical use in busy clinical settings. This highlights the importance of a risk factor-based approach, knowing the absolute risk estimates of each CHA₂DS₂-VASc score ranges. Furthermore, the stroke risk in different cohorts can vary, and calibration of prediction models is warranted.³⁰

We demonstrate that the stroke risk varies and overlaps between what is considered a low and intermediate stroke risk for patients with a single non-sex CHA₂DS₂-VASc stroke risk factor, among which management guidelines are inconclusive and recommend that physicians should consider initiating oral anticoagulation therapy (Class IIa recommendation). Indeed, not all CHA₂DS₂-VASc criteria carry equal weight in relation to their stroke risk, so with one point scored on CHA₂DS₂-VASc the potential stroke risk depends on the risk factor relevant to the particular patient.³¹ With our risk calculator CARS, a more personalized approximation of the stroke risk estimates can be achieved for shared decision-making with the patient. For clinicians, our results and CARS can potentially be used as a guide when there is uncertainty regarding the stroke risk.

Strength and limitations

Currently, the estimates of the predicted stroke and thromboembolism risk in patients with AF are based on observational studies with the inherent limitations. The major strengths of our study are the nationwide cohort and the use of validated national health care registers. In Denmark, health care is free of charge and drug prescriptions are partially reimbursed, which limits selection bias. Arguably, over-the-counter use of acetylsalicylic acid might modify stroke risk, but in Denmark, only 8% of purchases of low-dose acetylsalicylic acid are not registered by prescription (33), and acetylsalicylic acid has not been proven effective in stroke prevention in AF. We only included patients admitted with AF, thereby excluded patients with AF diagnosed in the primary care setting. Assuming that primary care patients have inherently lower risks, our model may overestimate the risk of stroke in the low-risk sets. We could not determine why the patients did not receive oral anticoagulation immediate after diagnosis, but it is possible that some of our patients could have contraindications for anticoagulation and as such, some patients could be a selected subset of the general AF population.

Our retrospective study is of an observational nature and subject to potential bias by misclassification. For instance, our validated definition of hypertension has a high positive predictive value⁶ but as the negative predictive value is unknown and patients might thus have been wrongly classified as not having this risk factor. The definition and interpretation of the different CHA₂DS₂-VASc score risk components can vary. For example, definition of vascular disease may contribute to the discrepancies found in different studies of vascular disease as an additional risk factor, as the case-mix across health care systems may differ and not all of those diagnoses carry equal stroke risk. Additionally, our study is most likely to represent a Scandinavian population and therefore might not represent other populations.

In this study, the absolute risk of stroke was used to quantify the individual risk. We also defined our outcome as stroke and thromboembolic risk, but the absolute risk did not change substantially when only looking at the stroke risk. The 1-year absolute risk is the probability that a randomly selected disease-free individual gets the disease within 1 year. Previous studies of predicted stroke risk have estimated the rate of occurrence. Compared to the 1-year absolute risk, the year rate need not to increase with time period (e.g. when considering a 5-year rate instead of a 1-year rate). While this may be a reasonable approximation when the follow-up period is very short, and the study is about healthy individuals, it is not reasonable with regards to mortality. This is relevant given that stroke risk is a dynamic process, and relying on baseline risk factors to predict events many years later is bedevilled by risk changes with increasing age and incident risk factors.³²

Conclusion

For patients with AF, the personalized 1-year risk of stroke varied widely among patients with identical CHA₂DS₂-VASc scores, reflecting large age differences and uneven weights of the stroke risk factors. Calculation of the individual risk using a risk factor-based approach as opposed to using average risk for a particular CHA₂DS₂-VASc score point may improve risk estimates. Furthermore, CARS can be an assisting tool to communicate the stroke risk for a more evidence-based shared decision-making of whether to initiate oral anticoagulation therapy.

Data availability

The data underlying this article cannot be shared publicly due to the privacy of individuals that participated in the study.

Supplementary material

Supplementary material is available at European Heart Journal – Cardiovascular Pharmacotherapy online.

Conflict of interest: P.T.E. is supported by a grant from Bayer AG to the Broad Institute focused on the genetics and therapeutics of cardiovascular diseases. P.T.E. has also served on advisory boards or consulted for Bayer AG, Quest Diagnostics, and Novartis. G.G.: Research grants from Bristol-Myers Squibb, Pfizer, Boehringer Ingelheim, and Bayer. G.Y.H.L.: Consultant for Bayer/Janssen, BMS/Pfizer, Medtronic, Boehringer Ingelheim, Novartis, Verseon, and Daiichi-Sankyo. Speaker for Bayer, BMS/Pfizer, Medtronic, Boehringer Ingelheim, and Daiichi-Sankyo. No fees are directly received personally. J.B.O.: Speaker for Bristol-Myers Squibb, Boehringer Ingelheim, Bayer, and AstraZeneca. Consultant for Bristol-Myers Squibb, Boehringer Ingelheim, Novartis Healthcare, and Novo Nordisk. Funding for research from Bristol-Myers Squibb and The Capital Region of Denmark, Foundation for Health Research. C.T.-P.: Research funding from Bayer and Biotronic. All other authors declared no conflict of interest.

Funding

This work was supported by the Danish Heart Foundation, Copenhagen, Denmark (18-R125-A8552).

References