Association of of Atrial Fibrillation Clinical Phenotypes With Treatment Patterns and Outcomes: A Multicenter Registry Study

Key Points

Question  How different are groupings derived from cluster analysis from traditional classifications of patients with atrial fibrillation?

Findings  This cluster analysis identified 4 statistically driven clusters: (1) a cluster with low comorbid burden; (2) a younger/behavioral disorder cluster; (3) a cluster who resembled patients with tachycardia-bradycardia and had implanted devices; and (4) a cluster defined by atherosclerotic comorbidities. These groupings are not driven by conventional classifications and have different risks for adverse clinical outcomes.

Meaning  Cluster analysis highlighted the considerable heterogeneity among patients with atrial fibrillation; cluster classification may identify differences in phenotypes, variations of clinical practice, and potential targets for improving the quality of care.

Importance  Atrial fibrillation (AF) is usually classified on the basis of the disease subtype. However, this characterization does not capture the full heterogeneity of AF, and a data-driven cluster analysis reveals different possible classifications of patients.

Objective  To characterize patients with AF based on a cluster analysis and to evaluate the association between these phenotypes, treatment, and clinical outcomes.

Design, Setting, and Participants  This cluster analysis used data from an observational cohort that included 9749 patients with AF who had been admitted to 174 US sites participating in the Outcomes Registry for Better Informed Treatment of Atrial Fibrillation (ORBIT-AF) registry. Data analysis was completed from January 2017 to October 2017.

Exposure  Patients with diagnosed AF who were included in the registry.

Main Outcomes and Measures  Composite of major adverse cardiovascular or neurological events and major bleeding, as defined by the International Society of Thrombosis and Hemostasis criteria.

Results  Of 9749 total patients, 4150 (42.6%) were female; 8719 (89.4%) were white and 477 (4.9%) were African American. A cluster analysis was performed using 60 baseline clinical characteristics, and it classified patients with AF into 4 statistically driven clusters: (1) those with considerably lower rates of risk factors and comorbidities than all other clusters (n = 4673); (2) those with AF at younger ages and/or with comorbid behavioral disorders (n = 963); (3) those with AF who had similarities to patients with tachycardia-brachycardia and had device implantation owing to sinus node dysfunction (n = 1651); and (4) those with AF and prior coronary artery disease, myocardial infarction, and/or atherosclerotic comorbidities (n = 2462). Conventional classifications, such as AF subtype and left atrial size, did not drive cluster formation. Compared with the low comorbidity AF cluster, adjusted risks of major adverse cardiovascular or neurological events were significantly higher in the other 3 clusters (behavioral comorbidity cluster: hazard ratio [HR], 1.49; 95% CI, 1.10-2.00; device implantation cluster: HR, 1.39; 95% CI, 1.15-1.68; and atherosclerotic comorbidity cluster: HR, 1.59; 95% CI, 1.31-1.92). For major bleeding, adjusted risks were higher in the behavioral disorder comorbidity cluster (HR, 1.35; 95% CI, 1.05-1.73), those with device implantation (HR, 1.24; 95% CI, 1.05-1.47), and those with atherosclerotic comorbidities (HR, 1.13; 95% CI, 0.96-1.33) compared with the low comorbidity cluster. The same clusters were identified in an external validation in the ORBIT AF II registry.

Conclusions and Relevance  Cluster analysis identified 4 clinically relevant phenotypes of AF that each have distinct associations with clinical outcomes, underscoring the heterogeneity of AF and importance of comorbidities and substrates.

Atrial fibrillation (AF) is associated with increased risks of heart failure, cognitive impairment, thromboembolism, and death. Classification of AF is often based on disease subtype (paroxysmal, persistent, or long-standing persistent/permanent), left atrial size, or thromboembolism risk (eg, the score that a patient receives on the CHA₂DS₂-VASc scale, an instrument designed to predict the risk of stroke in patients with AF).¹ These categories are often used to guide patient management and treatment decisions. However, AF is a very heterogeneous condition caused by a variety of underlying processes and disorders, and it might not be optimally characterized with simplified risk scores.²

Cluster analysis is a data-driven approach commonly used in other fields of research. It measures differences between individual cases on the basis of their characteristics and finds clusters of patients who are more similar to each other than the patients in other clusters. This statistical approach has been applied in a variety of places in medicine, and has been demonstrated to improve characterization of disease phenotypes.^3-7

In a large cohort of patients with AF, we sought to use cluster analysis to identify novel groups of patients with AF who share similar phenotypes and to evaluate whether these cluster phenotypes are associated with different treatment patterns and outcomes.

The Outcomes Registry for Better Informed Treatment of Atrial Fibrillation (ORBIT-AF) is a prospective, nationwide multicenter registry of patients with new-onset and ongoing cases of AF. Patients were enrolled by a diverse group of health care professionals, including internists, cardiologists, and electrophysiologists.⁸ Eligible patients were those 18 years and older who had electrographically documented AF, who were able to provide informed consent, and who adhered to local follow-up. Patients with AF due to a reversible cause (eg, pulmonary embolism, acute thyrotoxicosis, or postoperative status), solitary atrial flutter without AF, or a life expectancy of less than 6 months were excluded.

Health care centers participating in the ORBIT-AF registry used a web-based case report form to record patient demographics, medical history, medications, vital signs, laboratory data, imaging parameters, and electrocardiographic parameters. The data collection process also included the CHA₂DS₂-VASc scale; the Anticoagulation and Risk Factors in Atrial Fibrillation (ATRIA) scale; the ORBIT scale, a 5-item scale that predicts bleeding in patients with AF; and the European Heart Rhythm Association Scale, which quantifies AF symptoms. Follow-up data were collected at approximately 6-month intervals for up to 24 months. Major clinical outcomes assessed at follow-up included all-cause mortality, cardiovascular death, myocardial infarction, hospitalization, stroke or non–central nervous system (CNS) systemic embolism, new-onset heart failure, and bleeding.

This study was approved by the institutional review board at Duke University, which also served as the coordinating center for the study, and each participating center obtained local institutional review board approval. Informed written consent was obtained from each participant.

For the purpose of this analysis, the primary outcome measure was major adverse cardiovascular or neurological events (MACNE), which was defined as a composite of cardiovascular death, myocardial infarction, stroke/non-CNS systemic embolism, or transient ischemic attack. In addition, individual components of MACNE, all-cause death, new-onset heart failure, all-cause hospitalization, cardiovascular hospitalization, bleeding hospitalization, and major bleeding were also assessed. Major bleeding was defined by the International Society of Thrombosis and Hemostasis criteria as follows⁹: fatal bleeding and/or symptomatic bleeding in a critical area or organ and/or bleeding causing a fall in hemoglobin level of 2 g/dL or more or leading to transfusion of 2 or more units of whole blood or red blood cells.

The ORBIT-AF registry enrolled 10 137 patients with AF between June 2010 and August 2011 at 176 sites within the United States. After excluding patients with missing follow-up information (n = 388), the final cohort included 9749 patients from 174 sites.

The statistical methodology of cluster analysis has been described previously.³ The process of identifying patient clusters included multiple steps. First, we selected 60 candidate baseline variables that represented key characteristics of patients with AF (eAppendix 1 in the Supplement). Variable reduction was then performed to group the list of candidate variables into disjoint variable clusters (eAppendix 2 and eMethods in the Supplement). This process was performed separately on continuous and binary variables, resulting in 3 clusters of continuous variables and 11 clusters of categorical variables (eTable 1 in the Supplement).

A summary score for each individual patient was derived based on each of the 14 identified variable clusters. We standardized the 14 scores to have a mean of 0 and an SD of 1.¹⁰ The standardized value measures how many SDs a given data point was above (positive value) or below (negative value) the mean. From this, the Ward minimum variance method of clustering was used to identify patient clusters. A tree diagram is provided to display the semipartial R²obtained at each interaction of the clustering process (Figure 1). Small values of the semipartial R² indicate that the merged clusters were similar, and large values indicate the combination of 2 dissimilar (heterogeneous) clusters. Because this is the first experience to our knowledge with cluster analysis in the population of patients with AF, the determination of the numbers of clusters was not prespecified. The goal was to determine clusters that were homogeneous, based on the semipartial R², and illustrative of particular types of patients, all blinded to outcomes and treatment patterns. Examination of the tree diagram indicated that the groupings became more heterogeneous after being expanded to 6 clusters. Both 6-cluster and 4-cluster models were examined. The 4-cluster model retained a semipartial R² of 0.04 and formed much clearer patterns of patient groups than 6-cluster model. Therefore the 4-cluster model were used in this study.

To evaluate the robustness of the analysis, external validation was performed using the ORBIT-AF II Registry. The ORBIT-AF II Registry has a similar structure to the ORBIT-AF Registry, but enrolled a mutually exclusive population. The enrollment period of the ORBIT-AF II Registry was from February 2013 to April 2016, a more recent period than that of the ORBIT-AF Registry (which enrolled patients from June 2010 through August 2011). In contrast to ORBIT-AF, ORBIT-AF II also included only patients who had either a new diagnosis of AF within the previous 6 months or who had initiated an oral anticoagulant treatment regimen (excluding vitamin K) for AF.

We applied the same clustering technique to the patients enrolled in the ORBIT-AF II Registry (n = 12 679). Similar to the main analysis, we obtained 3 clusters of continuous variables and 11 clusters of categorical variables through the variable reduction process. A semipartial R² of 0.04, which was the same value we used in the main analysis, was used to determine the final number of clusters, resulting in identification of 4 clusters.

Baseline characteristics were compared among 4 identified clusters. The key characteristics that differentiated the clusters were determined on the basis of the variables that were markedly different among clusters. Conventional AF classifications, including degree of left atrial enlargement, AF diagnostic subtype, and CHA₂DS₂-VASc risk score, were also compared for the 4 clusters. Continuous variables are presented as medians with interquartile ranges or means with SDs, and the differences across clusters were assessed using the Kruskal-Wallis test. Categorical variables are reported as counts and percentages, and differences were assessed using the χ² test.

The association between clusters and clinical outcomes was assessed using the unadjusted and adjusted Cox proportional hazards models. Models were adjusted by CHA₂DS₂-VASc score for MACNE, cardiovascular death, and thromboembolic event outcomes or by ORBIT bleeding score for major bleeding and bleeding hospitalization outcomes. Models were also adjusted for medications (anticoagulant agent, antiplatelet agent, angiotensin-converting enzyme inhibitor, angiotensin 2 receptor blocker, and β-blocker, statin, and antiarrhythmic agent) and the specialty of the physician treating the patient (internal medicine/primary care, cardiology, electrophysiology, and neurology). The hazard ratio (HR) for each cluster is presented along with 95% confidence intervals and corresponding P values. For MACNE, a Kaplan-Meier plot with the log-rank test was also derived and assessed.

All variables had less than 5% missing data, except for estimated glomerular filtration rate (7.3%), hematocrit (10.2%), left ventricular ejection fraction (LVEF) (10.0%), and left atrial diameter (13.9%). To account for missing data, multiple imputation was used. Imputed values were obtained by the Markov chain Monte Carlo method or regression methods. The deidentified aggregate data were analyzed by the Duke Clinical Research Institute using SAS software (version 9.3; SAS Institute). Data analysis was completed from January 2017 to October 2017.

Clinical variables according to the 4 identified clusters are detailed in Table 1 and eTable 2 in the Supplement. Key characteristics of each patient cluster were as follows.

The low comorbidity cluster (n = 4673) was the largest cluster, with approximately twice as many patients as any other cluster. Patients in the low comorbidity cluster were the second youngest cluster, with a median age of 73 years (interquartile range [IQR], 65-81 years), and had considerably lower rates of risk factors and comorbidities than all other clusters, including the lowest rates of diabetes (1129/4673; 24.2%), sleep apnea (798/4673; 17.1%), prior myocardial infarction (108/4673; 2.3%), heart failure (921/4673; 19.7%), chronic obstructive pulmonary disease (518/4673; 11.1%), and prior stroke or transient ischemic attack (564/4673; 12.1%). Reflecting the relatively good health status of the people in this cluster, their median LVEF (60%; interquartile range [IQR], 55%-65%) was higher than those in any other clusters (younger/behavioral disorder cluster: 55%; IQR, 50-60; device implantation cluster: 55%; IQR, 50-60; atherosclerotic-comorbid cluster: 55%; IQR, 40-60) (P < .001). They also had the lowest mean (SD) CHA₂DS₂-VASc and ORBIT bleeding scores, which were 3.39 (1.69) and 1.95 (1.51), respectively, and the second-lowest mean (SD) ATRIA bleeding score, which was 2.61 (1.96).

These patients (n = 963) were, on average, the youngest (median age, 69 years; IQR, 61-77) and the most likely to be male (700/963; 72.7%). The key and distinguishing characteristics of this cluster were the highest rates of liver disease (169/963; 17.6%), alcohol abuse (356/963; 37.0%), drug abuse (118/963; 12.3%), and current smoking (485/963; 50.4%). Patients with any 1 of these medical conditions were categorized exclusively in the younger/behavioral disorder cluster. They also had the highest body mass index (calculated as weight in kilograms divided by height in meters squared) among the 4 clusters (median, 30; IQR, 26-35). While patients in the younger/behavioral disorder cluster also had the highest rates of chronic obstructive pulmonary disease (275/963; 28.6%) and obstructive sleep apnea (220/963; 22.9%), the prevalence rates of nonpulmonary comorbidities were relatively low. Notably, their renal function was the most normal of the 4 clusters, as measured by estimated glomerular filtration rate (median, 73.3 mL/min/1.73 m²; IQR, 57.7-88.2 mL/min/1.73 m²). Reflecting the younger age of patients in this cluster, the risk scores were relatively lower, and they had the lowest mean (SD) ATRIA bleeding score (2.44 [1.96]), and the second-lowest mean (SD) CHA₂DS₂-VASc and ORBIT scores, which were 3.42 (1.75) and 1.98 (1.59), respectively.

The 1651 patients in this cluster were characterized by their resemblance to patients with tachycardia-bradycardia. The unique characteristic of these patients appeared to be their high prevalence of device implantation owing to the sinus node dysfunction or atrioventricular node ablation, which was present in 1504 of the 1650 participants for whom data was available (91.2%). The majority of implanted cardiac electronic devices in this cluster were pacemakers. Patients in this cluster also exhibited the highest European Heart Rhythm Association symptom scores (eAppendix 1 in the Supplement) and 313 of 1644 patients for whom these data were available (18.9%) reported either severe AF symptoms (276/1644; 16.8%) or disabling AF symptoms (37/1644; 2.3%) that affected or discontinued their normal daily activities.¹¹ They had the highest mean (SD) ATRIA bleeding scores (3.75 [2.11]), and the second-highest mean (SD) CHA₂DS₂-VASc and ORBIT bleeding scores (4.57 [1.61] and 2.83 [1.66], respectively).

This cluster (n = 2462) included a higher percentage of male (1694/2462; 68.8%) patients and a higher median age (77 years; IQR, 69-82 years). A major feature of patients in the atherosclerotic-comorbid AF cluster is that 1973 of these 2462 patients (80.1%) had history of coronary artery disease, and they also had the highest prevalence of prior myocardial infarction (982/2462; 39.9%) and coronary revascularization, whether surgical (884/2462; 35.9%) or percutaneous (1042/2462; 42.3%). They also had the lowest LVEF (median, 55%; IQR, 40-60) among the 4 clusters and highest rates of heart failure (1273/2462; 51.7%). They also had the highest prevalence of most other risk factors and comorbidities, including hypertension (2137/2462; 86.8%), hyperlipidemia (2092/2462; 85.0%), diabetes (930/2462; 37.8%), chronic kidney disease (1088/2333; 46.6%), anemia (660/2462; 26.8%), prior gastrointestinal bleeding (360/2462; 14.6%), and peripheral vascular disease (516/2462; 21.0%). Reflecting the presence of many comorbidities, they had the highest mean (SD) CHA₂DS₂-VASc and ORBIT scores (4.75 [1.55] and 2.99 [1.73], respectively) and had the second highest mean (SD) ATRIA bleeding score (3.65 [2.26]).

We examined the relationship between the 4 identified clusters and conventional AF classifications including degree of left atrial enlargement, AF subtype, and CHA₂DS₂-VASc score. The differences in the distribution of left atrial size and AF subtype were not clinically meaningful, whereas the CHA₂DS₂-VASc scores did vary significantly across the clusters (P < .001; details appear in eFigure 1 in the Supplement). In the device implantation cluster and atherosclerotic-comorbid cluster, the proportions of patients with a risk score of 4 or higher were each approximately 75% (per eTable 2 in the Supplement); however, those in the low-comorbidity cluster and younger/behavioral disorder cluster were less than 50%.

Table 2 lists the differences in treatment patterns by clusters. With respect to management strategy, patients in the device implantation cluster were more likely to be managed with a rate control strategy than those in the other 3 clusters (1169/1651 or 70.9%, vs 66.6% [95% CI, 65.3%-68.0%] in the low comorbidity cluster, 67.9% [95% CI, 65.0%-70.9%] in the younger/behavioral disorder cluster, and 69.9% [95% CI, 68.0%-71.7%] in the atherosclerotic-comorbid cluster; P = .003). Class 1c antiarrhythmic medications were more frequently prescribed to those patients in the low comorbidity cluster, whereas sotalol was most commonly prescribed to the patients in the device implantation cluster (to 121/1651 patients, or 7.6% [95% CI, 6.4%-8.9%], compared with 6.3% [95% CI, 5.6%-7.0%] in the low comorbidity cluster, 5.5% [95% CI, 4.1%-6.9%] in the younger/behavioral disorder cluster, and 4.9% [95% CI, 4.1%-5.8%] in the atherosclerotic-comorbid cluster; P = .003). Reflecting the highest prevalence of coronary artery disease and heart failure, patients in the atherosclerotic-comorbid cluster were more likely to be receiving β-blockers, statins, diuretics, angiotensin-converting enzyme inhibitors, and antiplatelet agents (aspirin and/or clopidgrel) than those in the 3 other clusters. In contrast, among patients who have a CHA₂DS₂-VASc score of 2 or higher and therefore should be undergoing anticoagulation therapy, patients in the younger/behavioral disorder cluster and those in the atherosclerotic-comorbid cluster were less likely to be treated with an oral anticoagulant (74.5%; 95% CI, 71.5%-77.4% and 76.5%; 95% CI, 74.8%-78.2%, respectively) than those in the low comorbidity cluster and those in the device implantation cluster (79.3%; 95% CI, 78.1%-80.6% and 79.9%; 95% CI, 77.9%-81.8%)(P < .001).

The occurrence of MACNE across the different clusters is illustrated in Kaplan-Meier curves in Figure 2. The occurrence of MACNE was different across the 4 clusters. Patients in the low comorbidity cluster had the lowest risk of MACNE (2.58 events per 100 patient-years), followed in ascending order by those in the younger/behavioral disorder cluster (3.97 events per 100 patient-years), those in the device implantation cluster (5.10 events per 100 patient-years), and those with atherosclerotic-comorbid AF, who were at highest risk (6.12 events per 100 patient-years; P < .001). A similar trend was observed in each individual component of the MACNE composite (Table 3). In addition, the differences in outcomes across clusters were still apparent after adjustment for CHA₂DS₂-VASc scores, medications, and provider subspecialty (P < .001; eFigure 2 in the Supplement). Compared with the low comorbidity cluster, the adjusted risks of MACNE were significantly higher in the 3 other clusters (younger/behavioral disorder cluster: HR, 1.49; 95% CI, 1.10-2.00; device implantation cluster: HR, 1.39; 95% CI, 1.15-1.68; and the atherosclerotic comorbid cluster: HR, 1.59; 95% CI, 1.31-1.92).

For major bleeding and bleeding hospitalization events, patients in the device implantation cluster exhibited the highest risk (Table 3). Patients in the low comorbidity cluster consistently demonstrated the lowest risk for all outcomes. After adjusting for the ORBIT bleeding score, medications, and physician subspecialty, compared with the low comorbidity cluster, the adjusted risks were generally higher in the 3 other clusters, and patients in the device implantation cluster had the highest risk profile (HR, 1.24; 95% CI, 1.05-1.47 for major bleeding; HR, 1.37; 95% CI, 1.14-1.64 for bleeding hospitalization) (eFigure 2 in the Supplement).

Clinical variables according to the 4 identified clusters from the external validation using the ORBIT-AF II Registry are detailed in eTable 1 and eTable 3 in the Supplement. The identified clusters from the validation cohort were almost identical to clusters we identified in the main analysis.

We applied cluster analysis in a large cohort of patients with AF to identify alternative and informative clinical phenotyping, and there are 4 major findings. First, this cluster analysis led to the identification of 4 clinically relevant and recognizable phenotypes. Second, conventional measures of AF duration, including AF disease subtype (paroxysmal, persistent, or long-standing persistent/permanent) and left atrial size, did not drive cluster formation. Third, treatment patterns, including rhythm control and oral anticoagulation, varied among the clusters. Finally, these distinct clusters were associated with different risk for major cardiovascular and neurologic adverse events.

Using a phenotype-mapping algorithm, we were able to take advantage of the detailed patient-level variables in our AF cohort and find patterns of association, which allowed a novel grouping of patients with AF. Surprisingly, conventional measures of AF, including AF subtype, were not defining features of clusters. The identified clusters often emphasized overlooked variables in conventional classifications, such the presence of cardiac-implanted electronic devices, or the combination of liver disease, alcohol abuse, drug abuse, and current smoking. While persistent forms of AF are associated with worse outcomes,¹² classification of AF disease subtype has several limitations, including its dynamic nature¹ and difficulties with quantification of each type.¹³ These newly described cluster phenotypes confirm that AF is a very heterogeneous condition caused by a variety of underlying conditions, and its diversity is not always well described using conventional classifications.

The clusters did resemble several recognizable groups of patients with AF. For example, patients in the cluster dominated by the presence of cardiac-implanted electronic devices resembled patients with tachycardia-bradycardia syndrome and more frequently had AF with accompanying sinus node dysfunction requiring pacemaker implantation. Patients in the atherosclerotic-comorbid cluster were much more likely to have AF complicated by ischemic heart disease. They exhibited the highest rates of comorbidities, including coronary artery disease, prior myocardial infarction or revascularization, and heart failure due to reduced left ventricular systolic function. Ischemic heart disease is a known risk factor for AF. Left atrial ischemia increases focal drivers and perpetuates AF.¹⁴ Unsurprisingly, patients in the atherosclerotic-comorbid cluster also had high rates of cardiovascular risk factors, including hypertension, dyslipidemia, diabetes, and chronic kidney disease. Since cardiovascular risks factors are often modifiable and coronary artery disease can be prevented, the identification of poor outcomes in the atherosclerotic-comorbid cluster highlights the importance of measures to prevent AF and its contributory states. In addition, the key characteristics of patients in the younger/behavioral disorder cluster were liver disease, alcohol abuse, drug abuse, current smoking, and obesity, which is a reasonable combination of patient characteristics that was not anticipated before clustering. Liver disease, alcohol abuse, and drug abuse are variables included in the HAS-BLED bleeding score,¹⁵ as are validated risk factors for bleeding events. Also, Angoulvant et al¹⁶ demonstrated that active smoking was independently associated with a higher risk of major bleeding. The results of this cluster analysis also suggest that smokers are at an increased risk for adverse cardiovascular events.

Cluster analysis can be helpful by identifying unique phenotypes and patient characteristics within heterogeneous populations, such as patients with AF. It also can elucidate differences and variation in practice patterns across a population. Given the increased risk of bleeding with concomitant use of antiplatelet and anticoagulant agents,¹⁷ high-rates of antiplatelet therapy might explain the underutilization of anticoagulation therapy in atherosclerotic-comorbid AF. Although individual bleeding risk and thromboembolic risks need to be balanced, guideline-recommended anticoagulation should be properly implemented to reduce stroke and optimize patient outcomes. Current guidelines by the American College of Cardiology, American Heart Association Task Force on Practice Guidelines, and the Heart Rhythm Society recommend anticoagulation therapy in patients with AF who have a CHA₂DS₂-VASc score or 2 or more, regardless of antiplatelet prescription.¹ Similarly, in the European Society of Cardiology guidelines, antiplatelet monotherapy is not recommended for stroke prevention in patients with AF, regardless of stroke risk.¹⁸ In contrast, the suboptimal use of anticoagulation therapy in the younger-behavioral disorder cluster of patients with AF may be explained by the preponderance of certain bleeding risk factors. Although the mean ATRIA and ORBIT bleeding risk scores were relatively lower, the high rates of alcohol and/or drug abuse and liver disease are the notable features in this cluster and known to be associated with bleeding events.¹⁵ These examples illustrate how cluster analysis might identify potential areas for improving quality of care in management of AF among patients with similar thromboembolic risk.

Clarifying the distinguishing characteristics and key components of each cluster may help us understand how we should intervene clinically to improve outcomes within these phenotypes. For example, given that the atherosclerotic comorbid AF cluster had the highest ischemic event rates despite a reasonably high anticoagulation use, risk factor modification and/or treatment of residual risk would be an optimal clinical approach. For patients in the younger/behavioral disorder cluster, lifestyle modification might be the most effective means to improve outcomes. Moreover, the data suggest that appropriate blood pressure management should be universally emphasized, given the large proportion of hypertension even in patients classified as having low-comorbidity AF.

In future work, the efficacy of these hypotheses, interventions, and potential treatment heterogeneity should be evaluated and tested. Furthermore, clarifying the interaction between clusters and treatment, such as anticoagulation therapy and ablation or lifestyle modification of stroke risk, could substantially enhance the clinical implications of phenotype cluster analysis.

Our findings should be interpreted in the context of several potential limitations. First, we consider this analysis to be hypothesis generating. We are not proposing a formal classification system for AF. Our objective was to demonstrate that application of a relatively novel statistical approach to a large population of patients with AF can yield clinically relevant patient groupings that highlight factors that may be overlooked and absent from conventional classifications. Such findings may lead to improvements in AF classification. Second, clustering algorithm results are dependent on the underlying population, available clinical variables, and associated patterns of care in the community. Thus, findings might differ if additional variables are available or included. Although ORBIT-AF is a nationwide registry of patients with AF cared for by a diverse group of health care professionals, our registry was not a population sample, but rather relied on voluntary participation from sites and patients. Thus, the patients enrolled might not be entirely representative of the broader populations. Therefore, a clustering algorithm might not be able to identify the similar clusters and may instead clarify different type of clusters (eg, a valvular-related cluster) in other, more diverse groups. Generalizability of our findings should be explored in additional cohorts.

Finally, the ultimate selection of 4 clusters was based in part on investigator discretion. The incorporation of more clusters might yield differing groupings and clinical outcomes.

This analysis demonstrates that unsupervised cluster analysis can identify meaningful, clinically relevant categories and phenotypes of patients with AF. These clusters exhibit significant differences in underlying comorbidities and differential risk of adverse outcomes. Cluster phenotypes also demonstrated independent prognostic value, especially for cardiovascular outcomes, even after accounting for CHA₂DS₂-VASc risk scores. Our results highlight the high degree of heterogeneity within the population of patients with AF, and the possibility that cluster classification can help identify important differences in AF subgroups and subsequent health outcomes. Such insight may inform treatment decisions and/or inform future studies of this disease.

Corresponding Author: Jonathan P. Piccini, MD, MHS, Electrophysiology Section, Duke Center for Atrial Fibrillation, Duke University Medical Center, Duke Clinical Research Institute, PO Box 17969, Durham, NC 27710 (jonathan.piccini@duke.edu).

Accepted for Publication: October 23, 2017.

Published Online: November 12, 2017. doi:10.1001/jamacardio.2017.4665

Author Contributions: Drs Inohara and Piccini had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Drs Inohara and Piccini contributed equally to the study.

Study concept and design: Inohara, Pieper, Blanco, Freeman, Allen, Fonarow, Ezekowitz, Kowey, Reiffel, Naccarelli, Peterson, Piccini.

Acquisition, analysis, or interpretation of data: Inohara, Shrader, Pieper, Thomas, Singer, Allen, Fonarow, Gersh, Kowey, Reiffel, Naccarelli, Chan, Steinberg, Peterson, Piccini.

Drafting of the manuscript: Inohara, Shrader, Blanco, Reiffel.

Critical revision of the manuscript for important intellectual content: Shrader, Pieper, Thomas, Singer, Freeman, Allen, Fonarow, Gersh, Ezekowitz, Kowey, Reiffel, Naccarelli, Chan, Steinberg, Peterson, Piccini.

Statistical analysis: Shrader, Pieper, Thomas, Peterson.

Obtained funding: Blanco, Piccini.

Administrative, technical, or material support: Blanco, Naccarelli, Steinberg, Peterson, Piccini.

Study supervision: Pieper, Blanco, Allen, Ezekowitz, Kowey, Steinberg, Peterson, Piccini.

Conflict of Interest Disclosures: All authors have completed and submitted the ICMJE Form for Disclosure of Potential Conflicts of Interest. Dr Inohara reports receiving grants from JSPS during the conduct of the study, plus grants from Pfizer Health Research Foundation and Miyata Cardiac Research Promotion Foundation outside the submitted work. Dr Thomas has participated in research with Novartis, Boston Scientific, Gilead Sciences Inc, and Janssen Scientific. Dr Singer has acted as a consultant and advisory board member for Boehringer Ingelheim, Bristol-Myers Squibb, Merck, Johnson & Johnson, Pfizer, and Medtronic. Dr Singer also reports research grants from Boehringer Ingelheim and Bristol-Myers Squibb, personal fees from Johnson & Johnson during the conduct of the study, and grants from Boehringer Ingelheim, Bristol-Myers Squibb, and Medtronic, plus personal fees from Boehringer Ingelheim, Bristol-Myers Squibb, CVS Health, Johnson & Johnson, Merck, and Pfizer outside the submitted work. Dr Freeman has acted as a consultant and advisory board member for Janssen Scientific; he also reports personal fees from Janssen Pharmaceuticals and American College of Cardiology National Cardiovascular Data Registry outside the submitted work. Dr Allen has received grants from Patient-Centered Outcomes Research Institute, American Heart Association, and National Institutes of Health National Heart, Lung, and Blood Institute as well as personal fees from Novartis, Janssen, and Boston Scientific outside the submitted work. Dr Fonarow has acted as a consultant and advisory board member for Janssen Pharmaceuticals and reports personal fees from Janssen Pharmaceuticals and St. Jude outside the submitted work. Dr Gersh is a member of a data safety monitoring board for Mount Sinai St. Luke’s, Boston Scientific Corporation, Teva Pharmaceutical Industries, St. Jude Medical, Janssen Research & Development, Baxter Healthcare Corporation, and Cardiovascular Research Foundation; he has also acted as a consultant or advisory board member for Janssen Scientific Affairs, Cipla Limited, Armetheon Inc, and Medtronic and reports personal fees from Mount Sinai St. Luke’s, Boston Scientific Corporation, Teva Pharmaceutical Industries, St. Jude Medical, Janssen Research & Development, Baxter Healthcare Corporation, Cardiovascular Research Foundation, Janssen Scientific Affairs, Cipla Limited, Armetheon Inc, and Medtronic during the conduct of the study. Dr Ezekowitz has acted as a consultant or advisory board member for Boehringer Ingelheim, Daiichi Sankyo, Pfizer, Bristol-Myers Squibb, and Janssen Scientific Affairs; reports personal fees from Bristol-Myers Squibb, Boehringer Ingelheim, Armetheon, Pfizer, Sanofi, Portola, Daiichi Sankyo, Medtronics, Johnson & Johnson, and Janssen Scientific Affairs; and has received grants from Pfizer and BMS. Dr Kowey has provided ad hoc consulting services to Johnson & Johnson for several related and nonrelated projects. Dr Reiffel has received a research grant from Janssen Pharmaceuticals and research support from Boehringer Ingelheim Pharmaceuticals Inc and GlaxoSmithKline; he also reports acting as a consultant for Sanofi, Gilead Sciences Inc, CV Therapeutics, GlaxoSmithKline, Merck, Cardiome Pharma Corp, Boehringer Ingelheim Pharmaceuticals Inc, and Medtronic Inc. Dr Reiffel has received speakers’ bureau income from Sanofi and Boehringer Ingelheim Pharmaceuticals Inc. Dr Naccarelli has received a research grant from Janssen; acted as a consultant or advisory board member for Janssen and Daiichi Sankyo; received personal fees from Janssen during the conduct of the study; and reports receiving grants from Janssen, and personal fees from Janssen, GlaxoSmithKline, and Daiichi Sankyo outside the submitted work. Dr Chan is an employee of Janssen, as well as a consultant for Optum Rx and Johnson & Johnson. Dr Steinberg reports grants from Janssen Pharmaceuticals during the conduct of the study and grants and other support from Janssen Pharmaceuticals, as well as other support from BMS-Pfizer and Bayer outside the submitted work. Dr Peterson reports receiving a research grant from Janssen Pharmaceuticals and Eli Lilly and is a consultant for Janssen Pharmaceuticals and Boehringer Ingelheim. Dr Piccini reports receiving a research grant from the Agency for Healthcare Research and Quality, ARCA Biopharma, Boston Scientific, Gilead Sciences, Janssen Pharmaceuticals, Johnson & Johnson, ResMed, Spectranetics, and St Jude Medical; and acting as a consultant and advisory board member for BMS/Pfizer, GlaxoSmithKline, Janssen Pharmaceuticals, Johnson and Johnson, Medtronic, and Spectranetics. No other disclosures were reported.

Funding/Support: This project was supported in part by cooperative agreement 1U19 HS021092 from the Agency of Healthcare Research and Quality and JSPS Overseas Research Fellowship. The Outcomes Registry for Better Informed Treatment of Atrial Fibrillation is sponsored by Janssen Scientific Affairs LLC.

Role of the Funder/Sponsor: With the exception of Janssen Scientific Affairs LLC, the funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. Janssen Scientific Affairs LLC had no role in the design and conduct of this study, but hosted the electronic data collection system, while the executive committee of the registry was responsible for the design of the registry; they also reviewed the publication, but did not participate in the drafting or critical revision of the manuscript.

Disclaimer: Drs Fonarow and Thomas are the Associate Editor for Health Care Quality and Guidelines and the Assistant Editor for Statistics of JAMA Cardiology but did not participate in any discussions or decisions regarding publication of this manuscript.

Additional Contributions: The Duke Clinical Research Institute conducted the operations, site management, data analysis, and interpretation of the data of this study, in conjunction with the executive committee. We thank the Outcomes Registry for Better Informed Treatment of Atrial Fibrillation Patients and Investigators; specific individuals are acknowledged in the eAcknowledgements in the Supplement.