Unbiased Characterization of Atrial Fibrillation Phenotypic Architecture Provides Insight Into Genetic Liability and Clinically Relevant Outcomes

Abstract

BACKGROUND:

Atrial fibrillation (AF) is a common and clinically heterogeneous arrhythmia. Machine learning algorithms can define data-driven disease subtypes in an unbiased fashion, but whether these AF subgroups align with underlying mechanisms, such as polygenic liability to AF or inflammation, and associate with clinical outcomes is unclear.

METHODS:

We identified individuals with AF in a large biobank linked to electronic health records and genome-wide genotyping. We applied an unsupervised coclustering machine learning algorithm to 35 curated and uncorrelated clinical features to identify distinct phenotypic AF clusters. The clinical inflammatory status of the clusters was defined using measured biomarkers (CRP, ESR, WBC, Neutrophil %, Platelet count, RDW) within 6 months of first AF mention. Polygenic risk scores for AF and for cytokine levels were used to assess the genetic liability of clusters to AF and inflammation, respectively. Clinical outcomes were collected from electronic health records up to the last medical contact.

RESULTS:

The analysis included 23 271 subjects with AF, of which 6023 had available genome-wide genotyping. The machine learning algorithm identified 3 phenotypic clusters that were distinguished by increasing prevalence of comorbidities, particularly renal disease and coronary artery disease. Polygenic liability to AF across clusters was highest in the low comorbidity cluster. Clinically measured inflammatory biomarkers were highest in the high comorbidity cluster. There was no difference between groups in genetically predicted levels of inflammatory biomarkers. Cluster assignment was associated with mortality, stroke, bleeding, and use of cardiac implantable electronic devices after AF diagnosis.

CONCLUSIONS:

Patients within AF subgroups identified by clustering were distinguished by comorbidity burden and associated with risk of clinically important outcomes, polygenic liability to AF, and clinical inflammation, but not with genetically predicted inflammatory cytokine levels. Our study empirically demonstrates the complementary roles of comorbidity and genetic liability as major drivers of AF phenotypic variability using hypothesis-free methods.

Atrial fibrillation (AF) is the most common arrhythmia encountered in clinical practice,¹ and is associated with increased mortality,² stroke,³ heart failure,³ myocardial infarction,^4,5 dementia,⁶ and decreased quality of life.⁷ AF is a complex, dynamic, and heterogeneous condition, and there is considerable variation in the age at onset, arrhythmia burden, presence of risk factors, severity of symptoms, and incidence of complications among affected individuals.^8–18 Heterogeneity is likewise evident in response to AF treatments, including antiarrhythmic drugs and catheter-based ablation.^19,20 This interindividual variation suggests complex gene-environment interactions and the existence of multiple AF subtypes with differing mechanistic underpinnings.

Two sources of risk that may contribute to AF heterogeneity, and by extension, AF subgroups, are genetic susceptibility to AF and inflammation, which itself may be due to acquired or genetic factors. Indeed, polygenic liability conferred by common variants is associated with the risk of AF and multiple clinical outcomes.^21–26 Additionally, multiple lines of evidence support a significant role for inflammation in the pathophysiology of AF. For example, elevated levels of inflammatory biomarkers²⁷ and polymorphisms in genes encoding inflammatory cytokines^28–31 are associated with AF risk, and inflammatory infiltrates are observed in atria of patients with AF.³² Moreover, many AF risk factors (eg, heart failure, coronary artery disease [CAD], hypertension, obesity) and biologic mechanisms (eg, fibrosis, thrombogenesis) are themselves associated with increased inflammation,^25,26,33 and AF itself can promote inflammation and perpetuate atrial remodeling and arrhythmia.^34,35 However, the contributions of genetic susceptibility to AF and inflammation, both acquired and genetically driven, to AF phenotypic subgroups have not been reported.

The emergence of advanced machine learning methods has facilitated data-driven subtyping of complex, heterogeneous cardiovascular diseases, such as hypertension,³⁶ heart failure,³⁷ and pulmonary arterial hypertension.³⁸ Machine learning -derived phenotypic AF subgroups have been identified in posthoc analyses in observational registries.^39,40 These studies, however, did not explore the mechanistic or genetic basis of subgroups. Therefore, there is no consensus regarding the classification of AF subtypes based on their underlying mechanisms, particularly with respect to patients in clinical care.^41,42

We hypothesized that contemporary machine learning methods applied to a large clinical AF population would identify distinct subtypes with different polygenic liability to AF, inflammatory state, and clinical outcomes.

Methods

The study protocol was approved by the Vanderbilt University Medical Center institutional review board (no. 181403). Because of the sensitive nature of the data collected for this study, as well as the required privacy regulations of the data source used in this study to safeguard the rights of the research participants, requests to access the data set from qualified researchers trained in human subject confidentiality protocols may be sent to the corresponding author. All methods and code used in this study are available from the corresponding author on reasonable request.

Results

Cohort Characteristics

We identified 23 271 subjects with AF meeting the inclusion criteria (Figure S1). The cohort median age was 68 years (interquartile range [IQR], 59–76), 42.6% were female, and 90.4% self-reported as White. The most prevalent comorbidity was hypertension (62.3%), followed by renal disease (36.8%), CAD (35.8%), and fluid-electrolyte disorders (25.9%). Heart failure was also common (18.4%), and 12.1% of patients had a history of left ventricular systolic dysfunction (Table 1). The type of AF on first AF mention was able to be determined in 14 035 participants. Of those with AF type available, the majority had paroxysmal AF (9878 or 70.4%). Information on the rank order of AF diagnosis during visits associated with first AF mention was available in 14 894 participants, and AF was the primary code in 6095 participants (40.9%) for these visits. Of those, 693 participants had AF first mention capture as the primary visit code during an inpatient visit.

Table 1.

Baseline Characteristics of Included Participants at Time of First AF Mention in EHR

The median follow-up after AF first mention was 5.4 years (range, 0–26.1 years). Mortality during follow-up was 14%, with a median time to death of 4.5 years (IQR, 1.6–7.8) after first AF mention. Approximately half of the subjects (55.2%) received antiarrhythmic drug therapy, usually initiated shortly after diagnosis (median, 14.6 days; IQR, 0.0–401.5). Cardioversion was performed in approximately one-fifth of the subjects (19.9%), and AF ablation in 8.1%. Device placement was common, with pacemaker implantation in 13.8% of subjects and 9.3% undergoing implantable cardioverter defibrillator implantation. Stroke occurred in 12.3% of subjects, while bleeding events were rare (Table 2).

Table 2.

Outcomes of the Included Participants After First AF Mention in EHR

Feature Exploration

To gain additional insight into the phenotypic architecture of our cohort before clustering, we conducted a series of exploratory analyses. We first assessed pairwise correlations between features to identify patterns among variables, detect pairs of highly correlated features that could affect clustering, and inform a priori estimates of the expected number of feature groups derived by clustering. The selected features were not highly correlated (all r<0.5), and all were retained for clustering analyses (Figure 1A).

Figure 1.

Overall exploration and correlation of clinical variables used as predictors in the clustering algorithm. A, Observed correlation between clinical predictors used for clustering. There were no highly correlated features (r>0.5). B, Principal component (PC) analysis of predictors reveals PC1 is mainly determined by comorbidity burden, while PC2 is mostly determined by age at time of atrial fibrillation (AF) diagnosis and body mass index (BMI). There is an association between PC1 and death during follow-up (P<0.001). CAD indicates coronary artery disease; COPD, chronic obstructive pulmonary disease; PAD, peripheral arterial disease; LVSD, left ventricular systolic dysfunction; ICD, implantable cardioverter defibrillator; OSA, obstructive sleep apnea; DM, diabetes; and HTN, hypertension.

We then conducted data dimensionality reduction using principal component analysis and evaluation of variable contributions to components using a loading plot (Figure 1B). The loading plot revealed that comorbidity burden, particularly left ventricular systolic dysfunction, fluid-electrolyte disorders, CAD, and renal disease, were the main contributors to principal component (PC) 1, while age group was the main contributor for PC2. Analysis of PC associations with mortality revealed that PC1 values were associated with increased risk of death (PC1 interquartile odds ratio, 1.99 [95% CI, 1.91–2.07]).

Unbiased Clustering of Features at AF Diagnosis

Clustering was performed using a coclustering algorithm, where subjects and features are considered simultaneously to identify blocks of subjects with similar characteristics.^43,44 In this framework, the 2 key parameters that need to be determined a priori and prespecified to the algorithm are the number of subject clusters and the number of feature clusters. We determined both parameters using complementary evaluations.

To determine the number of subject clusters, visualization using nonlinear dimensionality reduction with Uniform Manifold Approximation and Projection suggested >2 clusters, with 3 or 4 being the most likely number of clusters (Figure S2). We then compared size and participant distribution and contribution by increasing the subject cluster number (Figure S3) from 2 to 6 clusters. When increasing subject cluster numbers from 2 to 4 clusters yielded stable subject assignments and cluster size. Although 5 and 6 clusters produced very small, unstable groups (Figure S3). Quantitative metrics (Silhouette score, Dunn index) also peaked at 2 subject clusters and declined thereafter (Table S1). Taken together, we selected 3 subject clusters as the most balanced solution between statistical evidence and clinical interpretability. For the number of feature clusters, we explored 2 to 5 clusters based on prior feature exploration results (Figure 1A). Across these settings, >94% of subjects remained assigned to the same subject cluster regardless of the number of feature clusters, indicating that the impact of this parameter was minimal. We, therefore, specified 3 feature clusters, also supported by visual grouping of feature presence in Figure 1A.

After clustering was performed, the clusters were distinguished by increasing burden of comorbidities (Figure 2A), particularly renal disease, fluid-electrolyte disorders, and CAD, as shown in Figure 2B. Based on comorbidity burden, we labeled the groups as low, mid, and high comorbidity clusters. Table S2 and Figure S4 show that the distribution of other comorbidities demonstrated a similar pattern across clusters, but the magnitude was less pronounced than renal disease, CAD, and fluid-electrolyte disorders.

Figure 2.

Clustering assignment and feature prevalence at atrial fibrillation (AF) diagnosis time by cluster. A, Heatmap of cluster assignment by participant and feature. Each column represents a participant, and each row a feature. White denotes the absence of a feature at AF first mention time, while red denotes the presence of the feature. Renal disease, fluid-electrolyte disorders, and coronary artery disease (CAD) are the most striking features that determine participant cluster assignment. B, Comorbidity burden for renal disease, fluid-electrolyte disorders, and CAD distribution across participant clusters. BMI indicates body mass index; COPD, chronic obstructive pulmonary disease; DM, diabetes; HTN, hypertension; ICD, implantable cardioverter defibrillator; LVSD, left ventricular systolic dysfunction; OSA, obstructive sleep apnea; PAD, peripheral arterial disease; Rheum, rheumatologic diseases; and VHD, valvular heart disease.

Association of Clusters With Clinical Outcomes

Unadjusted survival differed significantly among comorbidity clusters (median survival, 23.2 years [95% CI, 22– >25] for low cluster, 16.7 years [95% CI, 15.1–18.5] for mid cluster, and 9.5 years [95% CI, 8.3–10.9] for high cluster), with a dose-response effect observed between increasing cluster comorbidity burden and worse survival (Figure 3A). A similar relationship was seen with mortality when the 3 key variables determining cluster assignment (renal disease, CAD, fluid-electrolyte disorders) were considered (Figure 3B). Median survival was >25 years when all these variables were absent, 20.2 years (95% CI, 19.1–22.3) with one variable present, and 13.6 years (95% CI, 12.5–15.1) with 2 or 3 variables present. Age at first AF mention was also different between groups (median age at first AF mention was 68.1 years [IQR, 59.2–75.8] for low cluster, 68.3 years [IQR, 59.8–73.3] for mid cluster, and 64.6 years [IQR, 56.3–72.8] for high cluster, P<0.001). When adjusting for age and CHA₂DS₂-VASc score at the time of first AF mention, cluster assignment remained significantly associated with survival (adjusted median survival, 24.7 years [95% CI 22.3– >25] for low cluster, 16.9 years [95% CI, 15.1–18.6] for mid cluster, and 7.8 years [95% CI, 5.14–10.6] for high cluster; Figure 3C). Furthermore, the increased hazards of death with increasing cluster comorbidity burden were also time-dependent, with absolute differences between clusters being more pronounced during the initial follow-up period (Table S3).

Figure 3.

Survival by participant cluster assignment and by burden of renal disease, coronary artery disease, and fluid-electrolyte disorders. A, Kaplan-Meier curve demonstrating observed survival by participant cluster assignment. B, Kaplan-Meier curves demonstrating observed survival by burden of renal disease, coronary artery disease (CAD), and fluid-electrolyte disorders. C, Predicted survival by participant cluster assignment adjusted for age and CHA₂DS₂-VASc score at time of first atrial fibrillation (AF) mention. The solid line represents the point estimate and the shaded area the 95% point-wise CI.

Like survival, there was a significant difference in event-rates for stroke, bleeding, pacemaker placement, implantable cardioverter defibrillator placement, AF ablation, and cardioversion (Figure 4; Figure S5). In general, patients in the high cluster were more likely to have episodes of stroke or bleeding and implantable cardioverter defibrillator implantation, while patients in the low cluster were more likely to have ablations and cardioversions. Patients in the mid cluster were most likely to have pacemaker implantation. Antiarrhythmic drug therapy was comparable across clusters (Figure S5).

Figure 4.

Unadjusted clinical outcomes during follow-up by participant cluster assignment. A, Kaplan-Meier curve demonstrating observed stroke-free survival by participant cluster assignment. B, Kaplan-Meier curves demonstrating observed major bleeding-free survival by participant cluster assignment. C, Observed cumulative cardioversion event rate by participant cluster assignment. D, Observed cumulative atrial fibrillation (AF) ablation event rate by participant cluster assignment. E, Observed cumulative implantable cardioverter defibrillator (ICD) implantation event rate by participant cluster assignment. F, Observed cumulative pacemaker implantation event rate by participant cluster assignment.

Association of Clusters With AF Polygenic Risk, Measured Inflammatory Biomarkers, and Genetically Predicted Cytokine Levels

There was a significant difference among clusters with respect to genetic liability to AF assessed by the normalized AF Polygenic risk score after adjustment for age, sex, and the first 4 PCs of ancestry, with the highest genetic susceptibility to AF seen in the low comorbidity cluster and the lowest genetic susceptibility to AF in the high comorbidity cluster (Figure 5).

Figure 5.

Normalized atrial fibrillation (AF) polygenic risk score by participant cluster adjusted for age, sex, and the first 4 principal components of ancestry. Normalized (Z-score) of AF Polygenic risk score (AF PRS) by participant cluster group adjusted for the first 4 principal components of ancestry, age at first AF mention, and sex. Data are presented as median predicted values after adjustment (dot) and pointwise 95% CI (bars).

We compared clusters with respect to clinically ascertained levels of C-reactive protein, erythrocyte sedimentation rate, and white blood cell count measured within 6 months of AF first mention. There was a significant association between AF cluster and measured inflammatory biomarkers (Table 3): all biomarkers (P<0.001 for each) had the lowest values in the low comorbidity cluster and the highest values in the high comorbidity cluster, with intermediate values in the mid cluster.

Table 3.

Measured Inflammatory Biomarker Levels by Cluster

To assess genetic liability to inflammation, an exploratory PheWAS (Phenome-Wide Association Studies) of genetically predicted circulating cytokines identified 13 markers (IFN-GR1, IL-10RB, IL-17F, IL-17RD, IL-1RL2, IL-1RN, IL-23R, IL-34, IL-37, IL6R, IL-6ST, TNFRSF-11B, TNFRSF-6B) to take forward in this analysis. After accounting for multiple comparison testing, there were no significant differences among clusters with respect to genetically predicted levels for these inflammatory cytokines and cytokine-associated circulating proteins. The normalized genetically predicted cytokine levels overall and by cluster, as well as their nominal P values and FDR, are presented in Supplemental File 5.

Sensitivity Analysis, Association of Different Numbers of Prespecified Subject Clusters With Mortality

We evaluated the behavior of predictor assignment by subject cluster and association with mortality if the prespecified subject cluster number provided to the coclustering algorithm was 2 or 4 instead of 3. Like in the main analysis, in both cases, clusters were distinguished by increasing burden of comorbidities, but again, renal disease, fluid-electrolyte disorders, and CAD had the most pronounced differences (Figures S6 and S7). Also, as observed in the main analysis, cluster assignment was associated with survival, with a dose-response effect observed between increasing cluster comorbidity burden and worse survival (Figure S8).

Discussion

We conducted an analysis of 23 271 subjects with AF from a large tertiary medical center that identified 3 phenotypic disease clusters associated with mortality, stroke, and bleeding. Cluster membership was largely driven by comorbid diseases, in particular the presence of renal disease, CAD, and fluid-electrolyte disturbances. Greater comorbidity burden across groups was associated with higher levels of clinically measured inflammatory biomarkers and inversely associated with a polygenic predisposition for AF. There was no detectable difference among groups with respect to genetic predisposition to inflammation as ascertained by genetically predicted cytokine levels.

Many of our findings are in agreement with prior studies reporting data-driven subgroups of AF, and thereby expand and strengthen the existing evidence base.^39,40 For example, 2 earlier studies (n=9749 and 2458 subjects, respectively), both using observational cohorts and hierarchical clustering, identified 3 to 4 clusters, including clusters comprising subjects with young age or low comorbidity burden and a cluster with CAD as the driving feature. Additionally, both studies reported associations between cluster assignment and major adverse cardiovascular events. Like these studies, we observed a similar number of clusters, with CAD and other comorbidities (renal disease and fluid-electrolyte disturbances) as key drivers of cluster assignment and an association of cluster assignment with mortality, stroke, bleeding, cardioversion, ablation, and cardiac implantable electronic device implant. The present study adds to the evidence base of data-driven analyses of AF phenotypic architecture and supports the importance of comorbidities as determinants of interindividual variability in AF outcomes.

Our work also extends previous studies in important ways. One notable distinction is that the current report also provides new insights into potential mechanisms underlying AF subgroups, including clinically acquired and genetically determined inflammation as well as genetic susceptibility to AF itself. We observed higher levels of clinically measured inflammatory biomarkers in clusters with more comorbidities. However, there were no differences with respect to genetically predicted levels of inflammatory cytokines. This suggests that observed epidemiological associations between inflammation and AF are driven by acquired comorbidities and environmental factors rather than a genetic predisposition to increased inflammation. These observations could explain, at least in part, the lack of universal efficacy of anti-inflammatory agents as AF therapies,^45–48 and further support the importance of cardiovascular risk factor optimization in the management of AF. We also observed the greatest polygenic risk for AF in the group with the lowest comorbidity burden. Clinically, this cluster has many similarities with a group of patients previously referred to as having lone AF.

A second distinction of our study with prior ones is the size of our study population, which comprised over 23 000 subjects from a large medical center with dense phenotype data and more than 5 years of follow-up after AF first mention in the electronic health record. This population is more than twice the size of earlier reports and was sufficiently large to enable more detailed analyses of AF phenotypic architecture, including evaluation of a broader range of comorbidities, inflammatory markers, and polygenic risk. For example, we identified renal disease as an important determinant of cluster assignment, a finding that has not been previously reported. This observation warrants replication in other studies but suggests a previously underappreciated contribution of renal disease to AF phenotypic heterogeneity, especially in clinical populations.

Collectively, in the context of prior literature, our findings suggest that observed phenotypic variability in AF, in part, is distributed along a gradient with differing contributions of genetic risk and acquired comorbidity. This framework for understanding AF heterogeneity, incorporating both environmental and genetic factors, aligns with clinical experience and has been previously articulated in a theoretical context.⁴⁹ However, to our knowledge, our analysis is the first to empirically demonstrate the complementary roles of comorbidity and genetic liability as major drivers of phenotypic variability using hypothesis-free methods.

Limitations

Even though we took steps to improve the quality and completeness of electronic health record-derived data, including restricting the analysis to subjects receiving longitudinal care in our health system, misclassification and under-ascertainment of clinical phenotypes and outcomes, notably mortality, which may be under-estimated in the electronic health record, are possible due to misdiagnosis or underdiagnosis. We included patients with cardiac implantable electronic devices before AF diagnosis, which could lead to earlier detection of AF in the earliest disease stage. However, because cardiac implantable electronic devices were more frequently present before AF mention in the High group, one would expect this would have biased our results towards the null. Additionally, measured biomarkers were not prospectively collected, likely obtained during the time of acute events, and therefore are susceptible to bias. Only a portion of patients included had genetic data available, which limited the power of the genetically predicted cytokine levels analysis. It is also possible that genetically predicted cytokine levels did not capture sufficient genetic variability of cytokine levels; we attempted to minimize this issue, however, by only using genetically predicted cytokine levels validated by phenome-wide scanning in the analyses.

Conclusions

Patient subgroups identified by unsupervised clustering were associated with differing risk of mortality, stroke, bleeding, and implantable cardioverter defibrillator implant. These groups were distinguished by comorbidity burden and polygenic liability to AF, which were inversely correlated. Clinical inflammation, as reflected by higher measured biomarker levels, was associated with increasing comorbidity burden, but there were no differences in genetically predicted levels of inflammatory biomarkers, consistent with the idea that inflammatory burden is driven by acquired comorbidities. Additionally, our study empirically demonstrates the complementary roles of comorbidity and genetic liability as major drivers of AF phenotypic variability using hypothesis-free methods.

Article Information

Acknowledgments

Clinical data were obtained using Vanderbilt’s Synthetic Derivative. The Synthetic Derivative resource is supported by the CTSA award no. UL1TR000445 from the National Center for Advancing Translational Sciences. Genomic data were obtained using Vanderbilt’s de-identified biobank BioVU. Vanderbilt University Medical Center’s BioVU projects are supported by numerous sources: institutional funding, private agencies, and federal grants. These include the National Institutes of Health (NIH) funded Shared Instrumentation Grant S10OD017985, S10RR025141, and S10OD025092; CTSA grants UL1TR002243, UL1TR000445, and UL1RR024975. The contents of this publication are solely the responsibility of the authors and do not necessarily represent official views of the National Center for Advancing Translational Sciences or the National Institutes of Health.

Sources of Funding

This project was supported by Vanderbilt’s participation in the American Heart Association’s Atrial Fibrillation Strategically Focused Research Network (18SFRN34110369 and 18SFRN34230089). This project also received support from the National Institutes of Health (NIH) grants R01 HL149826 (Dr Roden), R01 HL142856 (Dr Mosley). Dr Davogustto receives funding from the American Heart Association (24CDA1271003).

Disclosures

None.

Supplemental Material

Materials and Methods

Tables S1–S4

Figures S1–S9

Supplemental Files 1 to 5