Unsupervised Cluster Analysis of Patients With Aortic Stenosis Reveals Distinct Population With Different Phenotypes and Outcomes

Introduction

The prevalence of aortic stenosis (AS) continues to increase.¹ Severe AS is fatal without aortic valve replacement (AVR) and the evolution of surgical AVR and transcatheter AVR has significantly improved the prognosis.^2,3 With the increased burden of comorbidities⁴ and expanded treatment options,⁵ phenotypes and outcomes of patients with AS may be more heterogeneous than expected.

Current guidelines categorize AS into 4 stages according to its severity based on echocardiographic aortic valve assessments and symptoms.⁶ However, the clinical course of AS is variable with complex interactions between the patients’ characteristics, associated cardiac and noncardiac diseases, factors critical when deciding the timing and types of intervention.^6,7 Moreover, there is growing evidence that extravalvular cardiac damages, such as myocardial fibrosis, can be substantially diverse among AS and associated with adverse outcomes.⁸ These imply that the current valve-oriented classification may not capture the heterogeneity within AS, and sophisticated phenotyping with multiple factors may have additive value.

Recently, machine learning has been adopted in cardiovascular research.⁹ Unsupervised cluster analysis categorizes the complex entities without investigators’ supervision by segregating samples into homogenous groups based on each cluster’s dissimilarities.⁹ This helps to unveil meaningful phenotypes within a disease that has been previously considered homogenous. Cluster analysis may be invaluable in phenotyping several cardiovascular diseases^10–12 but has never been adopted to AS.

We hypothesized that there might be clinically distinct AS clusters. We aimed to explore whether unsupervised cluster analysis can identify clinically relevant groups among AS with different outcomes and causes of death.

Methods

The study materials are available from the corresponding authors on reasonable request. The overall scheme of the study is depicted in Figure 1 and more detailed methods are available in Methods in the Data Supplement.

• Download figure
• Download PowerPoint

Study Design and Cohort

A prospective cohort of newly diagnosed patients with moderate or severe AS at a tertiary university hospital (Seoul National University Hospital, Korea) between 2013 and 2016 was used for derivation of the unsupervised clustering (n=398). We used another separate dataset of patients with AS with a different enrollment period, from 2010 to 2012 at the same institution (n=262), for validation of the clusters. Details on the cohort characteristics, inclusion/exclusion criteria, and definition of comorbidities are in Method in the Data Supplement.

The study protocol was approved by the institutional review board, and all participants in the derivation and validation cohort provided informed consent.

Outcome Assessment

All-cause mortality was the primary outcome. Secondary outcomes include cardiac mortality, noncardiac mortality, and death after AVR. Cardiac mortality was defined as either sudden cardiac arrest, death from heart failure or myocardial infarction, or death related to AVR. Noncardiac mortality was defined as mortality other than the cardiac causes.

Mortality with no identifiable cause of death from the death certificate was classified as indeterminate. Follow-up data were available in 94% of the derivation, and 100% of the validation cohort.

Variable Preparation for Cluster Analysis

Variables used for clustering were recruited from clinical or echocardiographic domains that are either routinely obtained in the assessment of AS, used for risk stratification, or have prognostic value.^7,13–15Table 1 summarizes phenotypic domains and variables used for the analysis.

The missing values were imputed with the missForest algorithm,¹⁶ with appropriate imputation error (Figure I in the Data Supplement).^16,17 Then, we selected pivotal variables for clustering through the dimension reduction using Pearson coefficient and Bayesian information criterion (BIC), which penalizes model complexity. A larger BIC indicates a stronger likelihood of the corresponding model. Details of variable preparation are described in Method in the Data Supplement.

Unsupervised Clustering

For the primary cluster analysis, we applied model-based clustering,^18,19 which has been broadly adopted in previous studies.^10–12 The number of clusters and specific geometric model were chosen based on BIC,^{10–12,18,19} as well as the integrated complete-data likelihood.^18,19 Clustering was performed independently from the outcome data. After allocating individuals to each cluster, we compared phenotypes and outcomes between clusters, followed by interpretation of its clinical relevance.

In addition, we used another popular clustering algorithm, agglomerative hierarchical clustering (Ward's method), to investigate whether it produces clusters of similar meaning. The number of clusters was determined based on 30 indices for hierarchical clustering. We also compared phenotypes and outcomes within these clusters. Details on model-based and hierarchical clustering are in Method in the Data Supplement.

Validation of the Findings From the Derivation Cohort

To validate the generalizability of the clustering, we used data from a separate validation cohort (n=262). The data was imputed and normalized using the same method as the derivation cohort (Figure II in the Data Supplement). Cluster prediction was determined by multivariate observations (the variables used for clustering) based on Gaussian finite mixture models derived from the model-based clustering.^10,19

Statistical Analysis

Continuous variables are presented as mean±SD and categorical variables as numbers (percentages). The difference between continuous variables was analyzed using the ANOVA or Kruskal-Wallis test, and for categorical variables, either χ² or Fisher's exact test. Kaplan-Meier curves were plotted with the duration from the enrollment to the last follow-up or death and compared with the log-rank test.

Cox proportional hazard analyses were performed to evaluate the association between the outcomes and clusters. In the derivation cohort, multivariate Cox models were adjusted only for basic characteristics (age, sex, and body mass index) to avoid possible overfitting with the small number of events. Patients with missing outcome data were excluded from the survival and Cox analyses.

To further evaluate the prognostic and discriminative utility of the clusters, we compared the predictability for 3-year outcomes between the model with and without the cluster variable.^20,21 The base model was built with the variables significant in univariate Cox analysis for all-cause mortality, with missing values <10%. We calculated C statistics, net reclassification improvement and integrated discrimination improvement for comparing prediction accuracy of the 2 models. The statistical inference for comparing 2 models was conducted by the perturbation-resampling method by Uno,^20,21 which uses an inverse probability of censoring weights. The truncation time was set at 3-year and the P-value was obtained with 1000 perturbation samples.²¹ This analysis was conducted with the survIDINRI package of R software. Additionally, we performed the same analysis using the CURRENT-AS risk score²² as a reference standard.

All analyses were done with R (Vienna, Austria) and its packages (Table I in the Data Supplement). A Pvalue <0.05 was considered statistically significant.

Results

Study Population

Between 2013 and 2016, 441 patients with AS were recruited. Among these, 43 patients were excluded because of a history of myocardial infarction or coronary artery bypass graft surgery (n=21), a history of valve surgery (n=12), or other valvular diseases greater than or equal to moderate degree (n=10). The remaining 398 patients constituted the final derivation cohort.

Variable Selection and Optimal Number of Clusters

From the initial 32 variables, 11 variables (6 clinical, 5 imaging parameters) were selected through the dimension reduction step (Table 1). A heatmap with the selected variables is shown (Figure IIIA in the Data Supplement). In model-based clustering, the VVE model with 3 clusters had the maximum BIC and integrated complete-data likelihood values (Figure IIIB and IIIC in the Data Supplement), which we concluded as the most optimal model and the number of clusters. The BIC of the final model using the 11 selected variables was substantially improved compared with the model with all variables (n=32), confirming the validity of the dimension reduction step (Table II in the Data Supplement). The relative importance of the 11 variables for cluster assignment was assessed by the McFadden pseudo-R² (Method in the Data Supplement), the rank of which was as follows: hemoglobin, tricuspid regurgitant jet velocity, creatinine, left atrial volume, E-wave velocity, left ventricular ejection fraction, body mass index, heart rate, A-wave velocity, platelet, and white blood cell count (Figure IV in the Data Supplement).

Comparison of Clinical and Echocardiography Parameters Between Clusters

The baseline characteristics of the clusters were compared (Table 2). Cluster 2 consisted predominantly of lean, elderly patients with more prevalent comorbidities, particularly end-stage renal disease, with the lowest glomerular filtration rate and hemoglobin level. Patients in cluster 3 were the youngest, least symptomatic, and had the least comorbidities among the 3 groups. In cluster 1, the most notable finding was the highest prevalence of atrial fibrillation.

Regarding echocardiographic evaluation, cluster 1 had significantly depressed left ventricular ejection fraction, more left ventricular hypertrophy, and more severe diastolic dysfunction (Table 3). In contrast, cluster 3 had the most preserved cardiac function and structure. Cluster 2 had mid-range values (ie, left ventricular mass index, left atrial volume) between cluster 1 and 3. Notably, there was no difference in AS severity between the 3 clusters. Overall, the 3 clusters could be characterized as follows: cardiac dysfunction for cluster 1, comorbidities for cluster 2, and healthy AS for cluster 3.

Clinical Outcomes of Each Cluster in the Derivation Cohort

Clinical outcomes were markedly different per cluster. During a median 2.4 years (interquartile range, 1.3–3.4 years) follow-up, there were 40 mortality cases (14 cardiac mortality, 20 noncardiac mortality, and 6 indeterminate; Table III in the Data Supplement). Cluster 2 had the highest all-cause mortality, followed by cluster 1 (P<0.001; Figure 2A). The cumulative incidence of cardiac mortality was the highest in cluster 1 and also significantly increased in cluster 2 compared with cluster 3 (P=0.005; Figure 2B). However, noncardiac mortality occurred predominantly only in cluster 2 (P=0.001; Figure 2C).

• Download figure
• Download PowerPoint

In unadjusted Cox analysis, when compared with cluster 3, all-cause mortality risk was higher in cluster 2, as well as in cluster 1 with marginal significance (Table 4). Cluster 1 had the strongest risk of cardiac mortality (HR, 6.44 [95% CI, 1.82–22.83]; P=0.004) but no association with noncardiac mortality, whereas the risk of noncardiac mortality was significantly elevated in cluster 2 (HR, 4.51 [95% CI, 1.78–11.45]; P=0.002). Cluster 2 also had a higher cardiac mortality risk (HR, 3.67 [95% CI, 0.92–14.72]; P=0.066). The association of cluster 1 with cardiac mortality and cluster 2 with noncardiac mortality was consistent in the adjusted Cox analysis (Table 4).

During follow-up, 196 (49%) received AVR, of which 32 were transcatheter AVR and 164 surgical AVR. There was no notable difference in the proportion and types of AVR between the clusters (Figure V in the Data Supplement). Cluster 2 had the worst survival after AVR, although not statistically different (Figure 2D).

Incremental Predictive Value of the Clusters for the Prediction of Outcomes

The baseline prediction models for 3-year outcomes were constructed using the risk factors identified from the univariate Cox analysis. Univariate and multivariate Cox analyses for these variables are shown in Table IV in the Data Supplement. The addition of the cluster variable to the base model showed a significant integrated discrimination improvement and net reclassification improvement for 3-year all-cause mortality (C statistics 0.762 versus 0.788; integrated discrimination improvement 0.029, P=0.020; net reclassification improvement 0.294, P=0.032), as well as for noncardiac mortality (Table 5). For cardiac mortality, the predictability was improved based on integrated discrimination improvement (Table 5). The result was consistent when the CURRENT-AS risk score²² was used as a reference standard (Table V in the Data Supplement).

Clinical Outcomes of the Clusters in the Independent Validation Cohort

A separate data of 262 patients with AS was used for validation of the findings from the derivation cohort. The difference between the 2 cohorts are described (Method and Table VI in the Data Supplement). The cluster prediction for the validation cohort was performed using the same 11 variables identified from the derivation cohort. As in the derivation cohort, cluster 2 in the validation cohort (V-cluster 2) included elderly patients with end-stage renal disease, and cluster 1 in the validation cohort (V-cluster 1) had reduced left ventricular ejection fraction and more frequent atrial fibrillation (Table VII in the Data Supplement).

The pattern of outcomes was reproduced in the validation cohort (Figure 3). During a median 4.3 years (interquartile range, 0.8–6.5 years) follow-up, 113 mortalities occurred (41 cardiac cause, 59 noncardiac cause, and 13 indeterminate). The all-cause mortality rate was the highest in V-cluster 2, followed by V-cluster 1 (Figure 3A). Both V-cluster 1 and 2 had increased cardiac mortality compared with V-cluster 3 (Figure 3B). However, the majority of noncardiac deaths occurred in V-cluster 2 (Figure 3C), as well as the post-AVR deaths (Figure 3D).

• Download figure
• Download PowerPoint

In Cox analysis of the validation cohort, V-cluster 1 was associated with an increased risk of cardiac mortality, whereas V-cluster 2 had increased risk of both cardiac and noncardiac mortality (Table VIII in the Data Supplement), findings consistent with that from the derivation cohort.

Alternative Cluster Analysis Using Hierarchical Clustering

We further investigated whether another alternative cluster algorithm produces similar results. The derivation cohort data were reanalyzed with agglomerative hierarchical clustering. Three clusters were identified in the hierarchical clustering by the majority rule (Figure VI in the Data Supplement). The characteristics of each cluster generally corresponded with the original clusters from model-based clustering, that cluster 2 in hierarchical clustering (H-cluster 2) had more frequent comorbidities, and H-cluster 1 was characterized by cardiac dysfunctions, while H-cluster 3 had neither (Table IX in the Data Supplement).

The overall trend of outcomes was similar to that of the model-based clustering. H-cluster 2 had the highest all-cause, noncardiac, and post-AVR mortality rate (Figure VII in the Data Supplement). Both H-cluster 1 and 2 had more cardiac mortality compared with H-cluster 3, although the higher rate of cardiac mortality in H-cluster 1 was less prominent (P=0.078).

Discussion

There are 3 main findings in the current study. First, unsupervised cluster analysis successfully demonstrated 3 groups of patients with moderate or severe AS with distinct phenotypes. Second, each cluster had markedly different patterns of outcomes and causes of death. Third, the result was replicated in the validation cohort. This study provides a guide on how patients with AS can be phenotypically categorized and what kinds of outcomes would be expected within these specific groups.

Cardiologists are apt to focus on echocardiography parameters of the valve in patients with AS. However, a variety of prognostic features has been identified beyond the valve, which is not included in the current staging.^7,8,14,15,23 A recent study reported that the AS classification based on extravalvular cardiac damages (ie, left ventricular hypertrophy, tricuspid regurgitation) has a significant prognostic value in predicting outcomes after AVR, while AS severity measurements (ie, aortic valve area) was not associated with adverse events.²³ While these diverse factors have been investigated individually through the conventional hypothesis-driven approach, a more integrative, data-driven cluster analysis can be powerful to explore heterogeneity.^9–12

In our cluster analysis, each of the 3 clusters demonstrated distinct phenotypes and can be interpreted as follows: cardiac dysfunction for cluster 1, comorbidities for cluster 2, and healthy AS for cluster 3. Notably, echocardiographic variables that were pivotal for clustering were tricuspid regurgitant jet velocity, left atrial volume, E-, and A-wave velocity, and left ventricular ejection fraction (Figure IV in the Data Supplement), whereas none of the AS severity indices were critical for clustering. Despite the similar degree of AS severity throughout the groups, cluster 1 was characterized by various structural and functional cardiac dysfunctions, which led to the worst cardiac consequences. The result implies the substantial prognostic value within these cardiac imaging markers in AS outside the diseased aortic valve.^14,23

Importantly, cluster 2 presented the worst prognosis in all-cause mortality, a difference mainly seen in noncardiac causes. Although noncardiac comorbidities, such as malignancy or infection, accounts for almost half of the actual deaths in AS,²⁴ this has not been paid enough attention. Factors associated with noncardiac death are age, low body mass index, anemia, and dialysis,²⁴ which were predominant features of cluster 2. These conditions, particularly end-stage renal disease, are also poor prognosticators even after AVR,²⁵ further supporting the lowest post-AVR survival of cluster 2. Collectively, our results highlight that there are specific types of death more related to specific groups of patients with AS, and noncardiac death should not be neglected, particularly in those with significant comorbidities.

With the similar phenotypes for the corresponding clusters, the pattern of outcome and the leading cause of death were reproduced in an independent cohort (n=262), providing the external clinical validity and generalizability.¹⁰ Notably, the marginally increased cardiac mortality risk of cluster 2 in the derivation cohort became more evident in the validation cohort, as well as the worst post-AVR prognosis (Figure 3B and 3D). This further supported the characteristic outcomes of the clusters (cluster 1—cardiac mortality, cluster 2—noncardiac and cardiac mortality, and death after AVR).

Our study suggests that a different therapeutic and surveillance strategy may be needed for each cluster, as a step toward precision-medicine in AS. In particular, cluster analysis with deeply phenotyped data could be utilized for future trials of AS. With rapid advances in AVR techniques, recent studies have tried to address the optimal timing of intervention, specifically for patients with moderate or asymptomatic severe AS.^26,27 Along with this clinical demand, there are ongoing efforts for a more sophisticated risk stratification using clinical factors,²² diverse imaging parameters,²³ and biomarkers,²⁷ to identify individuals for whom the benefits of AVR outweigh the risks.²⁷ The cluster-based analysis, which segregates patients into distinct phenotypes and relevant adverse outcomes, may offer a novel target for AVR among the heterogeneous AS patients for future trials. For instance, given a similar favorable post-AVR prognosis to the healthy AS group, cluster 1 (cardiac dysfunction group) might be a candidate for a more active AVR, while it may not completely resolve the dismal outcome in cluster 2 (comorbidities group; Figure 3D). The clinical implications of cluster-based approaches may also be enhanced by incorporating a broader range of data, including proteomics or cardiac magnetic resonance. These hypotheses need to be tested in future work.

Limitations

First, choosing the optimal clustering method can be difficult to determine. Although we found 3 distinct clusters using model-based clustering, which generally corresponds to the hierarchical clustering, other algorithms may yield different results. We adopted the model-based clustering as it is based on the likelihood and statistically more flexible. For instance, the number of clusters can be chosen by BIC or likelihood ratio test.^9–12 Second, patients were enrolled from a single institution, and the validation is warranted in datasets from other institutions. Additionally, there were several differences between the derivation and validation cohort. Nevertheless, the phenotypes and outcomes were reproduced in the validation cohort, verifying the robustness of our results. Third, the multivariate Cox analysis may be limited due to the small number of events in the derivation cohort. Last, some variables had missing values. However, they were imputed with an adequate statistical method,¹⁶ the appropriateness verified in previous studies.¹⁷

Conclusions

Unsupervised cluster analysis of patients with AS demonstrated 3 distinct phenotypes with significantly different outcomes. The result provides new insight into a novel grouping of patients with AS, emphasizing the role of comorbidities and extravalvular cardiac dysfunctions in association with different causes of death.

Footnotes