Predictors of post-operative cardiovascular events, focused on atrial fibrillation, after valve surgery for primary mitral regurgitation

Abstract

Aims

Primary mitral regurgitation (PMR) can be considered as a heterogeneous clinical disease. The optimal timing of valve surgery for severe PMR remains unknown. To determine whether unbiased clustering analysis using dense phenotypic data (phenomapping) could identify phenotypically distinct PMR categories of patients.

Methods and results

One hundred and twenty-two patients who underwent surgery were analysed, excluding patients with pre-operative permanent atrial fibrillation (AF), were prospectively included before surgery. They were given an extensive echocardiographic evaluation before surgery, and clinical data were collected. These phenotypic variables were grouped in clusters using hierarchical clustering analysis. Then, different groups were created using a dedicated phenomapping algorithm. Post-operative outcomes were compared between the groups. The primary endpoint was post-operative cardiovascular events (PCE), defined as a composite of: deaths, AF, stroke, and rehospitalization. The secondary endpoint was post-operative AF. Data from three phenogroups with different characteristics and prognoses were identified. Phenogroup-1 (67 patients) was the reference group. Phenogroup-2 (33 patients) included intermediate-risk male and smoker patients with heart remodelling. Phenogroup-3 (22 patients) included older female patients with comorbidities (chronic renal failure, paroxysmal AF) and diastolic dysfunction. They had a higher risk of developing both PCE [(hazard ratio) HR = 3.57(1.72–7.44), P < 0.001] and post-operative AF [HR = 4.75(2.03–11.10), P < 0.001]. Pre-operative paroxysmal AF was identified as an independent risk factor for PCE.

Conclusion

Classification of PMR can be improved using statistical learning algorithms to define therapeutically homogeneous patient subclasses. High-risk patients can be identified, and these patients should be carefully monitored and may even be treated earlier.

Introduction

The optimal timing of mitral valve surgery for severe primary mitral regurgitation (PMR) is still being debated even with the most recent guidelines on the management of valvular heart disease.^1,2

Although some risk factors of developing events, including post-operative atrial fibrillation (AF), have been identified, it remains difficult to predict which patients will develop post-operative AF and events before sending the patient to surgery.

Therefore, we hypothesized that PMR is a heterogeneous clinical disease that affects variable patient populations and, thus, different phenotypes of patients; available predictive parameters have low positive predictive values.^3,4 Machine learning, which can be defined as a process of using data to learn relationships between objects,⁵ was used in this study to determine the different phenotypes of severe PMR and to assign patients into these groups. Machine learning uses statistical learning algorithms to realize an unbiased hierarchical clustering analysis of phenotypic data. This statistical method is usually performed to analyse genetic data, but the approach was recently used to improve characterization of another heterogeneous cardiovascular syndrome.⁶ In this study, unbiased clustering analysis using dense phenotypic data, called ‘phenomapping’, resulted in the characterization of three groups of patients with different characteristics, defining a novel HFpEF-phenogroup classification.

The aims of this study were (i) to identify different phenogroups of patients with severe PMR, analysing clinical, demographic, electrocardiographic, and echocardiographic data, (ii) to search predictors of post-operative cardiovascular events (PCE), particularly AF, and (iii) to analyse outcomes in each phenogroup to improve management of this heterogeneous population.

Methods

Data collection and patients

From June 2007 to December 2015, data from 201 patients suffering from PMR due to leaflet prolapse, who underwent a pre-operative check-up in our institution, were prospectively collected. Demographic data, patients’ medical backgrounds, haemodynamic status, treatments, symptoms, operative data, and in-hospital outcomes were collected from patients’ medical records. All patients underwent pre-operative transthoracic 2D echocardiography (Vivid 7 or 9, GE Healthcare, Horten, Norway) with Doppler and tissue Doppler imaging. Echo recordings were retrospectively reanalysed, and measurements were performed by experienced blinded cardiologists (EchoPAC BT12, General Electric, Horten, Norway). Post-operative events were collected in hospitalization reports and by phone-calls with referent physicians. Among these 201 patients, 151 underwent surgery (mitral valve repair or replacement) at the end of the follow-up (December 2015). Patients with permanent pre-operative AF were excluded to account for confounders, as were patients presenting a non-severe PMR, a missing follow-up or extreme echocardiographic values (i.e. values above or below three deviations around the mean) (see Figure 1 and Supplementary Data online). Finally, 122 patients who underwent mitral valve surgery for severe PMR were included in this study. This study was conducted in accordance with institutional policies, national legal requirements, and the revised Declaration of Helsinki. The study was approved by an ethics committee (Person Protection Committee West V) (PIME 08/16-675).

Definitions

In this study, the PMR aetiology was mitral valve prolapse (Barlow disease and degenerative mitral regurgitation), Type 2 of Carpentier’s classification, defined according to the recommended criteria.⁷ Rheumatic heart diseases, restrictive lesions (due to drugs or radiation therapy), and infective endocarditis were not included.

Quantification of mitral regurgitation was assessed according to recommendations.⁸ Ventricular functions and volume measurements were also based on the recommendations.⁹ Global longitudinal strain (GLS) was measured from the three apical views using EchoPAC.

Paroxysmal AF was defined according to European Society of Cardiology (ESC) guidelines.¹⁰

Renal failure was defined as an estimated glomerular filtration rate <60 mL/min. The standard logistic European System for Cardiac Operative Risk Evaluation (EuroSCOREs 1 and 2) were calculated (www.euroscore.org).

Endpoints

After identification of three phenogroups (details below), post-operative outcomes were compared between groups: post-operative immediate AF (occurrence before or at 30 days), post-operative long-term AF (after 30 days), all-cause mortality, cardiovascular mortality, stroke, and cardiovascular cause of hospitalization. The primary endpoint was survival free of cardiovascular events. The secondary endpoint was survival free of post-operative long-term AF.

Statistical analysis

Phenotypic domains and clustering of variables

The phenotypic domains consisted of 64 variables, including clinical and demographic data, patients’ medical background (AF history, renal failure, and cardiovascular risk factors), EuroSCORE, treatments, symptoms, and echocardiographic parameters (indexed variables) (Supplementary Data online).

As described above, we excluded patients presenting extreme values, defined as values below or above three standard deviations around the mean. In case of asymmetric distribution, we performed either log, square-root, inverse, or power transformation to improve normality of variables, which was graphically checked.

We used the missForest algorithm¹¹ to impute missing data before the clustering analysis. To remove redundancies, we performed ascending hierarchical variable clustering using the ClustVarLV R package.¹² This algorithm is based on the aggregation of variables around latent components, with the capability to take into account the direction of correlations (i.e. position or negative association) between variables of mixed types.

Clustering of patients

The next step consisted of identifying clusters of similar patients from the previous latent components, which provided summarized information of the original variable. We used hierarchical clustering with the dissimilarity matrix given by Euclidean distance and the Ward’s minimum variance method for aggregation. The final clusters were determined by consensus across a set of criteria used to select the optimal number of clusters.¹³

Comparison of clinical characteristics and survival among phenogroups

We compared the clinical, demographic, electrocardiographic, and echocardiographic characteristics between clusters using the Kruskal–Wallis test for continuous variables and the χ² test (or Fisher’s exact test when appropriate) for categorical variables. Statistical significance was considered as a two-sided P-value <0.05. In this case, we also computed pairwise comparisons using the Conover test for continuous variables and the Fisher’s exact test for qualitative variables. P-values were adjusted with the Bonferroni correction.

We used Cox regression models to calculate between-group differences in PCEs and post-operative AF. Group 1 was considered as the reference group for survival analysis. We analysed Schoenfeld residuals to test the assumption of proportional hazards and used the Kaplan–Meier method to calculate survival curves.

We used R statistical software, version 3.3.3, for all analyses.¹⁴

Results

Patients and description of phenogroups

Performing ascending hierarchical variable clustering, the 64 phenotypic variables were grouped in six clusters of variables (Supplementary Data online). Then, we identified three groups of patients with the biclustering procedure.

The baseline characteristics of patients are depicted in Table 1. Data from 122 patients (among 151 operated, Figure 1) were analysed (median age 63 years old; 68% male). Mitral valve repair was performed in 105 (86%) patients, whereas 17 patients received mitral valve replacements (14%).

Phenogroup-1 included 67 (median age 61 years old; 70% male) patients. Phenogroup-2 included 33 (median age 61 years old; 94% male) patients, whereas phenogroup-3 included 22 (median age 73 years old; 23% male) patients.

The baseline characteristics were imbalanced between groups. Phenogroups-1 and 2 were predominantly male (70% and 94% of patients, respectively), whereas phenogroup-3 was mostly composed of women (77% of patients, P < 0.001) (Table 1 and Supplementary data online, Table S1). Patients in Group 3 were older than in the others (73 ± 5 years old vs. 61 ± 10 years old in Group 2 and 61 ± 11 in Group 1; P < 0.0001 after pairwise comparison).

Regarding cardiovascular risk factors (besides age and overweight, mentioned above), patients in phenogroup-3 were more likely to have high blood pressure than those of the first phenogroup: 64% of patients in Group 3 had hypertension vs. 29% in Group 1 (P = 0.015) (Supplementary data online, Table S1). Patients in Group 3 were also more likely to have diabetes mellitus (Table 1).

Regarding comorbidities, chronic obstructive pulmonary disease (COPD) was more frequent in Group 2 (see Supplementary data online, Table S1). Before surgery, the predicted operative mortality (estimated using EuroSCORE) in phenogroup-3 was higher than in other groups: the EuroSCORE 1 was 8.8% in Group 3 vs. 4.2% in Group 2 (P < 0.0001) and 3.3% in the Group 1 (P < 0.0001). The EuroSCORE 2 was 2.3% in Group 3 vs. 1.1% in Group 2 (P < 0.0001) and 1.0% in the Group 1 (P < 0.0001).

Patients in Group 3 were more likely to have pre-operative paroxysmal AF: 45% of patients in Group 3 had a history of AF vs. 18% in Group 2 and only 13% in the Group 1 (P = 0.009) (Table 1).

Pre-operative echocardiographic characteristics

Regarding left ventricular (LV) systolic function: LV ejection fraction (LVEF) was comparable between groups, with a mean of 67 ± 7% (P = 0.127) (Table 2). There was a trend towards a lower GLS in phenogroup-2, but this difference was weakly significant: GLS was −19 ± 3.1% in Group 2 vs. −20 ± 3.0% in Group 3 and −21 ± 3.1% in the Group 1, P = 0.049 (see Supplementary data online, Table S2). In phenogroup-2, the left ventricle was more dilated than in the Group 1.

The left atrium (LA) was larger in Group 2: the indexed LA volume was 66 ± 25 mL/m² (which represent a severe LA dilatation) vs. 58 ± 19 mL/m² in Group 3 and 44 ± 19 mL/m² in the Group 1 (P < 0.001). The pairwise comparison highlighted LA enlargements in both Groups 2 and 3 compared with the Group 1, but there was no longer any statistically significant difference between Groups 2 and 3 (see Supplementary data online, Table S2). The right atrium was also enlarged in Group 2 (Table 2).

Differences between groups regarding quantitative evaluation of mitral regurgitation were observed. Indeed, a larger effective regurgitant orifice area (EROA) and a larger regurgitation volume were assessed in Group 2: the average EROA was 69 mm² vs. 55 mm² in Group 3 and 56 mm² in Group 1 (P = 0.010) (Table 2).

Regarding the right heart, pulmonary artery systolic pressure was estimated at ∼48 ± 13 mmHg in Group 3 vs. ∼42 ± 17 mmHg in Group 2 and ∼35 ± 9.7 mmHg in the Group 1 (P < 0.001) (Table 2). The right ventricle (RV) was more dilated in Group 2 (RV end-diastolic area = 22 ± 3.6 cm² vs. 17 ± 3.1 cm² in Group 3 and 15 ± 2.7 cm² in the Group 1; P < 0.001).

Post-operative outcomes and endpoints

Regarding the primary endpoint (Table 3 and Figure 2) compared with patients in the Group 1, patients in Group 3 were more likely to develop PCE.

Post-operative outcomes are detailed in Table 4. Regarding the secondary endpoint, 29 (24%) patients have developed post-operative long-term AF in the overall population. Post-operative long-term AF occurred significantly more often in phenogroup-3 than in phenogroup-1. There was no statistical difference between the Groups 1 and 2 or between the Groups 2 and 3 (Table 4 and Supplementary data online, Table S3). The risk of developing post-operative long-term AF was significantly increased in Group 3 (Table 3 and Figure 3).

Regarding mortality, only four patients died during the follow-up, including two patients in Groups 1 and 2 patients in Group 3. There was no difference between groups (P = 0.153).

The incidence of cardiovascular causes of hospitalization was significantly higher in phenogroup-2 compared with the first group. Other between-groups differences were not statistically significant.

Paroxysmal pre-operative AF was confirmed as a risk factor of developing both PCE and post-operative AF (Supplementary data online, Figures S1 and S2). The risk of developing post-operative AF (including both immediate and long-term AF) was significantly increased in patients with pre-operative AF: hazard ratio = 3.19(1.49–6.82) (P = 0.00284).

Discussion

In this study, using machine learning and hierarchical clustering analysis applied on dense phenotypic data, the main findings were as follows (i) PMR is a heterogeneous disease, (ii) using a phenomapping dedicated algorithm, three phenogroups of patients were identified, despite PMR heterogeneity, (iii) these phenogroups had different prognoses, suggesting different risk profiles, and (iv) pre-operative paroxysmal AF is an important risk factor of PCE.

A previous study, ‘Phenomapping for Novel Classification of Heart Failure With Preserved Ejection Fraction’,⁶ provided a novel classification of a cardiovascular disorder using phenomapping and was ‘the first study that applies machine learning techniques to resolve heterogeneity in a cardiovascular syndrome using dense phenotypic data’. Using a similar algorithm, our study demonstrates that it is also possible to identify different phenogroups in PMR disease. Indeed, all patients had criteria for severe PMR, but phenomapping showed that PMR is a heterogeneous disease. A novel classification of PMR patients was defined using correlations between phenotypic variables to find patterns in dense data. Once the three groups were identified, the differences between them were striking regarding clinical and echocardiographic characteristics.

The three phenogroups represented three archetypes of PMR according to their demographic, clinical, and echocardiographic characteristics:

• Phenogroup-1 represented low-risk patients (less risk factors, less dilatation of the ventricle…) but their main characteristics did not differ from those of the overall population.

• Phenogroup-2, which represented intermediate-risk patients, included predominantly male patients, aged approximately 61 years old, with large body surface areas, who were smokers and with histories of COPD. Their heart remodelling appeared to be more pronounced.

• Phenogroup-3, which represented high-risk patients, included predominantly women who were older, thinner, and smaller, with hypertension and comorbidities such as renal failure or a pre-operative history of AF. Their EuroSCORE was higher than those of the other groups. Few studies have explored gender influence on valvular disease. However, women seemed to have worse outcomes than men after mitral surgery,¹⁵ although this difference can be partially driven by the rate of mitral valve replacement in women. The heart remodelling in phenogroup-3 appeared to be surprisingly less marked than that in Group 2.

Regarding echocardiographic characteristics: LA peak systolic strain is confirmed as a predictor of cardiovascular events in mitral regurgitation¹⁶ and is associated with elevated filling pressures¹⁷. E-wave velocity was also the highest in Group 3 (150 ± 44 cm/s): it may be explained by both an important mitral regurgitation (increasing stroke volume) and a more severe diastolic dysfunction in these patients, as described previously.¹⁸ LV diastolic dysfunction in PMR is known to result in poor outcomes¹⁹ and seemed to predict better cardiovascular outcomes than LVEF (mean pre-operative LVEF was 67%, with no difference between groups). LVEF may remain in the normal range for a long period of time, whereas subclinical myocardial LV dysfunction may be present.²⁰ Therefore, it is necessary to look for other echocardiographic predictors of outcomes, such as increased LA volume,²¹ abnormal LV GLS,⁴ pulmonary hypertension,²² increased E/e’ ratio,²³ or abnormal LA peak systolic strain.¹⁶ Most of these parameters were altered in phenogroup-3 (except LV GLS). Of note, regurgitant volume and EROA were not good predictors of post-operative outcomes.

Phenogroup-2 had global cardiac chamber enlargement (LA and right atrial volumes, LV volume, and RV size) without obvious diastolic dysfunction or even elevated filling pressures. Although cardiac chamber enlargement suggests myocardial remodelling, it did not seem to result in an increased risk of PCE. The enlargement in the right chambers and the lowest RV global strain in phenogroup-2 can also be explained by the important number of patients with COPD (36% of patients had a history of COPD in phenogroup-2).²⁴ The risk of developing PCE was not significantly increased in phenogroup-2. We hypothesized that phenogroup-2 likely included intermediate-risk patients.

Pre-operative AF is known as an important predictor of cardiovascular outcomes,²⁵ which was confirmed in this study, where 45% of phenogroup-3 patients, who presented the highest risk of PCE occurrence, had pre-operative paroxysmal AF. The high incidences of pre-operative AF in this group can be explained by several factors. LA dilatation is considered as an important risk factor for the occurrence of AF,²¹ and indexed LA volumes in phenogroup-3 (mean 58 ± 19 mL/m²) were close to the cut-off of 60 mL/m².²¹ Nevertheless, left atrium was also dilated in Group 2, although these patients were less likely to have post-operative long-term AF than those in Group 3. Therefore, an enlarged LA is not sufficient to explain the highest prevalence of AF in Group 3. Phenogroup-3 patients were older¹⁰ and were also more likely to have hypertension and chronic renal failure.¹⁰ LA peak systolic strain was reduced in phenogroup-3, reflecting an LA dysfunction (reduced reservoir function).²⁶ Finally, pre-operative AF was a predictor of post-operative AF (Supplementary data online, Figure S2).²⁷

Perspectives

Determining optimal timing for surgery in severe PMR is still problematic, and is currently based on symptoms, severity of the regurgitation and its impact on LV volume and systolic function. Predictors of outcome currently considered in regurgitant valvular diseases are LV increased volumes and reduced LVEF, indicating an alteration of myocardial contractility. However, reduced LVEF is often a late consequence of valve dysfunction and may even imply irreversible myocardial injury.

Machine learning identified three phenogroups with different prognoses. Therefore, the management of patients with severe PMR could be improved, as high-risk patients can now be identified earlier. We suggest that high-risk patients should be carefully monitored (i.e. by more frequent visits to their usual cardiologist). Sinus rhythm should be maintained as long as possible, and paroxysmal AF can be detected by Holter monitoring. Pulmonary vein isolation could be performed during valve surgery in patients with pre-operative paroxysmal AF to prevent post-operative AF.¹⁰

Mitral surgery may be performed early in (asymptomatic) phenogroup-3 individuals to prevent the AF-occurrence risk and can be encouraged by early detection of the cardiac consequences of PMR by performing exercise echocardiography.²⁸

Study limitations

We acknowledge several limitations. Hierarchical clustering led to an imbalanced number of patients between groups. This imbalance impeded definite conclusions about the value of phenomapping in PMR and resulted in a lack of statistical power, particularly because of the small sample size. The retrospective design of our study also implies that we had to perform an imputation procedure to deal with missing values and to exclude several patients having extreme values due to mistyping in electronic health records. Therefore, our results should be considered as hypothesis generating and interpreted with caution.

To confirm the reproducibility of our results (i.e. the identification of phenogroups, outcomes, and risk-levels), we will have to further control our results with a validation cohort.

Conclusion

In this study, PMR was confirmed as a heterogeneous clinical disease, but this heterogeneity may be resolved using a dedicated algorithm to perform phenomapping analysis. Phenotypic data can be grouped in clusters using machine learning to identify three phenogroups of patients with different prognoses.

Acknowledgements

The authors thank the study nurses from the CIC-IT1414 but also the CORECT programmes and the scientific committee of the CHU-Rennes for their support.

Conflict of interest: none declared.

References