WHAT IS ALREADY KNOWN ON THIS TOPIC

Heart failure (HF) is a common complication of hypertrophic cardiomyopathy (HCM), yet prediction of worsening HF using clinical measures is limited. Moreover, the mechanisms by which patients with HCM develop worsening HF have not been elucidated.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICYIntroduction

Hypertrophic cardiomyopathy (HCM) is one of the most common inherited cardiac disorders, with a prevalence of 1 in 200–500 in the USA.1 HCM often causes heart failure (HF), which leads to substantial disability and reduced quality of life.2 Indeed, among patients diagnosed at <40 years of age (about one-third of the HCM population), 65% experience lifestyle-limiting HF during their lifetime3 and about 5% require transplant or die from HF.4 Identification of patients with HCM who are at high risk of having worsening HF is clinically important because the treating physicians can perform more frequent monitoring of modifiable factors that worsens HF (eg, atrial fibrillation,5 obesity6) and intensify medical treatments. Especially for patients with non-obstructive HCM whose symptoms are refractory to medications, heart transplantation is the only definitive treatment option7 and earlier referral may improve outcomes.4 The emergence of targeted therapies, including cardiac myosin inhibitors, also highlights the urgent need for accurate risk stratification, as treatment in earlier stages of HF could prevent further left ventricular (LV) remodeling8 9 and progression of disease.10 However, currently no tools are available to predict which patients will experience worsening HF in HCM. Furthermore, the underlying molecular mechanisms by which patients with HCM develop HF remain incompletely understood.11

Proteomics profiling is a novel technology that simultaneously measures the concentration of thousands of proteins in blood or tissue samples. It has been applied to develop algorithms to distinguish patients with HCM from controls.12 13 Plasma proteomics profiling has also been used to predict late gadolinium enhancement (LGE) on cardiac magnetic resonance imaging (CMR) among patients with HCM.14 However, proteomics profiling has never previously been used to predict worsening HF among patients with HCM. Thus, in this multi-centre, prospective cohort study, we aimed to: (1) determine whether comprehensive plasma proteomics profiling could predict worsening HF among patients with HCM in both training and external test sets for validation and (2) identify signalling pathways that are differentially regulated between patients with HCM who subsequently develop HF and those who do not.

MethodsStudy design and sample

A multicentre, prospective cohort study was conducted among adult patients, age ≥18 years, with a diagnosis of HCM. Patients were recruited from the inpatient and outpatient settings at Massachusetts General Hospital (MGH) and Columbia University Irving Medical Center (CUIMC), between October 2015 and September 2019. Patients with HCM who followed at MGH constituted the training set for derivation of the prediction model, whereas those who followed at CUIMC comprised the external test set for validation. Written informed consent was obtained from all participants, and the study was performed under protocols approved by the CUIMC Institutional Review Board (AAAR5873) and the Mass General Brigham Human Research Office/Institutional Review Board (2015P001953). The investigation conforms with the principles outlined in the Declaration of Helsinki.

Outcome measures

The primary outcome of worsening HF was defined as an increase in New York Heart Association (NYHA) functional class by at least one class during the follow-up period. NYHA class was reassessed at the time of each clinical visit. If a patient had a worsening of their NYHA class compared with baseline, the endpoint was considered positive at the first date the NYHA class was noted to have increased. Subsequent NYHA class assessments were not tracked in such cases. Secondary outcomes included (1) progression to either NYHA class III or IV symptoms in a cohort limited to those with NYHA class I or II symptoms at enrolment and (2) hospitalisation for HF. Outcomes were prospectively adjudicated by two attending cardiologists blinded to the proteomics profiling data. Patients who developed worsening symptoms due to non-HCM-related causes, such as severe aortic stenosis, viral myocarditis, liver failure, kidney disease or advanced lung disease, were not considered to have the primary or secondary outcomes of interest.

Development of a proteomics-based model to predict worsening HF

Given the high-dimensional nature of the proteomics data, the Boruta algorithm was first applied for feature selection to identify proteins that were important predictors of worsening HF. Randomly permuted shadow features are created by the Boruta algorithm and the original features are compared with the shadow features, resulting in selection of only those original features with a higher importance than the randomised shadow features. A hyperparameter tuning grid was developed to specify the best hidden component and threshold parameter. Using the important predictors identified by the Boruta algorithm, a random forest machine learning (ML) model was developed in the training set of patients from MGH using fivefold cross validation with the R caret package. The ML model was then applied to predict which patients would develop worsening HF in the external test set of patients from CUIMC. Area under the receiver-operating characteristic curve (AUC) in the test set was calculated using the ROCR package. For all secondary outcomes, a new model using the same proteins identified by the Boruta algorithm was developed in the training set and its performance was assessed in the test set.

Pathway analysis

To address the second objective of the study, an analysis of signalling pathways differentially regulated between patients with worsening HF and those without worsening HF was performed in the entire study cohort. The Mann-Whitney-Wilcoxon test was used to determine statistically significant (univariable p<0.05) differences between the concentrations of each protein in the group of patients who developed worsening HF compared with those who did not. Using proteins with different concentrations, pathway analysis was performed using STRING V.11 (String Consortium, Europe) to identify the canonical pathways that were differentially regulated between the two groups.15 Associations between the significant proteins and canonical pathways from the Kyoto Encyclopaedia of Genes and Genomes database were determined by examining the ratio of the number of proteins that map to a pathway divided by the total number of proteins that map to the pathway. Pathways with false discovery rate (FDR) <0.001 with at least five associated proteins were considered positive (ie, dysregulated).

Additional methods on univariable, sensitivity and survival analyses as well as comparison with a clinical ML model are available in the online supplemental methods.

Results

Initially, 407 patients with HCM underwent proteomics profiling; 18 patients were lost to follow-up, resulting in a total of 389 patients included in the final analytic cohort. There were 268 patients enrolled through MGH (the training set) and 121 patients enrolled through CUIMC (the test set). Baseline characteristics of the cohort stratified by the primary outcome of worsening HF are shown in table 1 and online supplemental table 1. There was no difference in the prevalence of pathogenic or likely pathogenic gene mutations between the two groups (table 1, online supplemental table 2). Patients in the worsening HF group had higher levels of troponins and B-type natriuretic peptide (online supplemental table 3).

Table 1

Baseline clinical characteristics of the study sample

During a median follow-up of 2.8 years (IQR: 1.8–5.1), 68 (17%) patients developed worsening HF symptoms (online supplemental table 4). Worsening HF occurred in 51 (19%) patients in the training set and 17 (14%) patients in the test set. Event rate was 5.89 per 100 patient years in the training set and 4.65 per 100 patient years in the test set. Three patients in the worsening HF group underwent heart transplantation during the follow-up period. In addition, six patients underwent alcohol septal ablation and 15 patients underwent septal myectomy.

The proteomic profiles were well-separated between patients who developed worsening HF and those who did not (online supplemental figure 2). There were 11 proteins identified as important in distinguishing patients who developed the primary outcome of worsening HF from those who did not based on the Boruta feature selection algorithm (figure 1). Using the 11-protein proteomics-based model derived from the training set, the AUC was 0.87 (95% CI: 0.76 to 0.98) in the test set (figure 2). Sensitivity was 88% (64%–99%), specificity was 84% (75%–90%), positive predictive value (PPV) was 47% (34%–89%) and negative predictive value (NPV) was 98% (91%–99%). When patients in the test set were classified into high-risk (32 patients) and low-risk (89 patients) groups according to the proteomics-based predictive model derived from the training set, the high-risk group had a significantly higher rate of developing worsening HF compared with the low-risk group (HR=22.3; 95% CI, 5.1 to 97.7; p<0.0001; figure 3).

Figure 1

Relative importance of proteins included in the proteomics-based model to predict worsening heart failure in patients with hypertrophic cardiomyopathy. ER, endoplasmic reticulum; ILGF, insulin-like growth factor.

Figure 2

Receiver-operating-characteristic curve to predict worsening heart failure in the external test set for validation, using the proteomics-based prediction model derived from the training set. AUC, area under the receiver-operating characteristic curve.

Figure 3

Kaplan-Meier curve for the risk of worsening heart failure in the external test set for validation, using the proteomics-based prediction model derived from the training set.

In sensitivity analyses excluding patients with cystatin-C, alanine aminotransferase and C reactive protein more than two times the SD above the mean, the AUC remained similar (online supplemental figure 3A–C).

There were 32 patients (13%) in the training set and 13 patients (14%) in the test set who developed the secondary outcome of progression to NYHA class III or IV symptoms from a baseline of NYHA class I or II. In a proteomics-based model derived from the training set using the same 11 proteins from the primary outcome model, the AUC was 0.88 (95% CI 0.79 to 0.97) for predicting progression to NYHA class III or IV symptoms in the test set (online supplemental figure 4). There were 39 patients (15%) in the training set and 17 patients (14%) in the test set who developed the secondary outcome of hospitalisation for HF. In a proteomics-based model derived from the training set with the same set of 11 proteins, the AUC was 0.80 (95% CI 0.69 to 0.92) for predicting HF hospitalisation in the test set (online supplemental figure 5).

In an ML model developed in the training set using all baseline clinical variables from table 1 and online supplemental table 1 that differed significantly between groups (p<0.05), the AUC was 0.68 (95% CI: 0.54 to 0.83) in the test set (online supplemental figure 6). Sensitivity was 71% (44%–90%), specificity was 62% (51%–71%), PPV was 23% (17%–52%) and NPV was 93% (81%–95%). Based on the DeLong test, the proteomics-based model performed significantly better than the clinical model in predicting which patients with HCM would develop worsening HF (p=0.04). In a combined clinical and proteomics ML model developed in the training set using all significant baseline clinical variables and the 11 proteins from the proteomics model, the AUC was 0.89 (95% CI: 0.83 to 0.95) in the test set (online supplemental figure 7).

There were 1273 proteins that were differentially regulated between patients with and without the primary outcome of worsening HF with p<0.05 (online supplemental figure 8). Pathway analysis of these proteins revealed that the Ras-MAPK pathway (FDR<0.00001) and its upstream PI3K-Akt pathway (FDR<0.00001) were dysregulated in patients who subsequently developed worsening HF (figure 4). Other pathways upstream and downstream of the Ras-MAPK pathway, including ErbB, HIF-1, FcεRI, Rap1 and FoxO pathways, as well as pathways previously associated with cardiac hypertrophy (eg, NF-κB) were also dysregulated (figure 4).

Figure 4

Pathway analysis of proteins that were differentially regulated between patients who developed worsening heart failure and those who did not. Fcε RI, Fc-epsilon-RI; FoxO, forkhead box O; HIF-1, hypoxia-inducible factor; MAPK, mitogen-activated protein kinase; NF-κB, nuclear factor kappa-light-chain-enhancer of activated B cells; PI3K, phosphoinositide 3-kinase; VEGF, vascular endothelial growth factor.

DiscussionSummary of findings

In this multicentre, prospective cohort study of 389 patients with HCM, our model based on comprehensive plasma proteomics profiling of 4986 proteins demonstrated robust ability to predict worsening HF among patients with HCM. Importantly, using a small panel of 11 plasma circulating proteins, our model had high predictive ability which was reproducible in the independent test set for validation, suggesting high potential for application in clinical settings. We also identified key signalling pathways that were dysregulated in patients with HCM who developed worsening HF compared with those who did not—most notably the Ras-MAPK pathway, along with several of its associated upstream and downstream pathways. This is the first study to apply proteomics profiling to predict worsening HF among patients with HCM.

Results in context

HF is one of the most common adverse cardiovascular events among patients with HCM, and it is a strong driver of morbidity and mortality in this population.3 Based on findings from a recent study of ~5000 patients with HCM, among patients diagnosed with HCM at age <40 years, the lifetime risk of HF was 65% compared with a lifetime risk of HF of 20%–45% for a 40-year-old without HCM.3 The onset of HF is typically several years after the diagnosis of HCM, occurring most commonly between ages 50 and 70 years.3 This suggests that there is room for intervention if at-risk individuals can be identified prior to progression to HF.

Early identification of patients with HCM at risk for HF progression is essential to allow for implementation of targeted interventions and therapeutics known to improve HF outcomes or slow disease progression. Primarily, this population of patients can be monitored more closely for co-morbid conditions known to worsen HF outcomes. In a study of nearly one million patients hospitalised with HF, AF or both, patients with HF in AF were found to be at increased risk for mortality and prolonged hospitalisation compared with patients with isolated AF or HF.16 Similarly, obesity may be a modifiable risk factor for worse outcomes in HF. For instance, in a study of over 1600 patients with obesity with HF, bariatric surgery was found to be associated with a decreased rate of HF exacerbations.6

Importantly, the newest class of medications for the management of HCM, cardiac myosin inhibitors, may be efficacious in preventing LV remodeling8 9 and slowing disease progression10 in symptomatic patients with HCM. Among patients with obstructive HCM, mavacamten was associated with improved exercise capacity, LV outflow tract obstruction, NYHA functional class and health status.10 Among patients with non-obstructive HCM, for whom treatment options are even more limited, mavacamten use was associated with a reduction in biomarkers associated with myocardial wall stress, suggesting potential benefit in this population as well.17 It will be essential to identify patients with the greatest potential benefit with this new class of medications. Moreover, identification of a high-risk HCM subset will allow for the investigation of HF medication regimens in this specific cohort, as current management strategies for HF in patients with HCM are limited by a paucity of data.18 Finally, patients at the highest risk for HF progression may be more promptly referred for transplant evaluation with improved risk stratification.4

Our biomarker-based ML model demonstrated superior predictive ability compared with current, clinically available biomarkers. In a prior study assessing the use of high-sensitivity cardiac troponin T for predicting a composite of hospitalisation for HF, embolic stroke, ventricular arrhythmia, progression to NYHA class III or IV symptoms and cardiovascular death among patients with HCM, the AUC was 0.77, with a PPV of 32% and NPV of 93%.19 N-terminal pro-brain natriuretic peptide (NT-proBNP) demonstrated similar predictive ability for the outcome of all-cause mortality or transplant among patients with HCM, with an AUC of 0.78 (95% CI 0.73 to 0.84), PPV of 29% to 62% and NPV of 76% to 96%, depending on the NT-proBNP threshold used.20 Comparatively, our 11-protein model demonstrated higher predictive ability for the outcome of worsening HF, with an AUC of 0.87 (95% CI: 0.76 to 0.98), PPV of 47% and NPV of 98%, even in the independent test set for validation.

The Ras-MAPK pathway and its role in cardiac hypertrophy

Pathway analysis identified an association between dysregulation of the Ras-MAPK pathway and risk of worsening HF among patients with HCM in our cohort. Germline mutations in Ras signalling are known to cause RASopathies, which are a group of syndromes often associated with an HCM-like cardiac phenotype.21 Ras mRNA expression levels have also been found to be increased among patients with HCM.22 Prior studies from our group comparing HCM cases to controls with LV hypertrophy demonstrated upregulation of the Ras-MAPK pathway among patients with HCM, particularly in those with higher baseline NYHA class symptoms and left atrial diameter.12 13 Similarly, dysregulation of the Ras-MAPK pathway and related pathways was identified in prior studies using plasma proteomic profiles to predict LGE on CMR among patients with HCM.14 This is the first study, however, to prospectively demonstrate an association between dysregulation of the Ras-MAPK pathway and not only HCM pathogenesis but also progression of HF symptoms in HCM.

Strengths of the present study

Several strategies were employed in this study to reduce the rate of false-positive and false-negative findings and to improve the internal and external validity of the study. First, we validated the proteomics-based ML model in a test set of patients from a different institution than that of the training set, which serves to confirm the robustness and external validity of the model while reducing the possibility of model overfitting. Second, this study includes proteomics profiling of the greatest number of patients with HCM (n=389) with the most comprehensive proteomics profiling to date (4986 proteins), which enhances the internal validity of the study and reduces the likelihood of false-negative results, namely, missing important proteins and signalling pathways.12 13 23 Third, in order to reduce false-positive findings, we only considered pathways with FDR<0.001 to be positive. The use of pathway analysis further strengthens the biological plausibility of our study and reduces the likelihood of false-positive findings, as we identified proteins that are known to be connected biologically.24

Potential limitations

There are several potential limitations in our study. First, the study was performed at two high-volume, tertiary care centres, both of which serve as referral centres for patients with advanced HCM. This may limit the generalisability of our findings to patients with less advanced disease or those treated at smaller centres. Second, misclassification of the outcome was possible, although the prospective study design and rigorous event adjudication minimises the possibility. Third, the possibility of informative censoring could not be excluded. Fourth, the present study supports but does not prove causality between the differentially regulated proteins and HCM pathogenesis or disease progression. Fifth, samples are available only from one time point, and repeated measurements could be informative. Sixth, the frequency of pathogenic genetic variants was relatively low, likely because our patient population included sporadic or late-onset cases. Lastly, it is possible that some of the proteins used in this analysis were not derived from the heart.

Conclusions

In this multicentre prospective cohort study, using comprehensive proteomics profiling in HCM, we identified a small panel of plasma proteins with high predictive value for worsening HF in patients with HCM. Moreover, we identified signalling pathways that were dysregulated in patients who subsequently developed worsening HF compared with those who did not, thus implicating the role of these pathways in the progression to more severe disease. The present study serves as the first step to improve risk stratification with circulating biomarkers.

Data availability statement

Data are available upon reasonable request. The data that support the findings of the present study are available from the corresponding author upon reasonable request.

Ethics statementsPatient consent for publication

Not applicable.

Ethics approval

This study involves human participants and was approved by Columbia University Irving Medical Center Institutional Review Board (AAAR5873) and the Mass General Brigham Human Research Office/Institutional Review Board (2015P001953). Participants gave informed consent to participate in the study before taking part.

Acknowledgments

Preliminary findings from this study were presented as an abstract at the European Society of Cardiology Congress 2022: Lumish H, Liang LW, Hasegawa K, Maurer M, Fifer MA, Reilly MP, Shimada YJ. Use of serum proteomics profiling to predict worsening heart failure events in patients with hypertrophic cardiomyopathy. European Society of Cardiology Congress. arcelona, Spain, August 2022.