Reviews
Risk Stratification Towards Precision Medicine in Heart Failure ― Current Progress and Future Perspectives ―
2021 Volume 85 Issue 5 Pages 576-583
Details
2021 Volume 85 Issue 5 Pages 576-583
Clinical risk stratification is a key strategy used to identify low- and high-risk subjects to optimize the management, ranging from pharmacological treatment to palliative care, of patients with heart failure (HF). Using statistical modeling techniques, many HF risk prediction models that combine predictors to assess the risk of specific endpoints, including death or worsening HF, have been developed. However, most risk prediction models have not been well-integrated into the clinical setting because of their inadequacy and diverse predictive performance. To improve the performance of such models, several factors, including optimal sampling and biomarkers, need to be considered when deriving the models; however, given the large heterogeneity of HF, the currently advocated one-size-fits-all approach is not appropriate for every patient. Recent advances in techniques to analyze biological “omics” information could allow for the development of a personalized medicine platform, and there is growing awareness that an integrated approach based on the concept of system biology may be an excessively naïve view of the multiple contributors and complexity of an individual’s HF phenotype. This review article describes the progress in risk stratification strategies and perspectives of emerging precision medicine in the field of HF management.
Heart failure (HF) is one of the most common causes of hospitalization and healthcare expenditure worldwide, at more than US$100 billion.1 With aging of the population, the global prevalence of HF is on the rise, currently accounting for over 200,000 annual hospital admissions in Japan and 1 million in the US.2,3
Generally, in the management of patients with HF, understanding their prognosis is important. A physician’s understanding of the risks and sharing the trajectory of the disease with patients and their families are important aspects of building physician-patient relationships. Knowledge of future risks will also support decisions regarding optimal treatment strategies, especially in high-risk patients who have a higher incidence of adverse events such as death and hospitalization due to worsening HF, and is essential for determining when to introduce therapeutic strategies, such as pharmacotherapy, cardiovascular devices, closer monitoring, or advanced care planning in palliative care. Meanwhile, identifying low-risk patients helps reduce patient anxiety and curb unnecessary high-cost healthcare resource utilization, which is also important from a health economics perspective.
However, HF is a cardiovascular disease in which it is hard to predict the effectiveness of medications and devices and to develop personalized medicine based on prediction outcomes because of the repeated functional decline due to acute exacerbation and partial recovery following initial treatments, the possibility of death at the time of exacerbation, and sudden cardiovascular death during the course of the chronic compensated phase of HF.4 Although many studies on outcome prediction in HF have been published, outcome prediction using single factor, such as a clinical value or a single biomarker, is inadequate. This has led to the development of statistical prognostic models.
Notably, given the large heterogeneity of HF patients with multiple etiologies, cardiac phenotypes, and other comorbidities, the currently advocated “one-size-fits-all” approach (i.e., population medicine) is not appropriate for every patient. In fact, large randomized clinical trials have shown an overall beneficial treatment effect, but one that is not necessarily beneficial in every patient subgroup. Although a successful trial simply means that, on average, significantly more people benefited, many patients will never experience the endpoints for treatment success and treatment may have made no difference to their outcome. That is, the residual high mortality and event rates reflect groups of patients who do not receive sufficient benefit from existing therapies.
Recently, it was proposed that future significant improvements in clinical outcomes could be achieved when therapy is individualized and tailored to a patient’s biological profile. In 2015, the Precision Medicine Initiative announced by President Obama of the US increased the popularity of the precision medicine concept, and related research has accelerated in the field of cardiovascular disease. Precision medicine is a concept that further deepens personalized medicine by integrating a huge amount of biological “omics” information (e.g., genome, proteome, metabolome) that applies cutting-edge basic research to large-scale clinical data. Precision medicine would enable the selection of the optimal treatments for individual patients based on analyses of these next-generation concepts of real-world data.
In this review, we highlight recent advances in risk stratification and future perspectives towards precision medicine for HF.
To date, although numerous prognostic models have been developed from secondary analyses of a number of randomized clinical trials and real-world registries mainly from Europe and the US, there are actually few models that have sufficient ability to predict outcome (i.e., a C-index of ≥0.70) based on appropriate statistical analyses with a high-quality dataset and a large enough sample size. For example, among the models targeting patients with acute decompensated HF, the Organized Program To Initiate Lifesaving Treatment In Hospitalized Patients With Heart Failure (OPTIMIZE-HF) and the Get With The Guidelines Heart Failure (GWTG-HF) models are widely available.5,6 In each model, risk scores can be calculated from factors related to the prognosis, the predicted in-hospital mortality rate is calculated according to the total score, and it is possible to grasp the short-term risk of death at the time of admission for acute HF.
Given the fact that healthcare systems and hospitalization thresholds vary widely across world regions, it is likely that risk scores will perform differently in different populations. Indeed, our recent study suggested that these scores appeared moderately accurate and unbiased when applied to British patients but consistently overestimated mortality in Japan.7 In terms of chronic HF, the Seattle Heart Failure Model (SHFM) is a well-known classic model for predicting long-term mortality.8 The SHFM was derived from data for 1,153 patients in the Prospective Randomized Amlodipine Survival Evaluation (PRAISE) study and the risk can be estimated on a website; however, the performance of the SHFM is also slightly poorer when applied to Japanese patients, and recalibration would be preferable when using it.9
As noted above, it is worth considering how these models were developed and understanding the association between model characteristics and its performance. Ouwerkerk et al conducted aggregate data meta-analysis using 117 systematically identified HF prognostic models reported in 55 papers published worldwide.10 The mean C-index of the identified mortality models was modest at 0.66. Ouwerkerk et al also reported that the strongest predictors were blood urea nitrogen and serum sodium, and that the number of predictors included in the final models was significantly related to the C-statistics.10 As for study characteristics, those models derived from studies with a prospective design, claims data, and large sample size showed higher C-statistics than models from retrospective studies and data based on medical records. Meanwhile, diagnosis of acute or chronic HF, age, and the percentage of males in the derivation cohort were not significantly related to model performance.10 Rahimi et al also conducted an aggregate data meta-analysis reviewing 64 models in 48 studies;11 the reported C-statistics ranged from 0.60 to 0.89. Rahimi et al reported that mortality models had higher C-statistics than models using other outcomes.11 In contrast to the study of Ouwerkerk et al,10 Rahimi et al found that sample size (small [<1,000], medium [<5,000], or large), data source (trial data, primary data, patient records, or administrative data), and study design (prospective or retrospective) were not significantly associated with model performance.11
Recently, we performed a systematic review and external validation study to investigate whether bias or characteristics of the original derivation studies determined model performance by using published critical appraisal tools (i.e., the Quality In Prognosis Studies [QUIPS] and the CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies [CHARMS]).12 In that study, when externally applying the QUIPS and CHARMS tools to 224 systematically identified models from 6,354 eligible studies of a Japanese HF registry (n=3,452), only optimal sampling, assessed by an adequate and valid description of the sampling frame and recruitment details to collect the population of interest, was significantly associated with high model performance after adjustment for relevant study characteristics, such as data source, scale of the study, the stage of the illness, and the study year (standardized β=0.24; 95% confidence interval [CI] 0.07–0.40; P=0.01). Therefore, optimal sampling in the derivation studies would be a key determinant of the performance of HF prognostic models, rather than similarities in characteristics between studies, and should be taken into account when developing a prognostic model (Figure 1).
Conceptual image of model performance gap and its determinants between derivation and validation studies. The example model is a combination of two predictors. There is a model performance gap when tested in the original population (accuracy to identify the positive outcome=11/23) and in a different population (accuracy=8/23). The gap may be attributed partly to the difference of the two populations with different predictor-outcome relationships and different outcome incidence. Other potential determinants include characteristics of the original derivation study, such as source of data, participants and sample size, and quality of the original study. QUIPS and CHARMS checklist provide lists of potential determinants. Among these determinants, optimal sampling assessed by an adequate and valid description of the sampling frame and recruitment details to collect the population of interest could be a significant determinant of higher model performance. CHARMS, the CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies; QUIPS, the Quality In Prognosis Studies.
In the past several decades, experimental and clinical studies of biomarkers in HF have been published, primarily focusing on pathophysiological processes, such as myocardial stress or injury (e.g., B-type natriuretic peptide [BNP], N-terminal pro BNP [NT-proBNP], troponin), neurohormonal activation (e.g., renin, angiotensin II, endothelin-1), remodeling (e.g., matrix metalloproteinases, C-reactive protein), and comorbidities (e.g., cystatin C, albumin, iron, ferritin, erythropoietin, urinary N-acetyl-β-glucosaminidase, urinary sodium), and these prognostic values despite significant overlaps.13–20 Some markers of myocardial stress or injury, including BNP and troponin, can contribute to improved performance of prognostic models, and indeed the recently developed models have significantly better C-indices than those developed previously.21
Notably, the addition of these HF-related biomarkers to existing risk models can improve the performance of these models. For example, Sawano et al demonstrated that the Meta-analysis Global Group in Chronic Heart Failure (MAGGIC) score had a modest discrimination (C-index=0.71; 95% CI 0.67–0.74) and good calibration (R2=0.97), but there was constant overestimation for 1-year mortality if the model was applied to Japanese acute HF patients.22 Interestingly, when the BNP concentration was added to the original MAGGIC variables, the discriminative value of the model improved (C-index=0.74; 95% CI 0.70–0.78) and the overestimation of mortality disappeared.22
Furthermore, biomarkers related to malnutrition and nutritional score can improve the performance of prognostic models. We previously reported that a model that included serum albumin and total cholesterol had higher discriminative performance among the representative mortality prediction models, and that albumin was the strongest determinant of model performance.23 Importantly, the addition of the nutritional score or index (e.g., the Controlling Nutritional Status [CONUT] or the Geriatric Nutritional Risk Index [GNRI]) to the OPTIMIZE-HF model significantly increased the C-index from 0.75 to 0.77 and from 0.68 to 0.74, respectively.24,25 Moreover, the net reclassification was 21% for all-cause death, 27% for survival, and 49% overall after addition of the CONUT, and 27% for events, 33% for non-events, and 60% overall after addition of the GNRI.24,25
Shiraishi et al developed a simple model (SOB-ASAP score) for predicting the risk of in-hospital death using data from 4,351 patients with acute HF in 3 registries in Japan (Figure 2).26 BNP and serum albumin, in addition to age, systolic blood pressure, blood urea nitrogen, and serum sodium, were included in the final model to predict the risk of in-hospital death. This model showed excellent discrimination (C-index=0.82) and acceptable calibration. The predictive performance of the model was confirmed in an external validation dataset (n=1,682).
SOB-ASAP risk score.26 BNP, B-type natriuretic peptide; BP, blood pressure; BUN, blood urea nitrogen; NT-proBNP, N-terminal pro-B-type natriuretic peptide.
As mentioned above, most of the value described for newer biomarkers in HF relates to prognostication. Nevertheless, it should be noted that a biomarker is only useful if it informs specific change in clinical practice that leads to an improvement in prognosis. This precision strategy using biomarkers, mostly BNP and NT-pro BNP, has been tested in many trials. Although a meta-analysis of 8 trials indicated the superiority of BNP-guided treatment with marginal significance (hazard ratio [HR] 0.82; 95% CI 0.67–1.00),27 there were significant heterogeneities regarding the study design and demographics of eligible patients, such as age, HF phenotype (HF with reduced ejection fraction [HFrEF] vs. HF with preserved ejection fraction [HFpEF]), threshold of target BNP concentration, and change in BNP concentrations, which led to inconsistent results.28 The Guiding Evidence Based Therapy Using Biomarker Intensified Treatment in Heart Failure (GUIDE-IT) study examined whether an NT-proBNP-guided treatment strategy improves clinical outcomes compared with usual care in high-risk patients with HFrEF and found that not only was an NT-proBNP-guided therapy strategy no more effective than a usual care strategy in improving outcomes, but also that decreases in NT-proBNP concentrations did not differ significantly between the groups.29
Together, these results suggest that natriuretic peptides could provide more accurate prognostication, identify high-risk patients, and improve the performance of risk models for HF; however, the efficacy of natriuretic peptide-guided HF therapy would be limited in current clinical practice. For all other biomarkers, it remains unclear whether specific therapy changes should be made in response to an abnormal result. This strategy of a more personalized (i.e., precision) approach using biomarkers is already evident in the field of oncology, but is just beginning to gain momentum in HF.
In recent years, technology to analyze biological information has emerged, and it can be said that the possibility of constructing personalized medicine for heterogeneous diseases is approaching. In fact, there is broad heterogeneity in etiologies, demographics, comorbidities, and clinical outcomes of individuals with HF, thereby highlighting the urgent need for precision phenotyping in contrast with the currently advocated one-size-fits-all approach. Due to recent advances in panomics (e.g., genomics, transcriptomics, metabolomics, and proteomics; technologies and data analyses that provide in-depth clinical, biological, and molecular phenotyping), there is growing awareness that the integrated approach based on the concept of system biology may be an excessively naïve view of the multiple contributors and complexity of an individual’s HF phenotype. This concept may also help stratify the risk of HF progression in Stage A HF patients, despite a paucity of data. Indeed, a high genome-wide polygenic risk score calculated using genotype data for more than 350,000 single-nucleotide polymorphisms was independently associated with the development of ischemic stroke in the general Japanese population.30
Deep phenotyping can be defined as the precise and comprehensive analysis of phenotype abnormalities in which the individual components of the phenotype are observed and described.31 By performing network analysis of the enormous amount of data obtained as a result of deep phenotyping, it would be possible to cluster a group of patients with closer biological phenotypes. This novel clustering based on the concept of system biology would lead to the development of more accurate prognostic models and new biomarkers that can reflect the clustering with high accuracy and low cost. Furthermore, the newly developed clusters may provide appropriate candidates for clinical trials by predicting responses to drugs, leading to a reduction in the cost of trials (Figure 3).32,33
Phenotyping approach for precision medicine in heart failure. Although subjects could have similar endophenotypes in clinical and laboratory data (e.g., left ventricular ejection fraction, biomarkers), these may have heterogeneous biological backgrounds (Upper panel). Using a novel precision phenotyping, subjects undergo deep phenotyping with network analyses of multi-omics biological data in addition to clinical and laboratory data. This phenotyping approach may allow individuals to be reclassified into new clusters that have similar biological backgrounds, leading to the novel precision strategies such as the development of outcome prediction model, the optimization of medication and identification of new biomarkers (Lower panel).
Currently, HF guidelines recommend the use of multiple evidence-based medications consisting of angiotensin converting enzyme inhibitors (ACEIs), angiotensin receptor blockers (ARBs), angiotensin receptor-neprilysin inhibitors (ARNI), β-blockers, mineralocorticoid receptor agonists, hydralazine/isosorbide dinitrate, and ivabradine, which have been shown to improve not only cardiac function and clinical outcomes, including worsening HF and death, but also quality of life in patients with HFrEF.34–37 However, these guideline-directed medical treatments (GDMTs) are not always effective in every patient subgroup. For example, ACEIs/ARBs reduced the mortality rate in Stage 3B (HR 0.39; 95% CI 0.19–0.83) but not Stage 4 or 5 (HR 1.08; 95% CI 0.57–2.03) chronic kidney disease based on a propensity score-matched analysis in a Japanese hospitalized HF registry.38 Furthermore, another study demonstrated that all-cause death occurred in approximately 5% of European HFrEF patients within 1 year of those who received both ACEIs/ARBs and β-blockers and were successfully uptitrated (>50% of target dose).39
Given the vast heterogeneity in demographics and comorbidities in HF, there are the residual groups of patients with high mortality who do not receive sufficient benefit from GDMTs. To identify these patients, individualized responses to a drug can be studied with deep phenotyping using panomics in addition to comprehensive clinical data. Notably, several genetic variations are related to responses to the GDMTs (Table). Rigat et al reported that angiotensin-converting enzyme (ACE) gene insertion/deletion (I/D) polymorphisms were associated with the level of enzyme activity, and that deletion of the allele resulted in higher levels of ACE.40 In addition, the therapeutic response to ACEIs is smaller in patients who are ACE D/D carriers than in those who are ACE I/I carriers. These findings indicate that ACE genotype affects the response to ACEIs in patients with HF.41
| Genotype polymorphism | Phenotype | Response to GDMTs | References |
|---|---|---|---|
| ACE | |||
| I/D polymorphism | Deletion of the allele resulted in higher ACE levels |
Response to ACEIs is smaller in patients with ACE D/D than ACE I/I carriers |
40, 41 |
| Adrenergic receptors | |||
| Ile164 vs. Thr164 β2-adrenoceptor |
Ile164 was associated with lower response to β-blockers |
Greater mortality rate in Ile164 β2-adrenoceptor carriers treated with than without β-blockers |
42 |
| Gln27 vs. Glu27 β2-adrenoceptor |
Gln27 β2-adrenoceptor was resistant to agonist-promoted downregulation of adrenoceptors |
Greater response to carvedilol in Glu27 than Gln27 β2-adrenoceptor carriers |
42, 43 |
| I/D polymorphism in α2c-adrenoceptor |
Deletion of the allele (α2cDel322–325) was associated with dysfunction of autoinhibition of the central and peripheral nervous systems, resulting in increased norepinephrine release |
Patients with α2cDel322–325 and Arg389 β1-adrenoceptor had greater improvement in LVEF after β-blocker treatment than other genotypes |
45, 46 |
ACE, angiotensin-converting enzyme; ACEIs, ACE inhibitors; GDMTs, guideline-directed medical treatments; Gln, glutamine; Glu, glutamic acid; I/D, insertion/deletion; Ile, isoleucine; LVEF, left ventricular ejection fraction; Thr, threonine.
In terms of adrenergic receptor polymorphisms, several genetic variations have been reported to affect responses to β-blockers.42–46 For example, Kaye et al examined the β2-adrenoceptor genotype in HFrEF patients treated with carvedilol and found that the proportion of good responders to carvedilol (defined as an absolute improvement in left ventricular ejection fraction [LVEF] ≥10%) was significantly lower among those who were homozygous for the allele encoding the Gln27 polymorphism than among those who were homozygous or heterozygous for the Glu27 polymorphism (26% vs. 63%; P=0.003).42
Recently, the BIOlogy Study to Tailored Treatment in Chronic Heart Failure (BIOSTAT-CHF) was conducted at 9 institutions in 11 European countries, with more than 2,500 HFrEF patients enrolled. The aim of that study was to characterize the biological pathways related to good and poor responses to GDMTs for HFrEF using the concept of systems biology including omics (genome-wide association study [GWAS] and metabolome) and multibiomarker analyses).47 Although some of the findings primarily related to biomarkers, clinical phenotypes and GDMTs were reported;39,48–50 analyses related to precision models for the effectiveness of GDMTs are on-going.
It is of note that there are significant differences in the contemporary management and understanding of GDMTs for HFrEF between Asian and Western countries. A global registry of acute HF revealed a lower prevalence of GDMTs for HFrEF in Asian than Western countries.51 In terms of the prognostic implications of uptitration of GDMTs, uptitration of β-blockers rather than ACEIs/ARB is important in Asian and European populations.49 However, a prospective randomized stratified trial with 364 Japanese HFrEF patients demonstrated that the dose of carvedilol (2.5, 5, or 20 mg daily) was not associated with either the composite endpoint of all-cause death and hospitalization for cardiovascular diseases and HF or with the improvement in LVEF.52
Ethnic or individual genetic background may determine the responsiveness to GDMTs for clinical outcomes; therefore, we designed a prospective multicenter cohort study with the aim of establishing a platform of precision medicine for Japanese patients with HF (EpidemioLogical Multicenter Study for Tailored Treatment in Heart Failure; ELMSTAT-HF; UMIN Clinical Trials Registry ID: UMIN000039026) and started enrolling patients in January 2020. Our plan is to enroll more than 5,000 HF patients diagnosed using the updated guidelines34,36,37 from both inpatient and outpatient settings across 25 sites, and to follow them for at least 2-year to document adverse events, including death and worsening HF. Blood samples will be collected for HF-related multibiomarker, genome, and metabolome analyses at the time of enrollment, and GDMTs will be encouraged in all eligible patients. Furthermore, serial changes in hemodynamic parameters and cardiac structure and function will be evaluated by echocardiography after 9 and 18 months of GDMT. By performing deep phenotyping with genome and metabolome analyses in addition to clinical data and HF-related multibiomarkers, it will be possible to develop precision models for the responsiveness to GDMTs and clinical outcomes in Japanese patients with HFrEF. Such precision models should be instrumental in developing alternative therapies for patients with a suboptimal response, and therefore improve care for patients with HFrEF.
Precision Medicine for HFpEFHFpEF, which accounts for approximately half of all HF patients, has a similar prognosis to HFrEF and its prevalence continues to rise.53 Meanwhile, large clinical trials testing neurohumoral inhibition for HFpEF failed to show similar effects on mortality as seen for HFrEF,54 because, clinically, HFpEF presents as a diverse syndrome triggered by a variety of comorbidities, cardiac structural abnormalities, and elevated inflammatory mediators with extracardiac manifestations, such as hypertension, diabetes, obesity, and chronic kidney disease.55
In the Treatment of Preserved Cardiac Function Heart Failure with an Aldosterone Antagonist (TOPCAT) study, spironolactone significantly reduced primary endpoints in a subgroup of patients with elevated natriuretic peptide concentrations within 60 days prior to randomization, although no significant difference was observed in the entire study population.56 Thus, a trial to examine the efficacy of ARNI in patients with HFpEF mandated elevated natriuretic peptide concentrations as major inclusion criterion; however, ARNI failed to significantly reduce cardiovascular death or worsening HF compared with valsartan.57
As described above, the treatment strategy thus far has focused on a one-size-fits-all approach, which has worked relatively well for HFrEF despite the heterogeneity of the HFpEF syndrome. More tailored and personalized treatment strategies according to phenotyping would work for patients with HFpEF. Shah et al proposed a phenotype-specific treatment strategy based on 8 mechanism- and predisposition-derived phenotypes, including lung congestion/metabolic risk, arterial hypertension, renal dysfunction, coronary artery disease, chronotropic incompetence, pulmonary hypertension, skeletal muscle weakness, and atrial fibrillation.55 Furthermore, Shah et al performed unbiased machine learning cluster analysis (i.e., phenomapping) with dense phenotypic data from multiple domains (67 continuous variables) and found novel clusters of meaningful, clinically relevant categories of HFpEF patients.58 For more detailed clustering, the concept of system biology with deep phenotyping would be a breakthrough. The ELMSTAT-HF study will enroll HFpEF and HFrEF patients, and will therefore provide novel clustering towards precision medicine (e.g., identification of a cluster that responds well to specific drugs and therapies, and the development of an accurate mortality prediction model).
Perspectives on Data Science for Precision Medicine in HFAlthough the emerging omics techniques together with clinical data open up a new horizon of risk stratification towards precision medicine (i.e., from population to personalized medicine) resulting from an explosion of molecular data in modern biomedical research, these biomedical and omics datasets are complex and heterogeneous, and so it is challenging to extract meaningful findings from this vast amount of information using conventional statistical analyses. Recent progress in machine learning-based approaches has allowed accurate precision in the medical field;59 however, the performance of machine learning strongly depends on preprocessing to extract features.60 In this context, there is an increasing interest in the potential of deep learning techniques that include this feature extraction to develop precision models from these large datasets.60 To date, although no single universally applicable method among several deep learning approaches has been established for analyzing these extremely large datasets, we believe the increasing availability of these kinds of data will accelerate the progress of deep learning technology, which will play an important role in the field of precision medicine for HF in the near future.
None.
T.N. has received a Grant-in-Aid for Scientific Research from JSPS KAKENHI; research grants from the Takeda Science Foundation, the Japan Cardiovascular Research Foundation, the Japan Foundation for Aging and Health, and the Uehara Memorial Foundation; and honoraria from Daiichi Sankyo Co., Ltd. T.A. has received research grants from the Japan Agency for Medical Research and Development and Daiichi Sankyo Co., Ltd.; scholarship funds from Biotronik Japan Co., Ltd., Medtronic Japan Co., Ltd., Win International Co., Ltd., Medical System Network Co., Ltd., and Hokuyaku Takeyama Holdings, Inc.; and honoraria from Daiichi Sankyo Co., Ltd., Ono Pharmaceutical Co., Ltd., Boehringer Ingelheim Japan Co., Ltd., Bayer Pharmaceuticals Co., Ltd., and Bristol-Myers Squibb Co., Ltd. T.N. and T.A. are members of Circulation Journal’s Editorial Team. M.N. has no conflicts of interest to declare.
