Predicting Heart Failure: One Size Does Not Fit All

This article refers to ‘Nuclear magnetic resonance-based metabolomics identifies phenylalanine as a novel predictor of incident heart failure hospitalisation: results from PROSPER and FINRISK 1997’ by C. Delles et al., published in this issue on pages XXX. Due to aging and an increased incidence of cardiovascular risk factors such as hypertension, coronary artery disease and diabetes, the incidence and prevalence of heart failure (HF) will increase. Currently, one in five, both men and women, will develop HF during their lifetime.1 After diagnosis of HF, no cure is available in most cases. Thus, prevention of HF by treating its risk factors as well as identifying individuals at high risk is crucial. Current guidelines mention prevention and prediction of HF as a focal point and this remains an important unmet need within the cardiovascular health field.2 Nevertheless, predicting individuals at risk for developing HF is difficult. The advent of the -omics based studies might provide novel and possibly more clinically relevant risk prediction opportunities. Yet, how do we go forward? And how do lessons previously learned apply to research performed today? Predicting HF is not common place in clinical practice.3 To evaluate the clinical utility and predictive value of a risk prediction model, the C-index is an important tool. Generally, a C-index of >0.8 can be considered good. However, a good prediction model in clinical practice is highly dependent not only on statistical considerations per se, but also on clinical and cost-driven deliberations.4 Previously published prediction models from the Atherosclerosis Risk in Communities (ARIC, C-index 0.79) and the Framingham Heart Study (FHS, C-index 0.86) performed reasonably well. Yet, these models were either not externally validated (ARIC) or performed considerably worse in validation (FHS).5 Biomarkers have previously been successfully used to predict incident HF.6, 7 Particularly, the STOP-HF trial showed promise in using a biomarker-based approach to predict HF.8, 9 The STOP-HF trial was a parallel-group randomized trial, assigning 1374 participants with cardiovascular risk factors to either care-as-usual or screening with B-type natriuretic peptide (BNP) testing. Individuals with a BNP >50 pg/mL underwent echocardiography and were referred to specialist cardiovascular care. Compared to the control group, the intervention group showed reduced rates of new-onset HF. This suggests that screening for HF is feasible. What is important here and also for other more successful clinical risk scores, is that the STOP-HF trial included individuals at risk for developing HF.3 Indeed, a previous study showed that the predictive value of biomarkers could be improved by a priori risk stratification.10 By making use of this risk heterogeneity, risk prediction of incident HF by biomarkers can potentially be improved. Unfortunately, most models and novel predictors have not brought us closer to improving risk prediction in clinical practice, but novel predictors of HF can teach us about the aetiology and pathophysiology of HF. In this light, the article of Delles et al.11 in the current issue of the Journal is of particular interest. Here, the authors use a metabolomics-based approach to identify potential metabolites, which are associated with HF using data from the PROspective Study of Pravastatin in the Elderly at Risk (PROSPER) cohort. This cohort includes 5341 elderly (between 70 to 82 years) individuals of whom 182 were hospitalized for HF during the 2.5 year follow-up. The diagnosis of HF was determined by having either signs or symptoms of fluid overload, or a chest radiograph displaying fluid overload or an echocardiography showing diminished cardiac function. This may have led to some misclassification, since co-morbidities such as chronic obstructive pulmonary disease and anaemia can mimic the signs and symptoms of HF. The authors use high-throughput serum metabolomics, a technique which measures low molecular weight compounds (metabolites), which are usually substrates or by-products of enzymatic reactions, and identified 80 metabolites for subsequent analyses. A particular strength of the study is that the results were validated in an independent cohort of 7330 individuals (aged 25 to 74 years) of whom 133 were diagnosed with HF during the 5-year follow-up period from the FINRISK 1997 cohort. Another strength of this study is the relatively unbiased approach to selecting metabolites and the ability of the authors to validate part of their findings in an independent cohort. Furthermore, the authors used a cohort readily at high risk for developing HF due to the presence of pre-existing vascular disease or a significant risk of developing this condition. In the present study, the authors have successfully identified 13 novel metabolites associated with HF hospitalizations. After rigorous correcting for clinical confounders, only phenylalanine and acetate were significantly associated with HF hospitalizations during follow-up. While the association between phenylalanine and HF is not novel, the results of the present study clearly extend on contemporary knowledge by showing that phenylalanine predicts the development of HF. While a definitive function of phenylalanine in HF has yet to be determined, examining the known physiological function of phenylalanine might provide us with hints. Phenylalanine is a precursor to catecholamines and has previously been found to be increased in HF patients compared to controls.12 It has been suggested that increased levels of catecholamines in HF patients are the result of a decreased cardiac output and stress response.12 The data from the current study suggest that phenylalanine might indeed be involved in a causative mechanism for HF, but more rigorous studies confirming this finding and unravel the potential mechanisms are needed. While the authors showed a significant association between phenylalanine and incident HF, the increment in risk prediction is negligible: the authors show that a proportion of the non-cases are correctly reclassified, yet phenylalanine did not provide for an incremental improvement of the c-statistic, which significantly reduces its possible utility in clinical practice. Despite the obvious limitations, the study of Delles et al. 11 does provide us with an interesting possible novel target in the pathophysiology of HF, which, as the authors rightly point out, deserves further mechanistic study. The question remains how can we improve risk prediction for incident HF and what are the next steps. The answer to this question might lie in the heterogeneity of HF. Here, risk predicting for incident HF can be improved by recognizing the heterogeneity of the HF syndrome. This heterogeneity is present in two forms, namely the risk heterogeneity of the population studied as well as the disease heterogeneity (Figure 1). Indeed, it is well established that the HF syndrome consists of distinct subtypes with exclusive or partially reciprocal aetiology, pathophysiology and risk factors, which share a common and final presentation of signs and symptoms. Hence, it is increasingly recognized that the treatment paradigm of HF does not adhere to a one-size-fits-all approach.13 Similarly, the same can be said for prediction of HF and treating its risk factors. To move beyond the current results of predicting individuals at risk for HF, the focus should shift from predicting incident HF as a whole to predicting distinct subtypes within HF (disease heterogeneity), particularly within populations already at risk for developing HF due to the presence of one or more specific risk factors such as diabetes mellitus and coronary artery disease (risk heterogeneity). Here, a more personalized approach is needed to predict incident HF. Indeed, previous studies have readily shown that risk factors significantly differ for predicting HF with a reduced (HFrEF) and preserved (HFpEF) ejection fraction. Subsequent studies extended upon this, by showing that risk models for predicting HFrEF and HFpEF separately had relatively higher C-indexes than contemporary risk models predicting HF as a whole. These results were then successfully validated in an independent cohort.14 In a clinical setting, we are readily moving away from a one-size-fits-all approach in our understanding of the pathophysiology of HF as well as in the treatment of HF by having separate treatment recommendations for HFrEF, HFpEF and HF with mid-range ejection fraction.2 Particularly within HFpEF, calls are being made for phenotypic specific treatment.15, 16 The question remains, why are we still predicting HF using a one-size-fits-all approach? While the results of Delles et al. are compelling, phenylalanine might show additional predictive value when restricted to predicting incident HF in particular subgroups (risk heterogeneity) or predicting specific subtypes of HF (disease heterogeneity). Furthermore, a better understanding of how phenylalanine relates to particular subtypes of HF might provide us with additional vital information about its pathophysiological role in the HF syndrome. Future clinical as well as -omics based approaches in predicting incident HF might benefit from recognizing this particular heterogeneity of the HF syndrome by identifying subtypes of HF using more personalized medicine approaches, in populations at high risk vs. low risk for developing HF. Conflict of interest: none declared.