Will computerised medical records and statistical tools really help us move towards precision medicine of heart failure with preserved ejection fraction?

Therapeutic strategy in heart failure with preserved ejection fraction (HFPEF) remains a challenging problem. Indeed, the prevalence of this disease is constantly increasing, especially in most highincome countries in connection with the ageing of the population, and is a cause of significant morbidity and mortality. Up to now, with the applied medicine, and despite the welcome approval of Sodiumglucose cotransporter type 2 inhibitors (iSGLT2) in this context, we still face incomplete control of the symptoms and the disease progression. 4 Whereas the management of HFPEF has been greatly modified in recent years by new therapeutic classes, largescale clinical trials in HFPEF have never been able to show a benefit on mortality until recent trials on SGLT2 inhibitors. This is due both to a pathophysiology that is still imperfectly mastered, but above all, to the significant heterogeneity within this syndrome itself, whether in terms of age, sex, comorbidities, and clinical and echocardiographic profiles. In this context, many machine learning clustering models have emerged, made possible by the improvement of computer systems, making it possible to analyse a very large number of data and variables in order to highlight subgroups, more homogeneous in terms of pathogenesis, prognosis and therefore potentially more likely to benefit from a specific treatment. 6–9 The work presented here falls within this perspective. The initial prospective observational study (PURSUITHFpEF) allowed, in patients who had acute decompensated heart failure with an left ventricular ejection fraction of >50% in 26 Japanese centres, isolation of 4 phenotypes from 16 clinical and biological variables. The methodological followup was therefore to focus, in the same database, on the prognosis of these subgroups according to the use or not of four major cardiovascular drugs which can be used for alternative indications in these same patients (atrial fibrillation, hypertension, chronic kidney disease, prior acute myocardial infarction (AMI) or peripheral artery disease). This methodology is interesting and opens up new opportunities. The description of the methods looks rigorous and applicable with an objective that is clinically relevant. The chosen population and the fairly homogeneous sizes of the clusters make it possible to observe a large number of clinical events and is very close to that which can be encountered in daily practice. Potential confounding biases related to the nature of the study were taken into account in the statistical analysis. 11 Regarding the results, this study succeeds in highlighting clear disparities concerning the prognosis under treatment, clearly demonstrating the impossibility of studying HFPEF as a single disease. Among the most convincing results, this work highlights a possible benefit of mineralocorticoid antagonists in the phenogroup of patients with the most cardiovascular risk factors and, in particular, severe arterial hypertension, with frequent repercussions on the left ventricle. Spironolactone is currently used in a large proportion of patients with HFPEF. Based on TOPCAT, AldoHF and HOMAGE, spironolactone improved cardiac structure and function of patients with HFPEF. 12 Despite the results of randomised trials and the related controversies. It appears that the mineralocorticoid antagonists should be considered with great attention. In the iSGLT2 trials about HFPEF, these treatments were prescribed for 34.5% of the population. The present study is hypothesis generating only, and there is no external validation of the model. Among the questions raised by the results of this study, it should be underscored that an observational study, even on a large scale and methodologically well conducted, does not validate the benefit of a therapy. We are moving forward for a more personalised approach of medicin, but still guidelines are based on randomised clinical trials. It is interesting to wonder about its population, whose average age is relatively high, and among patients who have all already been hospitalised for heart failure. The study of these therapies in younger patients, at an earlier stage in the evolution of the disease, before the occurrence of comorbidities or irreversible cardiovascular repercussions, could reinforce the findings of this article. It could also be interesting, in an era of global patient management, to reinforce the role of echocardiographical diagnosis or multimodal imaging in clustering to better differentiate HFPEF damage. 16 The addition to the databases of the level of left ventricular impact by studying the longitudinal strain, the level of atrial dilatation, the presence of mitral or tricuspid regurgitation could make it possible to refine the phenotypes and thus optimise the subsequent support. However, adding more variables can alter the values of the clusters. Machine learning is a very highquality tool for the care of our patients, as this study tends to demonstrate that. Neverless, we should make sure that the phenogroups described could be applied somewhere else and overtime. The demographic characteristics of the HFPEF population is changing according to the socioeconomical situations and the demographic characteristics of the populations. The main problem of the next few years on this subject is probably to succeed homogenising the phenotypes on a larger scale in order to be able to have access to comparable data from one centre to another in order to improve the care of our patients. The polypill strategies and piles that could be linked to the phenotype is a wish for the future. However, despite the value of the present manuscript, a lot remains to be done to improve the clustering on large databases and to prospectively validate new therapeutic options and/or associations of therapeutic agents.