The Problem with Hospitalization Endpoints in Heart Failure Trials

Most early trials for heart failure (HF) assessed the effect of therapies on haemodynamic parameters, functional capacity and symptoms. Methodological limitations of these trials, as well the discovery that improvements in symptoms do not always correlate with better outcomes,1 resulted in the inclusion of cardiovascular safety as a mandated trial endpoint. The near parallel discovery that neurohormonal blockade could improve survival led to a paradigm in which new treatments received regulatory approval on this basis. As prognosis improved with successive therapies, event rates in clinical trials declined, meaning ever-larger sample sizes were required (Figure 1). Further confounding the situation, participant profiles have shifted from a population of younger people with a single condition, to one which is older and increasingly multimorbid. With this has come an increasing contribution of non-cardiovascular events, which dilute the treatment effect of HF specific therapies.2 So, what's the problem? Whilst hospitalizations (and their causes) are more difficult to assess than total mortality, preventing both death and hospitalization are important to patients. Including hospitalization events within composite endpoints is, therefore, a legitimate strategy to provide a broader assessment of treatment effects, whilst also having the advantage of increasing statistical power. Nevertheless, in time-to-first or first-and-recurrent event analyses, hospitalization events are regarded as being equally important to death although intuitively, this makes no sense. Whilst balancing the relative importance of mortality and hospitalization endpoints is challenging, this conflict is limited if the direction of the effect is concordant. Discordant endpoints complicate trial interpretation. In FINEARTS-HF (Finerenone Trial to Investigate Efficacy and Safety Superior to Placebo in Patients with Heart Failure), reductions in worsening HF events were not corroborated by reductions in cardiovascular death or total mortality.3 Whilst this trial is widely viewed as ‘positive’ and potentially practice changing, comparable reductions in HF hospitalizations in DIG (Digitalis Investigation Group) have been largely forgotten, perhaps in part due to the choice of all-cause mortality as this trial's primary endpoint.4 Hierarchical clinical composite endpoints attempt to take account of how mortality, hospitalization and patient-reported outcomes differ in their relative importance, whilst also increasing statistical power and including more information relevant to patients.5 However, subjective endpoints may dominate where event rates are low. In trials assessing tricuspid transcatheter edge-to-edge repair for right HF adverse clinical events were uncommon, and so the positive results were driven by unblinded quality-of-life assessments, which are challenging to assess since most scoring systems are influenced by comorbidities in this population.6, 7 The conflation of surrogate markers with mortality and hospitalization endpoints within hierarchical designs is another concern. For example, lower than anticipated event rates in DAPA-MI (Dapagliflozin in Patients with Myocardial Infarction) resulted in the amendment of the primary endpoint to a hierarchical clinical composite including ‘cardiometabolic’ endpoints including the new onset of diabetes or modest reductions in body weight.8 Many contemporary ‘positive’ trials report reductions in HF hospitalizations whilst the overall burden of hospitalization is unchanged. This has become increasingly common for trials conducted in populations with lower event rates, or competing risks of non-cardiovascular morbidity and mortality, such as those for HF with preserved ejection fraction. In EMPEROR-Preserved (Empagliflozin Trial in Patients with Chronic Heart Failure with Preserved Ejection Fraction), empagliflozin reduced HF hospitalizations, however 33 participants needed to be treated for 2 years to prevent one admission, and only 18% of total hospitalizations were due to HF.9 The net benefit to patients and healthcare systems of shifting the cause without reducing overall admissions remains unclear, and patients likely care little whether they are hospitalized for HF or for another reason. Hospitalizations are often complex, with overlapping causes and a spectrum of presentations ranging from acute pulmonary oedema to more gradual volume overload. Yet, trials require binary categorization: was the admission for HF, or not? The approach most widely adopted is for investigator-reported hospitalization episodes to be adjudicated by committees who impose strict criteria, usually requiring symptoms plus diuretic intensification, which may exclude cases where HF contributed indirectly. These criteria are not standardized and may differ between trials, complicating trial interpretation. Additionally, in the process of adjudication much information is lost and hospitalizations which include worsening HF but are precipitated by another cause (such as infection) may be simply ignored. The notion that we should differentiate between spontaneous and precipitated decompensation episodes may be counter-productive, since both are independently associated with increased mortality risk.10 Where the effect size of an intervention is large, the exact approach to endpoint accumulation, whether investigator-reported or centrally adjudicated may matter little.11 However, for many trials ignoring potentially relevant events may be detrimental to interpretation. One possible solution might be to approach adjudication in a non-binary fashion and assign a probability as to whether episode is related to the outcome of interest—in other words, instead of treating each event as counting or not, it incorporates the uncertainty of adjudication directly into the analysis. This approach increases statistical power and may affect the interpretation of trials with statistically fragile results.12 Another solution is to use routinely collected electronic healthcare data as an unbiased approach to endpoint collection. Whilst such routine data might risk misclassifying some events, this does not necessarily result in bias of the estimated treatment effect, so long as inaccuracies are randomly distributed across treatment arms—which should be the case in trials which are double-blinded. Adjudication is more problematic in open-label trials, where participants and investigators are aware of treatment allocation. If endpoint collection is biased, even a small increase in the number of false events in the control arm can significantly overestimate the true treatment effect. The logical solution to these limitations is to have adjudication committees blinded to treatment allocation. The assumption is that the committee's lack of direct knowledge of the participant, and the imposition of strict adjudication criteria improves the validity of these decisions. However, committees adjudicate using source data, including medical documentation and discharge letters which have been compiled by unblinded physicians. Therefore, in open-label trials even blinded adjudication does not mitigate against the risk of ascertainment bias, making less bias prone endpoints preferable. Unlike death, the decision to hospitalize a patient is a subjective judgement made by the treating physician. Although for some patients the need for hospitalization might be unequivocal, there may be differences in the reasons for and thresholds to hospitalize a patient between different physicians and healthcare systems. The knowledge of a participant's' treatment allocation has the potential to influence the decision to admit a patient to hospital or manage them in the community, whether consciously or not. This is a particular concern in unblinded trials of medical devices, where physicians may have lower thresholds to admit untreated patients. In RESHAPE-HF2 (A Randomized Study of the MitraClip Device in Heart Failure Patients With Clinically Significant Functional Mitral Regurgitation), reductions in hospitalizations and improvements in quality of life were not matched by changes in mortality or objective markers.13 One approach to mitigate against the subjective nature of hospitalization decisions might be to include other clinically meaningful events within an endpoint designed to capture worsening HF, such as urgent outpatient visits, or the intensification of diuretic therapy. This approach has been increasingly utilized in contemporary trials and provides additional nuance beyond binary hospitalization episodes.3 However, the potential downsides of this approach are that these events are even more prone to variability in practice patterns, are less prognostically relevant, and may increase the adjudication burden. All of these are factors could dilute the clinical interpretability of trial results compared to an approach which uses more robust outcomes. Bias is not limited to physicians, as patients aware of their assignment may alter their reporting and behaviour. In medical device trials, those randomized to the control arm are effectively ‘denied’ an intervention which many would consider as standard care. In EARLY TAVR (Evaluation of TAVR Compared to Surveillance for Patients with Asymptomatic Severe Aortic Stenosis), patients with objectively asymptomatic aortic stenosis randomized to usual care showed early and unexpected increases in hospitalizations,14 likely due to increased vigilance about their untreated disease, so called ‘subtraction anxiety’.15 Drug trials are less prone to this form of bias, but where the effects are obvious, blinding can still be compromised. Tirzepatide, for instance, causes rapid weight loss and gastrointestinal side effect, potentially unblinding participants making subjective endpoints more prone to bias.16 These are significant challenges which require physicians to look beyond the topline ‘positive’ results of mostly industry-sponsored trials. However, a better methodological approach is critical to avoid adopting costly new therapies that deliver only marginal or questionable benefit. S.S. is supported by a NIHR Clinical Lectureship. Conflict of interest: S.S. reports speaker's fees and honoraria from AstraZeneca. N.M. reports speaker's fees and honoraria from AstraZeneca, Novartis, BMS, Amgen, Cardiac Dimensions, Boehringer Ingelheim, Bayer and Novo Nordisk. K.K.W. reports speaker's fees and honoraria from Medtronic, Abbott, Cardiac Dimensions, Novartis, BMS, Pfizer, Bayer, and unconditional research grants from Medtronic. All other authors have nothing to disclose.