Methods

Individual participant-level data from seven trials of HFrEF (CHARM [Candesartan in Heart Failure Assessment of Reduction in Mortality and Morbidity]-Added, CHARM-Alternative, EMPHASIS-HF [Eplerenone in Mild Patients Hospitalization and Survival Study in Heart Failure], ATMOSPHERE [Aliskiren Trial to Minimize Outcomes in Patients with Heart Failure], PARADIGM-HF [Prospective Comparison of ARNI With ACEI to Determine Impact on Global Mortality and Morbidity in Heart Failure], DAPA-HF [Dapagliflozin and Prevention of Adverse Outcomes in Heart Failure], and GALACTIC-HF [Global Approach to Lowering Adverse Cardiac Outcomes Through Improving Contractility in Heart Failure]) and five trials of HFpEF (CHARM-Preserved, I-PRESERVE [Irbesartan in Heart Failure with Preserved Systolic Function], TOPCAT [Treatment of Preserved Cardiac Function Heart Failure With an Aldosterone Antagonist] Americas, PARAGON-HF [Prospective Comparison of Angiotensin Receptor Neprilysin Inhibitor With Angiotensin Receptor Blockers Global Outcomes in HF With Preserved Ejection Fraction], and FINEARTS-HF [Finerenone Trial to Investigate Efficacy and Safety Superior to Placebo in Patients with Heart Failure]) were included in the present study.13–24 The design, baseline characteristics, and results of these trials have been published in detail previously.13–38 Given the considerable regional heterogeneity in the conduct of the TOPCAT trial, we restricted the analysis to the TOPCAT America in this study.39 A summary of the designs is provided in Supplementary data online, Table S1. Each trial was approved by local ethics committees, and written informed consent was obtained from each patient. Serum rather than plasma potassium concentration was measured in all trials (serum potassium levels are slightly higher by approximately 0.1 to 0.4 mmol/L than plasma potassium levels).40 For each included study, patients from both the intervention and control arms were included in the present analysis. Exclusion criteria were a lack of baseline serum potassium level, potassium level >8.0 mmol/L considering the possibility of haemolysis, LVEF >40% or unknown in HFrEF trials, and LVEF <50% or unknown in HFpEF trials (see Supplementary data online, Figure S1).

Trial outcomes

The primary outcome of each trial is summarized in Supplementary data online, Table S1. The primary endpoints of each trial were as follows: In CHARM, EMPHASIS-HF, ATMOSPHERE, PARADIGM-HF, and PARAGON-HF, the primary endpoint was a composite of cardiovascular death or HF hospitalization (total events in PARAGON-HF and time-to-first event in the others). In TOPCAT, the primary endpoint was the same composite plus resuscitated sudden death (time-to-first). The primary composite outcome in I-PRESERVE was all-cause death or cardiovascular hospitalization (time-to-first). In DAPA-HF, the primary endpoint was a composite of worsening HF (hospitalization or urgent visit resulting in intravenous therapy) or cardiovascular death. The primary composite outcome in GALACTIC-HF was the composite of HF events or cardiovascular death (time-to-first). In FINEARTS-HF, the primary outcome was total worsening HF events and death from cardiovascular causes. All-cause death was a secondary endpoint in all trials except ATMOSPHERE, in which it was an exploratory endpoint. In the present study, we used all-cause death, the most objective and arguably the most important outcome, as the primary endpoint. We also separately analysed non-cardiovascular death and cardiovascular death, including the two major types of cardiovascular death—sudden death and pump failure death—as well as first HF hospitalization, the composite of first HF hospitalization or all-cause death, and the composite of first HF hospitalization or cardiovascular death. Each event and cause of death was adjudicated by a central endpoint committee according to the prespecified criteria in each trial.

Statistical analysis

The numbers of missing data in each trial are shown in Supplementary data online, Table S2. We divided the patients into two categories: HFrEF (LVEF ≤40%) and HFpEF (LVEF ≥50%). The following analyses were performed separately for patients with HFrEF and those with HFpEF. Based on serum potassium measurements at randomization in each study (only EMPHASIS-HF used at visit 1), patients were classified into the following potassium categories: <3.5 mmol/L, ≥3.5–<4.0 mmol/L, ≥4.0–<4.5 mmol/L, ≥4.5–<5.0 mmol/L, ≥5.0–<5.5 mmol/L, and ≥5.5 mmol/L. Baseline patient characteristics were reported as median (interquartile range) or mean ± standard deviation for continuous variables and as counts and percentages for categorical variables. According to potassium categories, continuous variables were compared by the Jonckheere–Terpstra trend test, and binary variables were compared by the Cochran–Armitage trend test.

The occurrence of each endpoint according to the potassium level was compared with Kaplan–Meier curves and the log-rank test. The incidence rates for each outcome were estimated per 100 patient-years. The associations between the potassium level, as a continuous variable, and the hazard ratios (HRs) of each major clinical outcome were modelled using restricted cubic splines with median potassium level as reference. The adjusted variables are listed below. The three knots (at 10th, 50th, and 90th percentiles) were placed at default positions according to the percentile of potassium level. In addition, the incidence rate for each individual and composite time-to-first outcome was evaluated across the spectrum of potassium level using a Poisson regression model, in which potassium level was examined using restricted cubic splines employing 4 knots. P for interaction was assessed using a likelihood ratio test, comparing models with and without the relevant interaction terms for potassium with baseline estimated glomerular filtration rate (eGFR) category (<60 vs ≥60 mL/min/1.73 m2) randomized treatment assignment, and use of mineralocorticoid receptor antagonists (MRAs). To evaluate the lowest risk of clinical outcomes, three approaches were employed: binary recursive partitioning,41 identification of the nadir of the HR, and the nadir of the incidence rate. Binary recursive partitioning was used to determine thresholds of clinical outcome risk, without prespecifying cut-off values, resulting in risk strata where the lowest relative hazard was observed. The nadir of HR was defined as the potassium level at which the spline-predicted adjusted HR was lowest, identified by calculating the minimum among the model-based predicted values. The nadir based on the spline-predicted incidence rate was also determined using the same approach to HR.

For time-to-first-event analysis, Cox proportional hazards models were used to compute HRs and 95% confidence interval (CI) according to potassium categories. In the Cox proportional hazard models, unadjusted HRs were stratified according to the trial and with the treatment assignment as a covariate. Adjusted HRs (aHRs) were further adjusted for age, sex, region, race, heart rate, New York Heart Association (NYHA) functional classes III or IV, systolic blood pressure, body mass index, LVEF, serum creatinine, haemoglobin, N-terminal pro-B-type natriuretic peptide (NT-proBNP, log-transformed), history of atrial fibrillation, history of myocardial infarction, history of diabetes, diuretic use, MRA use, and angiotensin-converting enzyme inhibitor/angiotensin receptor blocker/angiotensin receptor–neprilysin inhibitor use. Furthermore, additional analysis was made for the adjustment of the above variables and dose of oral loop diuretics, limited to more recent trials in which loop diuretic dose information was reported in a sufficiently detailed and standardized manner, including GALACTIC-HF, DAPA-HF, and PARADIGM-HF for HFrEF trials, and FINEARTS-HF and PARAGON-HF for HFpEF trials (see Supplementary data online, Table S3). For the dose of oral loop diuretics, total daily furosemide dose equivalents were calculated from medication records in each of the above trials. Bumetanide 1 mg, torsemide 20 mg, azosemide 60 mg, and ethacrynic acid 100 mg were considered equivalent to 80 mg of oral furosemide. In this calculation using the record, we only focused on: (1) oral diuretic use, (2) dosage units expressed in milligrams, (3) interpretable frequency of administration, and (4) exclusion of cases with total daily oral diuretic doses exceeding 1000 mg. In the cases with missing data, the missing indicator method was used for the NT-proBNP value and the dose of oral loop diuretics. In a sensitivity analysis, we conducted a two-stage meta-analysis in a random effects model for patients with HFrEF. We selected the potassium range of 4.0 to <4.5 mmol/L as the reference group, as it lies within both commonly used definitions of the normal range (3.5–5.0 and 4.0–5.0 mmol/L) in the previous studies and represents the midpoint or overlapping region of these prior categorizations. All statistical analyses were conducted using STATA version 18 (StataCorp LLC; College Station, TX, USA), and a P-value of < .05 was considered statistically significant.