Effect of hydrocortisone on mortality in patients with severe community-acquired pneumonia

Study design and oversight

The design of REMAP-CAP was reported previously [24, 25]. Patients eligible for the platform are assessed for eligibility to be randomized to interventions within one or more domains. We report here on patients randomized in the non-pandemic corticosteroid domain to fixed-duration hydrocortisone or control.

The trial is managed by a blinded International Trial Steering Committee (ITSC) and an unblinded independent DSMB. The trial had multiple international funders and sponsors. The funders had no role in study design, analysis, or reporting. The trial protocol was approved by relevant research ethics committees in each jurisdiction. Informed consent was obtained before randomization from all patients or their surrogates, or in a deferred fashion, in accordance with local legislation. The trial was conducted in accordance with Good Clinical Practice guidelines and the principles of the Declaration of Helsinki. All authors vouch for these data and analyses, as well as for the fidelity of this report to the trial protocol and statistical analysis plan. There are no confidentiality agreements that preclude the investigators publishing the findings.

Participants

Adult patients ≥ 18 years, who presented with CAP and were admitted within 48 h of hospital presentation to an intensive care unit (ICU) for respiratory or cardiovascular organ support, were eligible for enrollment in the platform. CAP was defined as signs and symptoms consistent with lower respiratory tract infection and radiological evidence of new-onset infiltrate of infective origin [26]. Respiratory organ support was defined as invasive or noninvasive mechanical ventilation or high flow nasal cannula if flow rate ≥ 30 L/min and FiO₂ ≥ 0.4. Cardiovascular organ support was defined as the intravenous infusion of any vasopressor or inotrope. Exclusion criteria included residents of a nursing home or long-term care facility, patients known to have been an inpatient in any healthcare facility within the last 30 days, presumption that death was imminent with lack of commitment to full support, and participation in REMAP-CAP in the prior 90 days. Additional exclusion criteria for the non-pandemic corticosteroid domain included a diagnosis of known or presumed COVID-19 infection, known hypersensitivity to corticosteroids, chronic systemic corticosteroid use, and a lapse of > 24 h since ICU admission. Further eligibility details are provided in the electronic supplementary material. The non-pandemic corticosteroid domain was recently extended to allow children and patients with moderate CAP (hospitalization without requirement for organ support). However, no children had been enrolled during the study period reported here and patients with moderate state CAP were ineligible for fixed-duration hydrocortisone.

Achieving a racially and ethnically diverse and sex-balanced sample was a goal of REMAP-CAP because of evidence of disparities in outcome and treatment effectiveness in pandemic and non-pandemic CAP [27, 28]. Participants (or their surrogates) self-reported their race/ethnicity and birth-assigned sex via fixed categories appropriate to their region.

Treatment allocation

The non-pandemic corticosteroid domain compares several interventions, including fixed-duration hydrocortisone, shock-dependent hydrocortisone (where hydrocortisone is only administered while the patient is receiving vasopressor therapy), fixed-duration dexamethasone, and control (no corticosteroid). Only the 7-day fixed-duration hydrocortisone arm met stopping criteria, and therefore this report only compares patients randomized to that arm or to control; other arms are ongoing and their results remain blinded. Sites participated in this domain if their investigators and clinicians were prepared, based on local equipoise, to randomize to ≥ 2 arms. If sites routinely gave corticosteroids for severe CAP, they could decline participation in the control arm, but still randomize to the different corticosteroid arms. This report therefore includes analysis of an additional subset of patients, those randomized within sites offering both fixed-dose hydrocortisone and control. Participants were randomized to each locally available arm, beginning with balanced assignment. Response-adaptive randomization was applied in a concealed fashion using allocation probabilities derived from each intervention's probability that it was most favorable, based on the accumulating evidence within the platform at each adaptive analysis [24].

Procedures

Because patients can be randomized to multiple interventions, REMAP-CAP has primarily used an open-label design, avoiding complex multiple placebo regimens. When a patient was randomized to one of the non-control arms, the clinical team was provided prescribing instructions for the intervention. All corticosteroids were supplied by each site's pharmacy. Other aspects of care were provided as per each site's standard of care. Data were collected on baseline characteristics, causative pathogens, corticosteroid use, and adverse events and outcomes by site investigators via an interactive web-based response technology with validation and logic checks. Although clinical staff were aware of their individual patient's treatment assignment, neither they nor the ITSC were provided any information about aggregate patient outcomes until a result was unblinded and released publicly.

Interventions

Participants randomized to the fixed-duration hydrocortisone arm were prescribed intravenous hydrocortisone bolus of 50 mg every 6 h for 7 days, without tapering. For patients discharged from hospital before the end of the 7-day course, hydrocortisone was discontinued at hospital discharge. In all patients, systemic corticosteroid therapy was permitted if a new clinical indication developed for which corticosteroids are an established treatment, such as post-extubation stridor, bronchospasm, or anaphylaxis. In addition to assignment to interventions in the non-pandemic corticosteroid domain, participants could be randomly assigned to other interventions within other therapeutic domains [24] (see the electronic supplementary material and www.remapcap.org).

Outcomes

The primary outcome was all-cause mortality, determined 90 days after randomization. Secondary outcomes were: ICU mortality, hospital length of stay (LOS), and ICU LOS, censored at 90 days; ventilator-free days (VFD), organ-support free days (OSFD), and proportion of intubated patients who receive a tracheostomy, censored at 28 days; progression to intubation and mechanical ventilation, extracorporeal membrane oxygenation (ECMO), or death, and readmission to the index ICU, each during the index hospitalization, censored at 90 days; destination at hospital discharge (home, rehabilitation hospital, nursing home or long-term care facility, or another acute hospitalization); and serious adverse events.

Statistical analysis

REMAP-CAP uses a Bayesian design with no maximum sample size. Planned adaptive analyses are performed and randomization continues until predefined statistical criteria are met for stopping a domain or intervention arm. The primary analysis of any intervention's effect is generated from an overarching Bayesian logistic model where the dependent variable is the primary outcome, 90-day mortality. The model is updated as the trial progresses, generating updated posterior probability distributions from the latest trial data and prior distributions, where the initial prior distributions for treatment effects were set as neutral. Full details of the overarching model and statistical analysis plan (SAP) governing the entire platform have been described previously [24]. In keeping with platform rules, following the DSMB's recommendation to cease enrollment into the fixed-duration hydrocortisone arm, blinded ITSC members wrote a specific SAP for this comparison of fixed-duration hydrocortisone to control. The plan was posted online (www.remapcap.org) before data lock and analysis (see the electronic supplementary material).

The primary estimate of the effect of hydrocortisone compared to control was generated from a model that compared the outcome of patients randomized to either hydrocortisone or control within the non-pandemic corticosteroid domain. The estimate was adjusted for other covariates that affect outcome: location (site, nested within country), age (categorized into 6 groups), sex, time period (3-month epochs), shock (yes/no), influenza (yes/no), and degree of respiratory compromise and support (not intubated, intubated, and intubated with severe hypoxemia [PEEP ≥ 5cmH₂0 and PaO₂/FiO₂ ratio < 200 mmHg]). As per the overarching design, to estimate the coefficients for these factors as robustly as possible, the primary model used information not only from patients randomized to hydrocortisone or control but rather from all randomized patients who met severe CAP criteria across all domains, regardless of whether patients were included in the corticosteroid domain or not, and inclusive of those patients whose CAP was due to COVID-19 and who were enrolled in domains open to both COVID-19 and non-COVID-19 patients. Thus, for example, any effect of ‘site’ could be estimated, and adjusted for, using information from all patients enrolled at that site, and not just those randomized to hydrocortisone or control. Although not all sites offered both control and hydrocortisone, the model's adjustment for site addressed potential differences in sites that offered one versus both arms.

Because patients were also being randomized (or not) within other domains and to other interventions, the model also accounted both for each patient’s random assignment to any other intervention in any other domain and for their eligibility for any domain or intervention (depending on site participation, baseline entry criteria, and patient or surrogate preference). Although the effect of hydrocortisone in patients with COVID was also evaluated in REMAP-CAP [18], the current estimate of the effect of hydrocortisone in non-pandemic CAP was estimated separately.

Because the primary model included information about assignment to interventions within domains whose evaluation is ongoing, it was run by the unblinded Statistical Analysis Committee (SAC) (electronic supplementary material), who conduct all protocol-specified trial update analyses and report those results to the DSMB. The model was run on all patients enrolled up to December 6th, 2023 for whom complete follow-up was ascertained by April 2nd, 2024. Vital status following randomization was determined through medical records, contact with the participant, their next of kin, or other health care professionals. Patients were analyzed according to the group to which they were assigned. Those missing the primary end point (n = 15) were ignored; there was no imputation of missing end point values. The primary analysis was expressed as an odds ratio (OR), where an OR < 1 implies benefit. Separate treatment effects of hydrocortisone compared to control were estimated in those with and without vasopressor-dependent shock at enrollment and those with and without confirmed or suspected influenza infection. The model used an additive factorial prior with a non-informative prior (mean 0, standard deviation 10) on the hydrocortisone main effect and informative priors centered on 0 (with standard deviation 0.15) for the additive log odds ratio interaction effects of hydrocortisone with shock and influenza. This prior structure assumes no knowledge of treatment effect at trial launch and allows separate estimates across influenza and shock strata by shrinking the posterior distributions for the hydrocortisone effects toward the overall effect. In pre-trial simulation, these priors performed well across scenarios where treatment effects were either homogenous or heterogenous across equal-sized subgroups.

The model was fit using a Markov chain Monte Carlo algorithm that drew iteratively (20,000 draws) from the joint posterior distribution, allowing calculation of ORs with their 95% credible intervals (CrI). The statistical triggers for an intervention were superiority (> 99% probability of OR < 1), futility (< 5% probability of > 20% reduction [OR < 1/1.2]), and inferiority (< 0.25% probability that an intervention was optimal in the domain).

Sensitivity analyses were conducted to generate estimates of the effect of hydrocortisone relative to control with full borrowing (pooled estimate across strata) and with no borrowing (independent estimates across strata). Additional sensitivity and secondary analyses were undertaken using only data from patients in the hydrocortisone and non-pandemic corticosteroid domain control groups. These analyses were conducted first on all patients randomized to either of these groups and second restricted to the subset of these patients who were enrolled at sites that offered both options (hydrocortisone or control; see SAP for detail in the electronic supplementary material). This last analysis most closely mimics a traditional 2-arm RCT. Wilcoxon Mann‐Whitney tests were used to compare the duration of cardiovascular support in the hydrocortisone and control groups. Data management and summaries were created using R version 4.1.2, the primary analysis was computed in R version 4.3.3 (2024-02-29), using the rstan package version 2.32.2.

Participant accrual, randomization, and baseline characteristics

Between March 11th, 2018, and December 6th, 2023, of 22,568 screened patients, 11,410 met platform enrollment criteria and were enrolled in REMAP-CAP. Of these, 3990 were eligible for inclusion in the primary model either, because they had non-pandemic CAP (n = 2606) or because they met COVID-19 criteria and were included in an intervention domain that included both pandemic and non-pandemic CAP (n = 1384). Of the 3990, 133 (3.3%) patients subsequently requested removal of all their data. Of the remaining 3857, by data lock, there were 3768 (97.7%) individuals for whom 90-day outcome had been ascertained, comprising the primary analysis cohort (Fig. 1).

Within the 2606 non-pandemic CAP cohort, 945 patients were enrolled in the corticosteroid domain, 29 of whom withdrew consent. Of the remaining 916, 658 patients were randomized to fixed-duration hydrocortisone (n = 536) or control (n = 122) at 101 sites in 18 countries. Of the 658 patients, 455 were randomized to hydrocortisone (n = 339) or control (n = 116) at the 70 sites that offered both options. Outcome was known for 100% of control patients, 97.2% (521/536) of patients assigned hydrocortisone, and 97.6% (331/339) of patients assigned hydrocortisone at sites where control was also available (Fig. 1). Two factors contribute to the high proportion of patients assigned to fixed-duration hydrocortisone. First, one third (n = 197, of whom outcome was known for 190 patients) of patients assigned to fixed-duration hydrocortisone were randomized at sites that offered > 1 corticosteroid intervention, but not control, due to perceived lack of equipoise. Second, the first adaptive analysis (conducted on 132 patients in the corticosteroid domain with known outcomes, at which point only control and fixed-duration hydrocortisone were available in the trial) used a dataset provided to the SAC with a transposition error from the data vendor for some sites, such that 13 and 15 patients in the hydrocortisone and control groups were incorrectly labeled as belonging to the opposite group. The adaptive analysis based on these erroneously labeled data led to updated response-adaptive randomization proportions that assigned more patients to fixed-duration hydrocortisone (electronic supplementary material).

The subsequent adaptive analysis used correctly-labeled data and triggered the futility threshold, halting further enrollment. The initial data error was only detected on February 23rd, 2024, after enrollment was closed and data were unblinded, at which point it was determined that the error was an isolated incident concerning only one data transfer to the SAC. Incorrect labeling only affected the corticosteroid domain, no other domains were impacted. All safety data provided to the DSMB throughout were correctly labelled. Nonetheless, the investigators filed an immediate incident report with relevant jurisdictions and implemented a corrective action plan (electronic supplementary material).

Baseline characteristics among both the entire group randomized to hydrocortisone or control (n = 658) and the subset randomized at sites where both options were available (n = 455) are presented in Table 1. Though the groups were broadly similar, patients randomized to hydrocortisone were slightly older and had slightly worse baseline respiratory dysfunction (higher mechanical ventilation rate and lower PaO₂/FiO₂ ratio) and cardiovascular dysfunction (slightly higher cardiovascular SOFA score). Although ethnicity was balanced, Asian sites (which only joined after the first adaptive analysis) contributed no patients to the control arm. Causative pathogens were evenly distributed among groups (Table 1 and electronic supplementary material).

Intervention fidelity

At the time of data lock, we had not been able to verify corticosteroid dosing for 3 patients, such that we had dosing information on 655 (99.5%) of the 658 participants randomized in the corticosteroid domain to fixed-duration hydrocortisone (n = 533) or control (n = 122). Among those assigned to fixed-duration hydrocortisone, 98.9% (527/533) received at least one dose of hydrocortisone, an additional 0.4% (2/533) received an alternative systemic corticosteroid, and only 0.8% (4/533) received no corticosteroid. The first dose of hydrocortisone was given before midnight of the first study day in 94.2% (502/533) of patients and the median (IQR) duration of hydrocortisone therapy was 7 (4–8) days. Among those assigned to control, 23% (28/122) received a systemic corticosteroid (16 of whom received hydrocortisone). For those receiving a corticosteroid, the median (IQR) duration of treatment was 4 (2–5.3) days. Only one patient (0.8%) among those assigned to the control arm received a full dose (> 6 days of 200 mg hydrocortisone-equivalent per day) of corticosteroids.

Primary outcome

By day 90, 78 (15%) of 521 patients assigned to hydrocortisone and 12 (9.8%) of 122 patients assigned to control had died (Fig. 2). The median adjusted OR for the effect of hydrocortisone, relative to control, on 90-day mortality, estimated from the primary model (which generates OR for each stratum with fixed borrowing) was 1.52 (95% CrI, 0.69 to 3.56) for vasopressor-dependent shock due to influenza; 1.53 (0.76 to 3.29) for vasopressor-dependent shock not due to influenza; 1.62 (0.73 to 3.84) for influenza without vasopressor-dependent shock, and; 1.63 (0.80 to 3.59) for severe CAP with neither influenza nor vasopressor-dependent shock. These OR yielded posterior probabilities of > 20% reduction in the odds of death at 90 days of 7.1%, 4.4%, 5.3%, and 3.3%, respectively. Probabilities of superiority (OR < 1) were 15.7%, 12.1%, 12.6%, and 9.2%, respectively, with corresponding probabilities of harm (OR > 1) 84.3%, 87.9%, 87.4%, and 90.8%, respectively (Table 2).

Pre-specified sensitivity and secondary analyses

Repeating the primary analysis, but generating a single treatment effect for hydrocortisone (as opposed to separate effects for each stratum), the overall median adjusted OR for death at day 90 with hydrocortisone was 1.56 (95% CrI, 0.80–3.31), yielding a 10% probability of superiority, a 3.4% probability of improving the odds of death at day 90 by > 20%, and a 90% probability of harm. Sensitivity analyses exploring the effect of less borrowing across strata generated considerably less certain probabilities of superiority or futility for the two influenza strata. However, the low probabilities of superiority and high probabilities of futility overall and in the non-influenza strata persisted, regardless of whether the models included borrowing across strata (see Table 2).

When restricting analyses to data from patients enrolled at sites where they could be randomized to either the fixed-dose hydrocortisone or the control arm, the median adjusted OR for death at day 90 with hydrocortisone was 1.55 (95% CrI, 0.75–3.63), yielding a 14.4% probability of superiority, a 4.4% probability of improving the odds of death at day 90 by > 20%, and a 85.6% probability of harm (electronic supplementary material).

Secondary outcomes

Secondary outcomes are reported in Table 3. ICU and hospital length of stay were similar in both arms, as were the proportions of intubated patients who received a tracheostomy; the rates of progression to intubation, mechanical ventilation, ECMO, or death; the readmission rates to the index ICU, and; the distributions of hospital discharge destinations. OSFDs are a composite of death and duration of support. Although the observed mortality rates were higher in the hydrocortisone arm, the median OSFD adjusted OR was 0.82 (0.56–1.22), yielding an 85.4% probability that hydrocortisone was superior to control. Restricted to sites where patients could be randomized to either arm, the median adjusted OR was 0.83 (95% CrI, 0.53–1.13), yielding a 91.3% probability that hydrocortisone was superior to control. This possible improvement in OSFDs appeared to be due to a shorter duration of cardiovascular support. The duration of cardiovascular support was median (IQR) 2(2–5) days in the hydrocortisone arm compared to control 3(2–6.75) days, p value = 0.05.

Serious adverse events were reported in seven patients (1.3%) randomized to the hydrocortisone arm and one patient (0.8%) randomized to the control arm. Because of the limited number of events, there was no formal statistical comparison (Table 3).

Limitations

This study has important limitations. First, the futility rule that stopped the study may have been premature. It only ruled out a relatively large effect (> 20% relative reduction in mortality), and the sample size was too small to explore subgroup effects. In particular, the limited number of patients in the influenza strata preclude any conclusions about the effect of steroids in these patients. Second, the open-label design meant clinicians were aware of the treatment assignment, which likely influenced their decision to prescribe corticosteroids in some control arm patients, although typically for a much shorter course. If steroids were beneficial, their use in the control arm would have reduced the effect size, but would be unlikely to cause futility and possible harm. Third, a miscoding error prior to the first adaptive analysis led to erroneous updated response-adaptive randomization proportions, which led more patients to be assigned to the fixed-duration hydrocortisone intervention. Though the miscoding may have caused fewer patients to be allocated to receive the optimal treatment, it is unlikely to have changed the overall outcome, as evidenced by the consistent findings across the sensitivity analyses. Fourth, one of the exclusion criteria, chronic corticosteroid therapy was not defined prospectively but by the physician in charge of the patient.

Conflict of interest

See submitted ICMJE forms for declared potential conflict of interests.

Ethical approval

All procedures performed in this study involving human participants were in accordance with the ethical standards of the institutional and/or national research committees and with the 1964 Helsinki Declaration and its later amendments.

Consent to participate

Informed consent was obtained before randomization from all patients or their surrogates, or in a deferred fashion, in accordance with local legislation.