IL-6 Receptor Antagonists in REMAP-CAP

Why Timing and Outcome Definition Matter

For critically ill patients with COVID-19, the question was not simply whether an IL-6 receptor antagonist could improve an outcome. It was whether starting treatment soon after intensive-care organ support began could improve the course of severe illness. The REMAP-CAP immune-modulation comparison addressed that question for tocilizumab and sarilumab against control. Interpreting its findings requires attention to both the early-treatment setting and the trial’s unusual primary outcome.

The trial was conducted within a platform that could evaluate multiple treatment comparisons. For this case, the important comparison is the randomized allocation to an IL-6 receptor antagonist or control among critically ill patients eligible for the immune-modulation domain. The control group represents the care available in the trial at that time, not an absence of all treatment. Usual care and other co-interventions could be given; glucocorticoids, for example, were commonly used.

Definition: In the REMAP-CAP primary outcome, a patient who died in hospital before day 90 was assigned a score of \(-1\). Otherwise, the score was the number of days alive and free of respiratory or cardiovascular organ support through day 21. Larger values represent a more favorable outcome. The death rule takes precedence even if a patient had been free of organ support for some days before dying.

This endpoint combines survival and duration of organ support into an ordered outcome. It is not a simple count of days alive: a patient who died in hospital before day 90 receives \(-1\), whereas a patient not meeting that death rule receives a value from 0 to 21. Values can therefore be tied, and a one-unit difference is not necessarily an equally important clinical change at every point on the scale.

The Trial Question and Early-Treatment Setting

REMAP-CAP enrolled critically ill patients receiving respiratory or cardiovascular organ support in an intensive care unit. Patients were randomized within 24 hours after organ support began. The treatment contrast therefore concerns patients treated early in their course of critical illness; it does not directly answer whether starting an antagonist several days later would have the same effect.

The trial compared tocilizumab and sarilumab with control. The published primary analysis used a Bayesian ordinal model for the organ support-free-days outcome, adjusting for prespecified factors and accounting for the platform structure. An adjusted cumulative odds ratio greater than 1 favored treatment: under the proportional-odds interpretation, it indicated greater odds of a more favorable outcome across the outcome thresholds. It was not a ratio of mean days, nor a ratio of mortality risks.

Worked Example: Scoring Patients on the Primary Outcome

Consider two entirely hypothetical patients. Patient A stops organ support at the end of day 6 and remains alive without organ support through day 21 and alive through day 90. Patient B is discharged from organ support on day 10 but dies in hospital on day 46.

For Patient A, the support-free days are days 7 through 21, inclusive. The count is

21 - 7 + 1 = 15 \text{ days}.

Patient A’s score is therefore 15. Patient B receives \(-1\), because death in hospital before day 90 determines the score. Counting Patient B’s support-free days before death instead would apply the endpoint incorrectly.

The timing restriction is part of the interpretation, not a minor enrollment detail. A patient who has just started organ support is at a different point in the illness trajectory from one who has required it for a prolonged period. Randomization supports a causal comparison for patients represented by the eligible trial population, but it does not by itself justify extrapolation to a different treatment window.

What REMAP-CAP Found

In the published primary analysis, both IL-6 receptor antagonists met the trial’s prespecified criterion for superiority to control on the organ support-free-days outcome. The adjusted cumulative odds ratio was 1.64 for tocilizumab, with a 95% Bayesian credible interval from 1.25 to 2.14. For sarilumab, the corresponding estimate was 1.76, with a 95% credible interval from 1.17 to 2.91. The posterior probabilities of superiority were reported as 99.9% for tocilizumab and 99.5% for sarilumab.

These estimates support a favorable effect of each treatment relative to control in the trial population and setting. The sarilumab interval is wider, reflecting less information for that comparison; its point estimate being numerically larger than tocilizumab’s is not evidence that sarilumab is more effective. Comparing two point estimates is not the same as estimating a randomized head-to-head contrast, and the trial was not designed to establish that one antagonist was superior to the other.

Interpretation: A 95% Bayesian credible interval gives a 95% posterior probability that the modeled treatment-effect parameter lies within the interval, conditional on the analysis model, prior distribution, and observed data. It does not mean that 95% of patients benefit, and it is not a probability that the null hypothesis is true.

The reported odds ratios are cumulative and adjusted. To see what “cumulative” means, imagine choosing a threshold on the ordered outcome, such as being above a particular number of support-free days. The model compares the odds of being above that threshold rather than at or below it. A proportional-odds model uses a common treatment odds ratio across these thresholds, while allowing the baseline odds to vary by threshold. That common-odds structure is an assumption of the model, not a claim that every patient benefits by the same number of days.

\text{Cumulative odds at threshold } c = \frac{P(Y>c)}{P(Y\leq c)}, \qquad OR = \frac{\text{treatment cumulative odds}}{\text{control cumulative odds}}.

Here \(Y\) denotes the ordered outcome score and \(c\) a threshold. With higher scores favorable, an odds ratio above 1 favors treatment. The model’s estimate combines information across the ordered categories; it does not give the probability of a particular patient recovering or a direct increase in average support-free days.

Reading the Odds Ratio Carefully

Worked Example: Interpreting an Adjusted Cumulative Odds Ratio

Use the published tocilizumab estimate, \(OR=1.64\). Its direct interpretation is that the adjusted cumulative odds of a more favorable outcome were estimated to be 1.64 times as high with tocilizumab as with control, under the model’s common-odds assumption. This is a 64% increase in odds, because

(1.64-1)\times 100\%=64\%.

It is not a 64-percentage-point increase in the chance of a favorable outcome. For illustration only, suppose the control cumulative odds at one particular threshold were 0.50. Applying an odds ratio of 1.64 would give treatment odds \(1.64\times0.50=0.82\). The corresponding probabilities would be

P_{\text{control}}=\frac{0.50}{1+0.50}=0.333, \qquad P_{\text{treatment}}=\frac{0.82}{1+0.82}=0.451.

Those hypothetical probabilities depend on the assumed control odds and refer to one threshold. They are not the observed trial probabilities and should not be presented as a translation of the adjusted trial estimate into a single overall recovery probability.

A useful justification for the interpretation follows from the odds definition. If the treatment and control cumulative odds at threshold \(c\) are \(O_T(c)\) and \(O_C(c)\), respectively, then \(OR(c)=O_T(c)/O_C(c)\). When a proportional-odds model assumes \(OR(c)=\theta\) for every threshold, \(\theta>1\) means higher odds of exceeding each threshold under treatment. This explains why the direction is favorable when larger endpoint values are better. It does not imply that the treatment effect is constant on the probability scale.

O_T(c)=\theta O_C(c), \qquad P_T(Y>c)=\frac{\theta O_C(c)}{1+\theta O_C(c)}.

Because the probability expression depends on the control odds \(O_C(c)\), the same odds ratio can correspond to different probability differences at different thresholds. The derivation also shows why an odds ratio is not a risk ratio and cannot be read as a fixed percentage-point benefit.

Timing, Scope, and Diagnostic Checks

Worked Example: Checking the Treatment Window

Suppose, hypothetically, that one critically ill patient starts organ support at 08:00 on Monday and is randomized at 18:00 the same day. The elapsed time is

18{:}00-08{:}00=10 \text{ hours}.

That patient is within a 24-hour window after organ support began. If another otherwise similar patient is randomized 30 hours after support began, then \(30-24=6\) hours beyond the window; that patient is outside the stated early-treatment eligibility window. This classification is about applicability of the trial comparison, not a prediction that treatment would or would not work for the second patient.

When reading the result, check the following before making a clinical or statistical claim:

Outcome coding: confirm that in-hospital deaths before day 90 receive \(-1\), rather than counting only deaths before day 21 as deaths on the endpoint.
Direction: higher scores are more favorable, so an odds ratio above 1 favors the antagonist.
Effect scale: the estimate is a cumulative odds ratio, not a mortality risk ratio, mean difference, or absolute risk reduction.
Population and timing: the evidence concerns critically ill, organ-supported patients treated within the trial’s early window.
Uncertainty and comparisons: examine the credible interval and avoid ranking drugs merely by their point estimates.

The model’s interpretation depends on its assumptions, including the adequacy of the ordinal outcome model, the proportional-odds structure, covariate adjustment, and the Bayesian prior specification. Randomization supports the treatment comparison within the trial, but missing outcome information, departures from the model, or important changes in background care can affect how well the estimate transports to another setting. The platform’s shared control can improve efficiency, but comparisons must still respect which patients and periods contributed to each randomized contrast.

The principal alternatives answer different questions. A binary endpoint such as mortality is easier to interpret as an event probability but discards information about time free of organ support. A mean difference in support-free days is direct in days but can be awkward for a distribution with many deaths assigned the worst score and substantial ties. The ordinal approach retains ordering across the outcome while relying on a common-odds assumption. A treatment estimate from observational data, meanwhile, would not have the same protection against confounding supplied by randomization.

Common Misreadings and Limitations

Worked Example: A Crude Odds Ratio Is Not the Trial Estimate

For a separate hypothetical two-category example, suppose 40 of 100 treated patients and 30 of 100 control patients have a favorable outcome. The treatment odds are \(40/60=0.667\), and the control odds are \(30/70=0.429\). The crude odds ratio is

OR_{\text{crude}} = \frac{40/60}{30/70} = \frac{40\times70}{60\times30} = 1.556.

This arithmetic describes a single binary outcome in these hypothetical counts. It is not the REMAP-CAP adjusted cumulative odds ratio: that estimate used the ordered organ support-free-days outcome, adjustment, and the trial’s Bayesian model. A superficially similar number does not make two analyses equivalent.

It is also incorrect to infer that the treatment effect increases with each day of delay, or that the study established an exact biological cutoff at 24 hours. The enrollment window defines the population and timing studied. Demonstrating that treatment effects differ by time since organ support began would require an appropriate treatment-by-timing analysis with sufficient information; the early-window result alone cannot establish such an interaction.

Similarly, success of both antagonists against control does not prove equal efficacy, and the numerically higher sarilumab estimate does not prove superiority over tocilizumab. The two comparisons have different uncertainty, and the relevant head-to-head contrast would need its own estimate and uncertainty. Finally, the primary outcome is a composite ordinal summary: its favorable shift is clinically important evidence, but it does not mean that every patient survived or avoided organ support.

Key takeaway: REMAP-CAP found that tocilizumab and sarilumab improved the distribution of organ support-free-days outcomes relative to control among critically ill patients treated soon after organ support began. Interpret the estimates as adjusted cumulative odds ratios, preserve the day-90 death rule, and do not extend the result automatically to delayed treatment or to a head-to-head ranking of the drugs.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A patient dies in hospital on day 70 after spending days 12 through 21 alive and free of organ support. What is the primary-outcome score, and why?
In this outcome, what does an adjusted cumulative odds ratio of 1.64 indicate? Name one interpretation it does not support.
A patient starts organ support at 06:00 Tuesday and is randomized at noon Wednesday. How many hours elapsed, and is this within a 24-hour window?
Why does a larger sarilumab point estimate than tocilizumab’s not establish that sarilumab is superior?

Solutions. (1) The score is \(-1\), because death in hospital before day 90 determines the score, regardless of earlier support-free days. (2) Under the model, the adjusted cumulative odds of a more favorable outcome are estimated to be 1.64 times those under control. It does not indicate a 64-percentage-point increase in recovery probability or a 64% reduction in mortality risk. (3) From 06:00 Tuesday to 06:00 Wednesday is 24 hours, then to noon is 6 more, for \(24+6=30\) hours. This is outside the 24-hour window. (4) The point estimates are uncertain and do not themselves estimate a randomized head-to-head contrast; comparing them requires an appropriate direct contrast and its uncertainty.