Mortality measured at a particular time point (landmark mortality) is often regarded as the gold standard outcome for randomised controlled trials in Intensive Care Medicine. An important limitation of many Intensive Care Medicine trials is that they hypothesize large and potentially implausible reductions in absolute mortality. This is a major problem in trial design for two reasons. Firstly, it makes false negative trial results more likely. Secondly, the less plausible a postulated mortality reduction is the more likely that a statistically significant mortality difference will represent a false positive. This is because a p-value is defined as the probability of finding a result equal to or more extreme than that actually observed, under the assumption that the null hypothesis is true. This means that the greater the pre-trial chance or prior probability that the null hypothesis is correct, the lower the chance that a p-value below a particular significance threshold will represent a true positive.
The biggest single problem with the current evidence base is that most hypotheses being tested have low prior probability. We need a new research paradigm to address this problem, particularly in relation to the fundamentals of Intensive Care Medicine. Intensive Care therapy is fundamentally about providing supportive care. Such care includes airway support, oxygen therapy, ventilation therapy, haemodynamic support, fluid therapy, temperature control, and nutritional therapy among others. Setting the goals for these therapies is what intensive care doctors do every day. At present, for most of these treatments, the level of evidence on which we are making our decisions is extremely limited. Moreover, the illusion of physiological gain may be leading us astray; making us believe we know the right thing to do when we really do not. We should be creating systems in our intensive care units that allow us to learn iteratively from every patient so that we can systematically reduce mortality over time by understanding how to optimise supportive care. Bayesian adaptive platform trials using response adaptive randomisation can improve the outcomes of patients with mathematical precision but require us, as doctors, to stop believing we know the answers when we really do not.