Oxygen Targets and the Future of Critical Care Clinical Research

In this issue of the Journal, van der Wal and colleagues (pp. 770–779) report the findings of the “ConservatIve versus CONventional oxygenation targets in Intensive Care patients” (ICONIC) trial, a randomized trial that compared the use of a lower oxygenation target (Pa_O2 = 55–80 mm Hg) with the use of a higher oxygenation target (Pa_O2 = 110–150 mm Hg) among 664 patients receiving invasive mechanical ventilation in nine ICUs in Europe (1). The primary outcome of mortality at 28 days did not differ significantly between the lower oxygenation target group (38.5%) and the higher oxygenation target group (34.7%), P = 0.30. Secondary outcomes and adverse events also did not differ between groups.

The ICONIC trial shares several strengths with prior trials of oxygenation targets: use of concealed allocation to prevent selection bias and randomization to balance baseline characteristics; selection of mortality as an objective, patient-centered outcome; and enrollment of a severely ill population with a high event rate. The trial also has unique strengths. First, most patients were enrolled within minutes of first receiving of invasive mechanical ventilation, capturing the period of highest risk for hyperoxemia and hypoxemia and minimizing contamination between groups from oxygen exposure after initiation of mechanical ventilation before enrollment. This was only achievable because, like multiple prior trials of oxygenation targets (2–4), ICONIC permitted enrollment before receiving informed consent, an ethically appropriate approach to comparing emergency treatments in common use in current clinical care. Second, the separation in Pa_O2 values achieved between the lower target group (median, 75 mm Hg; interquartile range, 70–83) and the higher target group (median, 115; interquartile range, 100–129) was greater than in most prior trials.

The ICONIC trial also has significant limitations. First, because of the coronavirus disease (COVID-19) pandemic, the trial was prematurely terminated after the enrollment of 664 (43.9%) of the planned 1,512 patients. Early termination increases the risk of failing to detect a true difference between groups (type II error). This is especially relevant because the absolute numerical difference between groups observed in the trial (3.8%) was both similar in magnitude to the difference that the trial was originally powered to detect (6.0%) and, if a true difference, large enough to justify changing oxygenation target practices. Second, 218 (24.7%) of the patients enrolled in the trial, who were randomized and treated with assigned oxygenation targets, were ultimately excluded from the analyses because they met exclusion criteria or because informed consent could not be obtained. These postrandomization exclusions weaken the main strength of the randomized design by reintroducing the opportunity for selection bias and violating the intention-to-treat principle.

The results of the ICONIC trial, along with the results of seven other recent trials of oxygenation targets, make clear that use of a lower versus higher oxygenation target does not result in a large difference in outcomes for critically ill adults receiving mechanical ventilation overall. But what message does the story of research on oxygenation targets convey about the current paradigm and future of critical care research?

To date, research on oxygenation targets has followed the same sequence as research on numerous other critical care interventions (Figure 1). First, a common treatment was administered in clinical care to millions of patients each year for decades without any attempt to determine how the use of the treatment affected patient outcomes. Second, observational studies identified variation in the way the treatment was administered, which suggested that 1) different approaches to treatment might be associated with differences in patient outcomes and that 2) clinical equipoise existed for a randomized trial (5, 6). Third, critical care research groups each independently conducted a moderate-sized randomized trial in their geographic region, enrolling a sufficient number of patients (e.g., 500–2,500) to rule out a large difference in outcomes between groups (e.g., 5–10%) in the trial population as a whole (average treatment effect). Historically, a moderate-sized randomized trial that does not detect a large difference in average treatment effect has been the end of the story for research on most critical care interventions. This has contributed to medical nihilism (“no interventions work in critical care”), research nihilism (“critical care trials are always negative”), and overreliance on nonevidentiary knowledge (“clinical experience is sufficient to determine which treatments benefit which patients”).

Figure 1.

How research on oxygenation targets may inform the historical and future paradigms of critical care clinical research.

For oxygenation targets, however, the research is continuing in two novel directions that, separately or together, may contribute to a new paradigm for critical care clinical research. First, on the basis of the premise that, for an inexpensive and common treatment such as oxygen therapy, even small differences in outcomes would be clinically meaningful, research groups from across the world are collaborating on a registry-embedded randomized trial that will enroll 40,000 patients—sufficient to detect a 1.5% average treatment effect between lower and higher oxygenation targets (7). Developing the efficiencies in trial design, conduct, and regulation required to enroll tens of thousands of patients in critical care trials would allow the rigorous comparison of a vast number of treatments that patients are currently receiving in clinical care, which may meaningfully affect outcomes but cannot realistically be expected to decrease mortality by 5–10%.

Second, randomized trials traditionally report the average effect of a treatment for all patients in the population, but clinicians need to know which treatment would be best for an individual patient on the basis of his or her unique characteristics. To address this, advanced analytic methods, including machine learning, are being applied to datasets from randomized trials of oxygenation targets to derive and validate estimates of the effect of treatment with lower versus higher oxygenation targets on outcomes for individual patients, considering all their individual characteristics simultaneously. Several recent studies using different methods of generating evidence-based individualized treatment effect estimates have found that treatment effect differed dramatically for different patients, even in critical care trials with no average treatment effect (8–10).

One vision for the future of critical care clinical research would unite these two novel approaches. When observational research identifies variability in clinical care, signifying an evidence gap and clinical equipoise, large pragmatic trials could be embedded within clinical care to enroll tens of thousands of patients who represent the full diversity of critically ill patients in practice. Using data from these trials about patients’ baseline demographics, comorbidities, vital signs, laboratory values, and other characteristics, machine learning methods could be used to derive and validate models for estimating the effects of treatment for individual patients. These models could be incorporated into decision support in the electronic health record to help guide clinicians to treatments that, as rigorous evidence from randomized trials indicates, will result in the best outcomes for their individual patients. In a best-case scenario, the study of oxygen therapy, one of the oldest and most routine treatments in critical care, could help bring about a future in which unprecedented trial designs (massive global trials) combine with unprecedented analytic methods (artificial intelligence) to deliver what clinicians and patients have needed all along: medicine that is both evidence based and personalized.