Personalized oxygen therapy for critically ill adults: a randomized trial

PROJECT SUMMARY / ABSTRACT Each year, 2-3 million critically ill adults in the United States receive invasive mechanical ventilation, of whom 30-40% die before hospital discharge. Approaches to patient care that decrease mortality for mechanically ventilated adults are urgently needed. For all mechanically ventilated adults, the fraction of inspired oxygen (FiO2) is titrated to maintain arterial oxygen saturation (SpO2). Using a higher SpO2 target (96-100%) provides a margin of safety against hypoxemia, but increases exposure to excess FiO2, hyperoxemia, and tissue hyperoxia. Using a lower SpO2 target (88-92%) minimizes these risks but may increase exposure to hypoxemia and hypoxia. We conducted a randomized trial of higher vs lower SpO2 targets among 2,541 mechanically ventilated critically ill adults (Semler, NEJM 2022). In addition to reporting the effect of a higher vs lower SpO2 target on mortality among all patients in the trial population (“average treatment effect”), we analyzed the trial using machine learning methods to derive a statistical model estimating the effect of a higher vs lower SpO2 target on mortality for each individual patient (“personalized SpO2 target”) (Buell, JAMA 2024). When we validated these personalized SpO2 targets in the dataset from a second randomized trial (Mackle, NEJM 2020), we found that the model accurately identified which individual patients would benefit from receipt of a higher vs lower SpO2 target. We calculated that, had each patient received the personalized SpO2 target predicted by the statistical model to be best for him or her, overall mortality would have been reduced by 6.4%. Before the use of personalized SpO2 targets can be widely implemented into clinical care, the key next step is to prospectively evaluate whether using the personalized SpO2 targets decreases mortality for mechanically ventilated critically ill adults, compared to usual care. The EXamination of PeRsonalizEd SpO2 TargetS (EXPRESS) randomized trial proposed in this application will compare use of a personalized SpO2 target vs usual care with regard to 28-day in-hospital mortality among 3,000 mechanically ventilated patients. The trial will address two specific aims. Aim 1 will determine whether use of a personalized SpO2 target decreases mortality compared with usual care. Aim 2 will evaluate for differential treatment effects across patient subgroups to test the hypothesis that the effect of a personalized SpO2 target vs usual care on 28-day mortality will be greater for patients with a greater predicted benefit from use of the personalized SpO2 target (Aim 2a), will be greater for patients with greater medical complexity and severity of illness (Aim 2b), and will not differ by patients’ age, sex, race/ethnicity, or socioeconomic status (Aim 2c). The results of the trial will inform care for millions of critically ill patients each year. The methodological innovation of using machine learning analyses of randomized trial data to deliver care that is both evidence-based and personalized has the potential to fundamentally change the way we generate and apply evidence in medicine.