By Samuel Nadler, MD, PhD
Clinical Instructor, University of Washington, Seattle
SYNOPSIS: Using a machine learning algorithm, a predictive model demonstrated that different oxygen targets for mechanically ventilated patients may improve outcomes.
SOURCE: Buell KG, Spicer AB, Casey JD, et al. Individualized treatment effects of oxygen targets in mechanically ventilated critically ill adults. JAMA 2024;331:1195-1204.
Optimal oxygenation strategies for patients in the intensive care unit (ICU) on mechanical ventilation remain an active area of research. Uncertainty regarding oxygen saturation (SpO2) goals led to studies comparing higher vs. lower targets summarized in a recent Cochrane review.1 Heterogeneity within study groups of these negative trials and subsequent subset analysis suggesting specific populations might have benefited raises the question as to whether more individualized SpO2 targets might yield a positive result. To address this question, the current study used machine learning to derive a predictive model of the effects of patient-specific data on outcomes from the PILOT study.2 This model was validated against the ICU-ROX study, a temporally and geographically distinct trial.3 The cohorts included 1,682 patients from the PILOT study and 965 patients from the ICU-ROX study. The authors reported the predicted treatment effects of lower vs. higher SpO2 targets could vary from a 27.2% absolute reduction in 28-day mortality with the use of lower SpO2 targets to a 34.4% absolute increase in 28-day mortality with higher SpO2 targets. This suggests that use of more individualized targets could have a significant impact on outcome in mechanically ventilated patients.
Using this machine learning approach, patient-specific characteristics modified predicted SpO2 treatment targets for optimal outcomes. Mean arterial pressure (MAP), heart rate, age, and arterial partial pressure of carbon dioxide (PaCO2) had the strongest influence on these outcomes. Increasing age and MAP correlated with higher mortality using lower SpO2 targets. Lower PaCO2 was associated with higher mortality with lower SpO2 targets. Overall, patients predicted to benefit from lower SpO2 targets had a higher prevalence of both hypoxic and non-hypoxic brain injuries as well as cardiovascular disease. In contrast, those patients predicted to benefit from higher SpO2 targets had a higher prevalence of sepsis and respiratory disease. In total, the use of these patient-level variables to modify SpO2 targets was associated with an absolute 6.4% lower 28-day mortality (95% confidence interval, 1.9% to 10.9%).
COMMENTARY
The current study is a thought-provoking analysis of two previously published randomized controlled trials involving higher vs. lower SpO2 targets in patients receiving mechanical ventilation. The authors used multiple machine learning algorithms to generate the best predictive model for which patients might benefit from higher vs. lower oxygen targets. They subsequently validated this model in a separate data set and showed that a more individualized approach might significantly reduce mortality. Interestingly, if the validation data set was used to generate the model and the baseline data set used to validate the model, a less robust prediction was achieved, which raises questions about generalizability. Because this was not a prospective study of the model’s ability to improve outcomes, this study is not strong enough to change practice and redefine SpO2 targets for specific populations.
Beyond predicting oxygenation targets, this study foreshadows changes in the way we use data and conduct clinical trials that will change our practice. This is a clear example of how machine learning and big data could be used to redefine patient populations in a study. Rather than randomization of groups to one intervention or another, a study might randomize for individualized targets for a particular intervention vs. generalized targets. In that paradigm, we would not simply consider a patient with sepsis as a candidate for higher SpO2 targets, but a septic patient of a certain age presenting with specific demographic characteristics, microbiology, laboratory values, or other patient-specific variables to have an individualized oxygenation target. These predictive models must be robust over time, since many variables can change with time and treatment. In addition, they must be generalizable to all patients. Particularly with oxygenation targets, we also know that skin color can modify the reliability of these measurements. With larger and more inclusive data sets, studies like this may help hone our treatment targets and improve patient care.
REFERENCES
- Klitgaard TL, Schjørring OL, Nielsen FM, et al. Higher versus lower fractions of inspired oxygen or targets of arterial oxygenation for adults admitted to the intensive care unit. Cochrane Database Syst Rev 2023;9:CD012631.
- Semler MW, Casey JD, Lloyd BD, et al. Oxygen-saturation targets for critically ill adults receiving mechanical ventilation. N Engl J Med 2022;387:1759-1769.
- ICU-ROX Investigators and the Australian and New Zealand Intensive Care Society Clinical Trials Group; Mackle D, Bellomo R, Bailey M, et al. Conservative oxygen therapy during mechanical ventilation in the ICU. N Engl J Med 2020;382:989-998.
