What If We Used Personalized Oxygen Saturation Targets?
June 5, 2024
So sorry for the late email today. We had a technical issue…think it’s fixed! ~Clay
Written by Alex Clark
Spoon Feed
This secondary analysis derives and validates a machine learning model that supports the use of personalized oxygen saturation targets (SpO2) in mechanically ventilated critically ill adults based on individual patient characteristics.
Personalized medicine ROX!
This secondary analysis uses two temporally and geographically distinct RCTs (PILOT & ICU-ROX trials) to derive and externally validate a machine learning model that aims to identify specific cohorts of patients that may benefit from lower vs. higher SpO2 targets.
In brief, the authors begin by using the PILOT trial to predict individualized treatment effects from oxygen therapy for each patient. Then using a priori predictor characteristics, they stratified all patients in the ICU-ROX trial into one of three quantiles: those predicted to benefit from (1) lower, (2) higher, or (3) either SpO2 target equally. Patients in the lower group were more likely to be older, male, or suffer from brain injury and/or cardiovascular disease. The higher group was more likely to have sepsis, respiratory disease, or abnormal vital signs. Patients in each quantile received either high or low SpO2 targets based on their actual randomization in the ICU-ROX trial. The primary outcome was the absolute difference in 28-day mortality between these groups.
Overall, patients predicted to benefit from lower SpO2 had lower 28-day mortality when randomized to the lower SpO2 group (-6.1%; 95%CI −4.3% to 16.5%), and those predicted to benefit from higher SpO2 had greater mortality with lower targets (13.0%; 95%CI 3.5% to 22.6%). Additionally, the authors predict an absolute reduction in 28-day mortality of 6.4% (95%CI 1.9% to 10.9%) if all patients had been treated with their “ideal” SpO2 target compared to their randomly assigned group.
How will this change my practice?
This secondary analysis suggests that machine learning algorithms have the potential to improve outcomes in mechanically ventilated critically ill adults. Patient heterogeneity combined with indiscriminate application of interventions may explain a lack of significant outcomes in prior RCTs. Due to limitations related to its retrospective nature and differences in trial design, this analysis is hypothesis generating. However, I suspect that machine learning will continue to play a powerful role in providing targeted, personalized care to critically ill patients.
Source
Individualized Treatment Effects of Oxygen Targets in Mechanically Ventilated Critically Ill Adults. JAMA. 2024 Apr 9;331(14):1195-1204. doi: 10.1001/jama.2024.2933. PMID: 38501205; PMCID: PMC10951851