麻豆果冻传媒影音

It may be possible to develop individualized peripheral oxygen-saturation (SpO2) targets in mechanical ventilation based on a patient's characteristics using predictive algorithms derived from machine learning, according to a secondary analysis of data from two unrelated randomized trials.

A machine learning model derived in one randomized trial and validated in another showed the ability to predict up to a 27.2% absolute reduction in 28-day mortality with use of a lower SpO2 target and up to a 34.4% mortality reduction with a higher SpO2 target, depending on patient demographic and clinical characteristics, reported Kevin G. Buell, MBBS, of the University of Chicago, and colleagues in .

The most important characteristics for predicting individual treatment effects were mean arterial pressure, heart rate, age, and arterial partial pressure of carbon dioxide (PaCO2). Absolute overall mortality in the validation cohort would have been 6.4% lower (95% CI 1.9-10.9) if all patients had been treated with the SpO2 target predicted to benefit them instead of random assignment.

The heterogeneity of treatment effect makes it difficult to determine the most appropriate SpO2 targets to maximize survival in patients receiving mechanical ventilation, the authors explained, hence the lack of outcome differences seen in previous trials that attempted to identify whether a higher or lower target would benefit a particular cohort of diverse patients.

The use of machine learning to derive and then validate a predictive algorithm to identify a target based on patient characteristics, however, would enable physicians to individualize targets to maximize a particular patient's likelihood of survival.

"For example, the use of a lower SpO2 target may decrease mortality for patients with acute brain injury, whereas use of a higher SpO2 target may decrease mortality for patients with sepsis and abnormally elevated vital signs," Buell and team wrote. "These findings suggest that the use of SpO2 targets that are individualized using machine learning analyses of randomized trials may improve outcomes for critically ill adults receiving mechanical ventilation."

Randomized trials have not yet provided much guidance in determining appropriate supplemental oxygen targets that minimize risk of both hypoxemia and hyperoxemia, noted Derek C. Angus, MD, MPH, of the University of Pittsburgh Schools of the Health Sciences and a senior editor at JAMA, in an .

"Unfortunately, there is no easy way to measure directly either tissue hypoxia or oxygen toxicity," Angus wrote. "Furthermore, any consequences may not be apparent until days or weeks later."

However, if this study's results, relying on "generally robust" methods, are "true and generalizable, then the consequences are staggering," Angus added. "If one could instantly assign every patient into their appropriate group of predicted benefit or harm and assign their oxygen target accordingly, the intervention would theoretically yield the greatest single improvement in lives saved from critical illness in the history of the field."

But he also pointed to important limitations to the study, such as potentially exaggerated heterogeneity in effect models, the need to rely on imperfect simulation models to test different patient outcomes, considerations in implementation, and the need for further validation with larger patient cohorts.

Buell and colleagues used data from the Pragmatic Investigation of Optimal Oxygen Targets (PILOT) trial to derive a machine learning model for individualized treatment targets, and then they validated that model using the Intensive Care Unit Randomized Trial Comparing Two Approaches to Oxygen Therapy (ICU-ROX) cohort.

PILOT was a cluster-randomized, cluster-crossover trial that compared a lower SPO2 target of 90%, an intermediate target of 94%, and a higher target of 98% among 2,541 patients who received mechanical ventilation in a U.S. academic medical center's ICU or emergency department from July 2018 to August 2021.

ICU-ROX was a multicenter, parallel-group, randomized trial that compared conservative-oxygen therapy (91-96%) and usual-oxygen therapy (91-100%) in 965 patients who received mechanical ventilation in 21 ICUs in Australia and New Zealand from September 2015 to May 2018.

Both trials' primary outcome was the number of days alive and ventilator-free through day 28, with a secondary outcome of 28-day mortality. Neither trial found differences in outcomes between the groups.

This secondary analysis included patients from the lower and higher target groups of both trials (excluding the intermediate target group of the PILOT trial) and used 28-day mortality as the primary outcome.

The variables they used to predict appropriate treatment targets in the algorithms included patients' age and sex, source of admission to the ICU, ICU admitting diagnoses, vital signs, serum creatinine levels, receipt of vasopressors or inotropes, time from mechanical ventilation to randomization, mode of mechanical ventilation, positive end-expiratory pressure, PaCO2, missing indicator for PaCO2, and predicted risk of 28-day mortality, the latter derived from the non-respiratory Sequential Organ Failure Assessment (SOFA) score in the PILOT trial and the Acute Physiology and Chronic Health Evaluation (APACHE) II score in the ICU-ROX trial.

"The predicted risk of 28-day mortality was included as a predictor variable in the models because variation in patients' baseline risk of the primary outcome across a trial population may contribute to heterogeneity of treatment effect," the authors explained.

The researchers found in the ICU-ROX validation cohort that patients predicted to benefit from a lower SpO2 target were older, more likely to be male, and more likely to have brain injury (both hypoxic and non-hypoxic) and cardiovascular disease. Patients randomized to the lower SpO2 group who were predicted to benefit from the lower target had 6.1% lower 28-day mortality (95% CI -4.3% to 16.5%).

Patients predicted to benefit from a higher SpO2 target were younger and had a greater prevalence of sepsis and respiratory disease, as well as a higher baseline heart rate, mean arterial pressure, respiratory rate, and temperature. When these patients were randomized to the lower SpO2 target group, mortality was 13% higher (95% CI 3.5-22.6) compared with those in the higher target group.

Comment