Machine learning enables personalized oxygen delivery for ventilated patients
Staff writer, HospiMedica International
Posted on June 10, 2024
Supplemental oxygen is one of the most commonly prescribed treatments worldwide, with 13 to 20 million patients requiring oxygen delivery via mechanical ventilators each year. Mechanical ventilators are a critical life support technology that facilitates the movement of air in and out of the lungs, functioning similarly to a bellows. Modern ventilators have come a long way from the old “iron lung” machines commonly imagined. Today’s devices are sophisticated, compact, digital devices that administer oxygen through a small plastic tube inserted down the throat. Despite these technological advances, determining the appropriate oxygen level for each patient is still largely based on estimation. Clinicians set oxygen levels using a device that measures SpO2 saturation, an indication of the oxygen saturation in the patient’s blood, but studies to date have not determined whether patients benefit more from higher or lower SpO2 targets.
To take the guesswork out of ventilators, a team at the University of Chicago Medicine (Chicago, Illinois, USA) employed machine learning models to explore how different oxygen levels affect outcomes based on individual patient characteristics. Their findings suggest that personalized oxygenation targets could significantly reduce mortality rates and revolutionize the practice of critical care medicine. Previous studies by different research groups have attempted to determine whether higher or lower oxygen levels are favorable, but generally, these studies have yielded inconclusive results. The UChicago Medicine researchers proposed that the neutral results of these trials do not mean that oxygen levels are irrelevant to patient outcomes, but rather that the effect of different oxygen levels may vary from patient to patient and, on average, may be zero in randomized trials.
Image: Personalized oxygen delivery could improve outcomes for ventilated patients (Photo credit: 123RF)
As personalized medicine becomes more prevalent, there is growing interest in leveraging machine learning to predict optimal treatments for individual patients. In the field of mechanical ventilation, these predictive models could determine a patient's ideal oxygen saturation level based on certain characteristics such as age, sex, heart rate, temperature, and reason for ICU admission. The team and their colleagues utilized data from a previous randomized trial to develop and refine their machine learning model. After initial development using U.S. data, the model was applied to patient data from Australia and New Zealand. Their findings showed that patients who achieved oxygen levels that the model deemed optimal experienced a 6.4% reduction in overall mortality. It is important to note that outcomes cannot be universally predicted based on a single characteristic. For example, not all patients with brain injury benefit from lower oxygen levels, despite data trends suggesting otherwise. This calls for comprehensive tools like machine learning models that integrate diverse patient data.
Although the algorithms are complex, the input variables are common clinical parameters, so it would be easy for medical teams to use such tools in the future. At UChicago Medicine, the algorithms have already been integrated directly into the electronic health record (EHR) system to support a range of clinical decisions. Researchers believe that ventilators could be managed in a similar way. Hospitals that do not have the resources to integrate such advanced machine learning tools into their EHRs could also develop web-based applications with online calculator-like functionality where clinicians can input patient characteristics and receive predictions. These applications require extensive validation, testing, and refinement before they can be implemented clinically, but the potential benefits justify investment in further research.
“If the results are real and generalizable, the results would be astounding,” says critical care specialist Derek Angus, M.D. “If we could instantly assign every patient to the appropriate group of predicted benefit or harm, and assign oxygen targets accordingly, this intervention would theoretically produce the largest single improvement in lives saved from critical illness in the history of the field.”
Related Links:
University of Chicago Medicine
